This Java native memory leak debugging war story from a Twitter engineer
is very interesting:
http://www.evanjones.ca/java-native-leak-bug.html
Tweet is https://twitter.com/epcjones/status/603295445067014144 .

It seems to be very important to check all locations where
GzipInputStream (and other InflaterInputStream impl) and
GzipOutputStream (and other DeflaterOutputStream impl) are used.

I assume that an InputStream opened from a resource URL originating from
a Jar file could also leak native memory.
(ClassLoader.getResource(...).openStream()). This makes it a very common
source of native memory problems in Java.

It could be a coincidence, but Tomcat 8.0.20 seems to have changes in
this area:
https://github.com/apache/tomcat/commit/6e5420c67fbad81973d888ad3701a392fac4fc71
(I linked to that commit in my email on May 14).

Lari

On Thu, May 14, 2015 at 10:23 AM, Daniel Jones <
daniel.jones(a)engineerbetter.com> wrote:

Hi Lari,

Thanks again for your input. Have you seen this problem with versions of
Tomcat before 8.0.20?

David and I think we've narrowed down the issue to a change from using
Tomcat 8.0.18 to 8.0.21. We're running more tests and collaborating with
Pivotal support. We also noticed that non-prod versions of our apps were
taking longer to crash, so it would seem to be activity-related at least.

Since it seems activity related, have you tried monitoring the number of
threads in the JVM?

While you can cap the number of threads Tomcat uses for processing
requests, I don't believe that you can cap the number of threads it's
possible to create in the JVM. The reason I mention this is because each
thread causes the amount of memory required to go up by the thread stack
size (Xss * <threads> = total thread memory). Perhaps as activity
increases, so does the thread count and that's pushing you over the limit.

Are you setting a custom -Xss value or using the default? From memory, the
default is pretty large. If you're not using a custom one, you might try a
smaller one, like 256k and see if that has any impact on the problem.

Alternatively, you could adjust the memory weightings in the build pack so
that heap consumes a smaller amount of the total memory and there's more
memory available for native / stack and other memory.

Do you know how Tomcat's APR/NIO memory gets allocated?

I'm not sure I follow your question, but in terms of NIO I suspect the JVM
is going to handle memory allocation not Tomcat. Given that, I it should
happen just like any other Java code that uses the NIO. APR is unlikely to
be an issue, see my next comment.

Is there a way of telling from pmap whether pages are being used for NIO
buffers or by the APR?

Unless you compile the APR native library and include it with your version
of the build pack, Tomcat's not going to use it. It'll use NIO by default,
with Tomcat version 8. You can confirm by looking at the logs when you
start your app. One of the first things it logs will be the protocol
handler. This is from a test app, where it's using NIO.

Ex:

```
[CONTAINER] org.apache.coyote.http11.Http11NioProtocol INFO
Initializing ProtocolHandler ["http-nio-63227"]
```

Dan

I wonder if the other folks that have reported CF out of memory errors
with later versions of Tomcat are seeing slow creeps in native memory
consumption?

On Mon, May 11, 2015 at 2:19 PM, Lari Hotari <Lari(a)hotari.net> wrote:

fyi. Tomcat 8.0.20 might be consuming more memory than 8.0.18:

https://github.com/cloudfoundry/java-buildpack/issues/166#issuecomment-94517568

Other things we’ve tried:

- We set verbose garbage collection to verify there was no
memory size issues within the JVM. There wasn’t.

- We tried setting minimum memory for native, it had no effect.
The container still gets killed

- We tried adjusting the ‘memory heuristics’ so that they added
up to 80 rather than 100. This had the effect of causing a delay in the
container being killed. However it still was killed.

I think adjusting memory heuristics so that they add up to 80 doesn't
make a difference because the values aren't percentages.
The values are proportional weighting values used in the memory
calculation:

https://github.com/grails-samples/java-buildpack/blob/b4abf89/docs/jre-oracle_jre.md#memory-calculation

I found out that the only way to reserve "unused" memory is to set a high
value for the native memory lower bound in the memory_sizes.native setting
of config/open_jdk_jre.yml .
Example:

https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25



This seems like classic memory leak behaviour to me.

In my case it wasn't a classical Java memory leak, since the Java
application wasn't leaking memory. I was able to confirm this by getting
some heap dumps with the HeapDumpServlet (
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy)
and analyzing them.

In my case the JVM's RSS memory size is slowly growing. It probably is
some kind of memory leak since one process I've been monitoring now is very
close to the memory limit. The uptime is now almost 3 weeks.

Here is the latest diff of the meminfo report.

https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example2-txt

From a Java perspective this isn't classical. The JVM heap isn't filling
up. The problem is that RSS size is slowly growing and will eventually
cause the Java process to cross the memory boundary so that the process
gets kill by the Linux kernel cgroups OOM killer.

RSS size might be growing because of many reasons. I have been able to
slow down the growth by doing the various MALLOC_ and JVM parameter tuning
(-XX:MinMetaspaceExpansion=1M -XX:CodeCacheExpansionSize=1M). I'm able to
get a longer uptime, but the problem isn't solved.

Lari



On 15-05-11 06:41 AM, Head-Rapson, David wrote:

Thanks for the continued advice.



We’ve hit on a key discovery after yet another a soak test this weekend.

- When we deploy using Tomcat 8.0.18 we don’t see the issue

- When we deploy using Tomcat 8.0.20 (same app version, same CF
space, same services bound, same JBP code version, same JRE version,
running at the same time), we see the crashes occurring after just a couple
of hours.



Ideally we’d go ahead with the memory calculations you mentioned however
we’re stuck on lucid64 because we’re using Pivotal CF 1.3.x & we’re having
upgrade issues to 1.4.x.

So we’re not able to adjust MALLOC_ARENA_MAX, nor are we able to view RSS
in pmap as you describe



Other things we’ve tried:

- We set verbose garbage collection to verify there was no
memory size issues within the JVM. There wasn’t.

- We tried setting minimum memory for native, it had no effect.
The container still gets killed

- We tried adjusting the ‘memory heuristics’ so that they added
up to 80 rather than 100. This had the effect of causing a delay in the
container being killed. However it still was killed.



This seems like classic memory leak behaviour to me.



*From:* Lari Hotari [mailto:lari.hotari(a)sagire.fi <lari.hotari(a)sagire.fi>]
*On Behalf Of *Lari Hotari
*Sent:* 08 May 2015 16:25
*To:* Daniel Jones; Head-Rapson, David
*Cc:* cf-dev(a)lists.cloudfoundry.org
*Subject:* Re: [Cf-dev] [vcap-dev] Java OOM debugging




For my case, it turned out to be essential to reserve enough memory for
"native" in the JBP. For the 2GB total memory, I set the minimum to 330M.
With that setting I have been able to get over 2 weeks up time by now.

I mentioned this in my previous email:

The workaround for that in my case was to add a native key under
memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is for a
2GB total memory).
see example
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the memory
bounds. I'm sure there is now also a way to get the keys without forking
the buildpack. I could have also adjusted the percentage portions, but I
wanted to set a hard minimum for this case.


I've been trying to get some insight by diffing the reports gathered from
the meminfo servlet https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy


Here is such an example of a diff:

https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example-txt

meminfo has pmap output included to get the report of the memory map of
the process. I have just noticed that most of the memory has already been
mmap:ed from the OS and it's just growing in RSS size. For example:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

The pmap output from lucid64 didn't include the RSS size, so you have to
use cflinuxfs2 for this. It's also better because of other reasons. The
glibc in lucid64 is old and has some bugs around the MALLOC_ARENA_MAX.

I was manually able to estimate the maximum size of the RSS size of what
the Java process will consume by simply picking the large anon-blocks from
the pmap report and calculating those blocks by the allocated virtual size
(VSS).
Based on this calculation, I picked the minimum of 330M for "native" in
open_jdk_jre.yml as I mentioned before.

It looks like these rows are for the Heap size:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

It looks like the JVM doesn't fully allocate that block in RSS initially
and most of the growth of RSS size comes from that in my case. In your
case, it might be something different.

I also added a servlet for getting glibc malloc_info statistics in XML
format (). I haven't really analysed that information because of time
constraints and because I don't have a pressing problem any more. btw. The
malloc_info XML report is missing some key elements, that has been added in
later glibc versions (
https://github.com/bminor/glibc/commit/4d653a59ffeae0f46f76a40230e2cfa9587b7e7e
).

If killjava.sh never fires and the app crashed with Warden out of memory
errors, then I believe it's the kernel's cgroups OOM killer that has killed
the container processes. I have found this location where Warden oom
notifier gets the OOM notification event:

https://github.com/cloudfoundry/warden/blob/ad18bff/warden/lib/warden/container/features/mem_limit.rb#L70
This is the oom.c source code:
https://github.com/cloudfoundry/warden/blob/ad18bff7dc56acbc55ff10bcc6045ebdf0b20c97/warden/src/oom/oom.c
. It reads the cgroups control files and receives events from the kernel
that way.

I'd suggest that you use pmap for the Java process after it has started
and calculate the maximum RSS size by calculating the VSS size of the large
anon blocks instead of RSS for the blocks that the Java process has
reserved for it's different memory areas (I think you shouldn't . You
should discard adding VSS for the CompressedClassSpaceSize block.
After this calculation, add enough memory to the "native" parameter in
JBP until the RSS size calculated this way stays under the limit.
That's the only "method" I have come up by now.

It might be required to have some RSS space allocated for any zip/jar
files read by the Java process. I think that Java uses mmap files for zip
file reading by default and that might go on top of all other limits.
To test this theory, I'd suggest testing by adding
-Dsun.zip.disableMemoryMapping=true system property setting to JAVA_OPTS.
That disables the native mmap for zip/jar file reading. I haven't had time
to test this assumption.

I guess the only way to understand how Java allocates memory is to look
at the source code.
from http://openjdk.java.net/projects/jdk8u/ , the instructions to get
the source code of JDK 8:
hg clone http://hg.openjdk.java.net/jdk8u/jdk8u;cd jdk8u;sh
get_source.sh
This tool is really good for grepping and searching the source code:
http://geoff.greer.fm/ag/
On Ubuntu it's in silversearcher-ag package, "apt-get install
silversearcher-ag" and on MacOSX brew it's "brew install
the_silver_searcher".
This alias is pretty useful:
alias codegrep='ag --color --group --pager less -C 5'
Then you just search for the correct location in code by starting with
the tokens you know about:
codegrep MaxMetaspaceSize
this gives pretty good starting points in looking how the JDK allocates
memory.

So the JDK source code is only a few commands away.

It would be interesting to hear more about this if someone has the time
to dig in to this. This is about how far I got and I hope sharing this
information helps someone continue. :)


Lari
github/twitter: lhotari

On 15-05-08 10:02 AM, Daniel Jones wrote:

Hi Lari et al,



Thanks for your help Lari.



David and I are pairing on this issue, and we're yet to resolve it. We're
in the process of creating a repeatable test case (our most crashy app
makes calls to external services that need mocking), but in the meantime,
here's what we've seen.



Between Java Buildpack commit e89e546 and 17162df, we see apps crashing
with Warden out of memory errors. killjava.sh never fires, and this has led
us to believe that the kernel is shooting a cgroup process in the head
after the cgroup oversteps its memory limit. We cannot find any evidence of
the OOM killer firing in any logs, but we may not be looking in the right
place.



The JBP is setting heap to be 70%, metaspace to be 15% (with max set to
the same as initial), 5% for "stack", 5% for "normalised stack" and 10% for
"native". We do not understand why this adds up to 105%, but haven't looked
into the JBP algorithm yet. Any pointers on what "normalised stack" is
would be much appreciated, as this doesn't appear in the list of heuristics
supplied via app env.



Other team members tried applying the same settings that you suggested -
thanks for this. Apps still crash with these settings, albeit less
frequently.



After reading the blog you linked to (
http://java.dzone.com/articles/java-8-permgen-metaspace) we wondered
whether the increased *reserved *metaspace claimed after metaspace GC
might be causing a problem; however we reused the test code to create a
metaspace leak in a CF app and saw metaspace GCs occur correctly, and
memory usage never grow over MaxMetaspaceSize. This figures, as the
committed metaspace is still less than MaxMetaspaceSize, and the reserved
appears to be whatever RAM is free across the whole DEA.



We noted that an Oracle blog (
https://blogs.oracle.com/poonam/entry/about_g1_garbage_collector_permanent)
mentions that the metaspace size parameters are approximate. We're
currently wondering if native allocations by Tomcat (APR, NIO) are taking
up more container memory, and so when the metaspace fills, it's creeping
slightly over the limit and triggering the kernel's OOM killer.



Any suggestions would be much appreciated. We've tried to resist tweaking
heuristics blindly, but are running out of options as we're struggling to
figure out how the Java process is using *committed* memory. pmap seems
to show virtual memory, and so it's hard to see if things like the
metaspace or NIO ByteBuffers are nabbing too much and trigger the kernel's
OOM killer.



Thanks for all your help,



Daniel Jones & David Head-Rapson



On Wed, Apr 29, 2015 at 8:07 PM, Lari Hotari <Lari(a)hotari.net> wrote:

Hi,

I created a few tools to debug OOM problems since the application I was
responsible for running on CF was failing constantly because of OOM
problems. The problems I had, turned out not to be actual memory leaks in
the Java application.

In the "cf events appname" log I would get entries like this:
2015-xx-xxTxx:xx:xx.00-0400 app.crash appname index: 1,
reason: CRASHED, exit_description: out of memory, exit_status: 255

These type of entries are produced when the container goes over it's
memory resource limits. It doesn't mean that there is a memory leak in the
Java application. The container gets killed by the Linux kernel oom killer (
https://github.com/cloudfoundry/warden/blob/master/warden/README.md#limit-handle-mem-value)
based on the resource limits set to the warden container.

The memory limit is specified in number of bytes. It is enforced using
the control group associated with the container. When a container exceeds
this limit, one or more of its processes will be killed by the kernel.
Additionally, the Warden will be notified that an OOM happened and it
subsequently tears down the container.

In my case it never got killed by the killjava.sh script that gets called
in the java-buildpack when an OOM happens in Java.

This is the tool I built to debug the problems:
https://github.com/lhotari/java-buildpack-diagnostics-app
I deployed that app as part of the forked buildpack I'm using.
Please read the readme about what it's limitations are. It worked for me,
but it might not work for you. It's opensource and you can fork it. :)

There is a solution in my toolcase for creating a heapdump and uploading
that to S3:

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy
The readme explains how to setup Amazon S3 keys for this:
https://github.com/lhotari/java-buildpack-diagnostics-app#amazon-s3-setup
Once you get a dump, you can then analyse the dump in a java profiler
tool like YourKit.

I also have a solution that forks the java-buildpack modifies killjava.sh
and adds a script that uploads the heapdump to S3 in the case of OOM:

https://github.com/lhotari/java-buildpack/commit/2d654b80f3bf1a0e0f1bae4f29cb85f56f5f8c46

In java-buildpack-diagnostics-app I have also other tools for getting
Linux operation system specific memory information, for example:


https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemorySmapServlet.groovy

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MallocInfoServlet.groovy

These tools are handy for looking at details of the Java process RSS
memory usage growth.

There is also a solution for getting ssh shell access inside your
application with tmate.io:

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/TmateSshServlet.groovy
(this version is only compatible with the new "cflinuxfs2" stack)

It looks like there are serious problems on CloudFoundry with the memory
sizing calculation. An application that doesn't have a OOM problem will get
killed by the oom killer because the Java process will go over the memory
limits.
I filed this issue:
https://github.com/cloudfoundry/java-buildpack/issues/157 , but that
might not cover everything.

The workaround for that in my case was to add a native key under
memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is for a
2GB total memory).
see example
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the memory
bounds. I'm sure there is now also a way to get the keys without forking
the buildpack. I could have also adjusted the percentage portions, but I
wanted to set a hard minimum for this case.

It was also required to do some other tuning.

I added this to JAVA_OPTS:
-XX:CompressedClassSpaceSize=256M -XX:InitialCodeCacheSize=64M
-XX:CodeCacheExpansionSize=1M -XX:CodeCacheMinimumFreeSpace=1M
-XX:ReservedCodeCacheSize=200M -XX:MinMetaspaceExpansion=1M
-XX:MaxMetaspaceExpansion=8M -XX:MaxDirectMemorySize=96M
while trying to keep the Java process from growing in RSS memory size.

The memory overhead of a 64 bit Java process on Linux can be reduced by
specifying these environment variables:

stack: cflinuxfs2
.
.
.
env:
MALLOC_ARENA_MAX: 2
MALLOC_MMAP_THRESHOLD_: 131072
MALLOC_TRIM_THRESHOLD_: 131072
MALLOC_TOP_PAD_: 131072
MALLOC_MMAP_MAX_: 65536

MALLOC_ARENA_MAX works only on cflinuxfs2 stack (the lucid64 stack has a
buggy version of glibc).

explanation about MALLOC_ARENA_MAX from Heroku:
https://devcenter.heroku.com/articles/tuning-glibc-memory-behavior
some measurement data how it reduces memory consumption:
https://devcenter.heroku.com/articles/testing-cedar-14-memory-use

I have created a PR to add this to CF java-buildpack:
https://github.com/cloudfoundry/java-buildpack/pull/160

I also created an issues
https://github.com/cloudfoundry/java-buildpack/issues/163 and
https://github.com/cloudfoundry/java-buildpack/pull/159 .

I hope this information helps others struggling with OOM problems in CF.
I'm not saying that this is a ready made solution just for you. YMMV. It
worked for me.

-Lari




On 15-04-29 10:53 AM, Head-Rapson, David wrote:

Hi,

I’m after some guidance on how to get profile Java apps in CF, in order
to get to the bottom of memory issues.

We have an app that’s crashing every few hours with OOM error, most
likely it’s a memory leak.

I’d like to profile the JVM and work out what’s eating memory, however
tools like yourkit require connectivity INTO the JVM server (i.e. the
warden container), either via host / port or via SSH.

Since warden containers cannot be connected to on ports other than for
HTTP and cannot be SSHd to, neither of these works for me.



I tried installed a standalone JDK onto the warden container, however as
soon as I ran ‘jmap’ to invoke the dump, warden cleaned up the container –
most likely for memory over-consumption.



I had previously found a hack in the Weblogic buildpack (
https://github.com/pivotal-cf/weblogic-buildpack/blob/master/docs/container-wls-monitoring.md)
for modifying the start script which, when used with
–XX:HeapDumpOnOutOfMemoryError, should copy any heapdump files to a file
share somewhere. I have my own custom buildpack so I could use something
similar.

Has anyone got a better solution than this?



We would love to use newrelic / app dynamics for this however we’re not
allowed. And I’m not 100% certain they could help with this either.



Dave



The information transmitted is intended for the person or entity to which
it is addressed and may contain confidential, privileged or copyrighted
material. If you receive this in error, please contact the sender and
delete the material from any computer. Fidelity only gives information on
products and services and does not give investment advice to retail clients
based on individual circumstances. Any comments or statements made are not
necessarily those of Fidelity. All e-mails may be monitored. FIL
Investments International (Reg. No.1448245), FIL Investment Services (UK)
Limited (Reg. No. 2016555), FIL Pensions Management (Reg. No. 2015142) and
Financial Administration Services Limited (Reg. No. 1629709) are authorised
and regulated in the UK by the Financial Conduct Authority. FIL Life
Insurance Limited (Reg No. 3406905) is authorised in the UK by the
Prudential Regulation Authority and regulated in the UK by the Financial
Conduct Authority and the Prudential Regulation Authority. Registered
offices at Oakhill House, 130 Tonbridge Road, Hildenborough, Tonbridge,
Kent TN11 9DZ.

--
You received this message because you are subscribed to the Google Groups
"Cloud Foundry Developers" group.
To view this discussion on the web visit
https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com
<https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com?utm_medium=email&utm_source=footer>
.
To unsubscribe from this group and stop receiving emails from it, send an
email to vcap-dev+unsubscribe(a)cloudfoundry.org.




_______________________________________________
Cf-dev mailing list
Cf-dev(a)lists.cloudfoundry.org
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev





--

Regards,



Daniel Jones

EngineerBetter.com

--
Regards,

Daniel Jones
EngineerBetter.com

_______________________________________________
cf-dev mailing list
cf-dev(a)lists.cloudfoundry.org
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev

On Thu, May 14, 2015 at 2:59 PM, Lari Hotari <Lari(a)hotari.net> wrote:

On 15-05-14 10:23 AM, Daniel Jones wrote:

Thanks again for your input. Have you seen this problem with versions
of Tomcat before 8.0.20?

I don't have proper data gathered from older than 8.0.20, so I cannot
compare.
I was just wondering when did 8.0.20 become available in JBP, I found
this date:
HEAD https://download.run.pivotal.io/tomcat/tomcat-8.0.20.tar.gz | grep
Last-Modified
Last-Modified: Tue, 03 Mar 2015 11:35:19 GMT

David and I think we've narrowed down the issue to a change from using
Tomcat 8.0.18 to 8.0.21. We're running more tests and collaborating
with Pivotal support. We also noticed that non-prod versions of our
apps were taking longer to crash, so it would seem to be
activity-related at least.

Do you know how Tomcat's APR/NIO memory gets allocated? Is there a way
of telling from pmap whether pages are being used for NIO buffers or
by the APR?

I don't think you can get the info from pmap. The malloc_info xml shows
better allocation stats, but only stats.
Is Tomcat using APR library or NIO by default in CloudFoundry? I'd
assume that NIO isn't used by default.

Sorry for the duplicate. Resending as my first reply didn't go back to the
list.

Since the Connector in server.xml is not specifically setting an
implementation it should use the NIO connector, which is the default in
Tomcat 8. A quick test on PWS confirmed this for me.


https://github.com/cloudfoundry/java-buildpack/blob/master/resources/tomcat/conf/server.xml#L22

It would be interesting to see if the BIO or APR connectors have similar
issues. BIO would be easy to test, just add
`protocol="org.apache.coyote.http11.Http11Protocol"` to the Connector tag
on line #22.

APR would be trickier as you'd need to compile the native library and pull
that into the environment.

Dan

Have you tried the "-Dsun.zip.disableMemoryMapping=true" JVM option to
rule out the possibility that zip/jar file access is causing the
trouble? There has been some bugs in the past in the JVM in that area:

http://javaeesupportpatterns.blogspot.com.es/2011/08/mmap-file-outofmemoryerror-and-pmap.html
. That has been fixed
http://bugs.java.com/bugdatabase/view_bug.do?bug_id=6280693 , but doing
a check with "-Dsun.zip.disableMemoryMapping=true" JVM option would be
interesting.
Mainly concerned about this commit:

https://github.com/apache/tomcat/commit/6e5420c67fbad81973d888ad3701a392fac4fc71

Since most commits weren't very interesting in this diff:
https://github.com/apache/tomcat/compare/075bc2d6...c0eb033f?w=1
Might make a difference to Jar file access. I'm not saying that this
commit is a problem, just seemed like a big change.

-Lari
_______________________________________________
cf-dev mailing list
cf-dev(a)lists.cloudfoundry.org
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev

I ran 3 tests over the weekend. These are the results:

- the copy of the application that ran against offline Java buildpack v3 /
tomcat 8.0.21 'crashed' 10 times over the weekend
- the copy of the application that ran against our custom Java buildpack /
tomcat 8.0.22 'crashed' 4 times over the weekend
- the copy of the application than ran against our custom Java buildpack /
tomcat 8.0.18 stayed up all weekend

We're certain this is down to the Tomcat version. Tomcat 8.0.22 doesn't seem
to be much better than 8.0.21 in respect of this memory issue.



--
View this message in context: http://cf-dev.70369.x6.nabble.com/Re-cf-dev-Cf-dev-vcap-dev-Java-OOM-debugging-tp25p112.html
Sent from the CF Dev mailing list archive at Nabble.com.

Thanks Chris.
We had given the MALLOC_ARENA_MAX setting a wide birth because Lari had
mentioned it was buggy on lucid64 & we're still on that stemcell right now.

We do have a test case running all weekend on Tomcat 8.0.22, which I'll try
to remember to post back to the list on.





--
View this message in context: http://cf-dev.70369.x6.nabble.com/Re-cf-dev-Cf-dev-vcap-dev-Java-OOM-debugging-tp25p91.html
Sent from the CF Dev mailing list archive at Nabble.com.

On 15-05-14 10:23 AM, Daniel Jones wrote:

Do you know how Tomcat's APR/NIO memory gets allocated? Is there a way
of telling from pmap whether pages are being used for NIO buffers or
by the APR?

The JVM has Native Memory Tracking (NMT), but I've never used it. Here
are some details about it:
http://hirt.se/blog/?p=401
http://docs.oracle.com/javase/8/docs/technotes/guides/vm/nmt-8.html
http://docs.oracle.com/javase/8/docs/technotes/guides/troubleshoot/tooldescr007.html#BABIIIAC
By using NMT, you can find out what memory regions are used by the JVM.
That would help finding out what areas of the pmap report aren't tracked
by NMT and NMT reports will give insight what the tracked memory is used
for.

I assume that won't track any allocations made in custom native
libraries like the Tomcat native library.
The native memory used by zip handling in the JVM doesn't seem to be
tracked either:
http://hg.openjdk.java.net/jdk8u/jdk8u/jdk/file/a006fa0a9e8f/src/share/native/java/util/zip/zip_util.c#l627
The "-Dsun.zip.disableMemoryMapping=true" setting changes how memory
allocation is done in the zip functions of the JDK. It doesn't get rid
of native memory allocations, but just makes it not use mmap. I'm not
sure if that leads to anything useful, but it's worth trying if you have
a good test environment.

Lari

On 15-05-14 10:23 AM, Daniel Jones wrote:

Thanks again for your input. Have you seen this problem with versions
of Tomcat before 8.0.20?

I don't have proper data gathered from older than 8.0.20, so I cannot
compare.
I was just wondering when did 8.0.20 become available in JBP, I found
this date:
HEAD https://download.run.pivotal.io/tomcat/tomcat-8.0.20.tar.gz | grep
Last-Modified
Last-Modified: Tue, 03 Mar 2015 11:35:19 GMT

David and I think we've narrowed down the issue to a change from using
Tomcat 8.0.18 to 8.0.21. We're running more tests and collaborating
with Pivotal support. We also noticed that non-prod versions of our
apps were taking longer to crash, so it would seem to be
activity-related at least.

Do you know how Tomcat's APR/NIO memory gets allocated? Is there a way
of telling from pmap whether pages are being used for NIO buffers or
by the APR?

I don't think you can get the info from pmap. The malloc_info xml shows
better allocation stats, but only stats.
Is Tomcat using APR library or NIO by default in CloudFoundry? I'd
assume that NIO isn't used by default.

Have you tried the "-Dsun.zip.disableMemoryMapping=true" JVM option to
rule out the possibility that zip/jar file access is causing the
trouble? There has been some bugs in the past in the JVM in that area:
http://javaeesupportpatterns.blogspot.com.es/2011/08/mmap-file-outofmemoryerror-and-pmap.html
. That has been fixed
http://bugs.java.com/bugdatabase/view_bug.do?bug_id=6280693 , but doing
a check with "-Dsun.zip.disableMemoryMapping=true" JVM option would be
interesting.
Mainly concerned about this commit:
https://github.com/apache/tomcat/commit/6e5420c67fbad81973d888ad3701a392fac4fc71

Since most commits weren't very interesting in this diff:
https://github.com/apache/tomcat/compare/075bc2d6...c0eb033f?w=1
Might make a difference to Jar file access. I'm not saying that this
commit is a problem, just seemed like a big change.

-Lari

Hi,

We've had a look through this and think it would be useful to give our
perspective for now.

To clarify the Memory heuristics, they are just weightings and not
percentages, they were never meant to add up to a hundred. The fact the
default settings add up to 105 is purely chance. Will see if the docs can
be improved around this.

The fact the JRE isn't detecting an OOM error and triggering the
killjava.sh script and because increasing the native memory prolongs the
OOM, we believe the leak is occurring outside the JRE. NIO, JNI, lots of
threads creation, etc... .There are other things running in the container
that could be consuming the memory.

If you suspect Tomcat then 8.0.22 is out and has been published for use in
the Java Buildpack, moving up to that might help things. We still haven't
seen a rash of memory problems from the wider Tomcat community for either
8.0.20 or 8.0.22 though so this would be unexpected.

It is actually possible to set the MALLOC_ARENA_MAX using an environment
variable.

cf set-env <APP> MALLOC_ARENA_MAX 2

This can also be specified in an applications manifest file.

Finally, a change has gone in to the master branch of the Java buildpack
that moves all of the memory heuristics code to an external program written
in Go. This means that scaling a Java application with cf scale no longer
requires a separate restage for the new memory setting to be applied as the
Go code will calculate them during every application start. During staging
you will also be able to see the memory settings that have been calculated
come out on the console. The plan is to release this new feature in version
3.1 of the buildpack in a week or so time.

We do still have an issue on our backlog to look at this and it shouldn't
be too long before we get to it.
https://www.pivotaltracker.com/story/show/94381284

Chris.

*On Mon, May 11, 2015 at 13:19 PM, Lari Hotari <Lari at hotari.net
<https://lists.cloudfoundry.org/mailman/listinfo/cf-dev>
**<mailto:Lari at hotari.net
<https://lists.cloudfoundry.org/mailman/listinfo/cf-dev>>> wrote:*

fyi. Tomcat 8.0.20 might be consuming more memory than
8.0.18:https://github.com/cloudfoundry/java-buildpack/issues/166#issuecomment-94517568

* Other things we’ve tried:

*>>* - We set verbose garbage collection to verify there was no
*>* memory size issues within the JVM. There wasn’t.
*>>* - We tried setting minimum memory for native, it had no
*>* effect. The container still gets killed
*>>* - We tried adjusting the ‘memory heuristics’ so that they
*>* added up to 80 rather than 100. This had the effect of causing a delay
*>* in the container being killed. However it still was killed.
*>I think adjusting memory heuristics so that they add up to 80 doesn't
make a difference because the values aren't percentages.
The values are proportional weighting values used in the memory
calculation:https://github.com/grails-samples/java-buildpack/blob/b4abf89/docs/jre-oracle_jre.md#memory-calculation

I found out that the only way to reserve "unused" memory is to set a
high value for the native memory lower bound in the memory_sizes.native
setting of config/open_jdk_jre.yml .
Example:https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25

>>* This seems like classic memory leak behaviour to me.

*>In my case it wasn't a classical Java memory leak, since the Java
application wasn't leaking memory. I was able to confirm this by getting
some heap dumps with the HeapDumpServlet
(https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy)
and analyzing them.

In my case the JVM's RSS memory size is slowly growing. It probably is
some kind of memory leak since one process I've been monitoring now is
very close to the memory limit. The uptime is now almost 3 weeks.

Here is the latest diff of the meminfo report.https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example2-txt

*From a Java perspective this isn't classical. The JVM heap isn't filling

*up. The problem is that RSS size is slowly growing and will eventually
cause the Java process to cross the memory boundary so that the process
gets kill by the Linux kernel cgroups OOM killer.

RSS size might be growing because of many reasons. I have been able to
slow down the growth by doing the various MALLOC_ and JVM parameter
tuning (-XX:MinMetaspaceExpansion=1M -XX:CodeCacheExpansionSize=1M). I'm
able to get a longer uptime, but the problem isn't solved.

Lari


On 15-05-11 06:41 AM, Head-Rapson, David wrote:

* Thanks for the continued advice.

*>> >>* We’ve hit on a key discovery after yet another a soak test this weekend.
*>>* - When we deploy using Tomcat 8.0.18 we don’t see the issue
*>>* - When we deploy using Tomcat 8.0.20 (same app version, same
*>* CF space, same services bound, same JBP code version, same JRE
*>* version, running at the same time), we see the crashes occurring after
*>* just a couple of hours.
*>> >>* Ideally we’d go ahead with the memory calculations you mentioned
*>* however we’re stuck on lucid64 because we’re using Pivotal CF 1.3.x &
*>* we’re having upgrade issues to 1.4.x.
*>>* So we’re not able to adjust MALLOC_ARENA_MAX, nor are we able to view
*>* RSS in pmap as you describe
*>> >>* Other things we’ve tried:
*>>* - We set verbose garbage collection to verify there was no
*>* memory size issues within the JVM. There wasn’t.
*>>* - We tried setting minimum memory for native, it had no
*>* effect. The container still gets killed
*>>* - We tried adjusting the ‘memory heuristics’ so that they
*>* added up to 80 rather than 100. This had the effect of causing a delay
*>* in the container being killed. However it still was killed.
*>> >>* This seems like classic memory leak behaviour to me.
*>> >>* *From:*Lari Hotari [mailto:lari.hotari at sagire.fi <https://lists.cloudfoundry.org/mailman/listinfo/cf-dev>] *On Behalf Of *Lari
*>* Hotari
*>* *Sent:* 08 May 2015 16:25
*>* *To:* Daniel Jones; Head-Rapson, David
*>* *Cc:* cf-dev at lists.cloudfoundry.org <https://lists.cloudfoundry.org/mailman/listinfo/cf-dev>
*>* *Subject:* Re: [Cf-dev] [vcap-dev] Java OOM debugging
*>> >>>* For my case, it turned out to be essential to reserve enough memory
*>* for "native" in the JBP. For the 2GB total memory, I set the minimum
*>* to 330M. With that setting I have been able to get over 2 weeks up
*>* time by now.
*>>* I mentioned this in my previous email:
*>>* The workaround for that in my case was to add a native key under
*>* memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is
*>* for a 2GB total memory).
*>* see example
*>* https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25 <https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25>
*>* that was how I got the app I'm running on CF to stay within the memory
*>* bounds. I'm sure there is now also a way to get the keys without
*>* forking the buildpack. I could have also adjusted the percentage
*>* portions, but I wanted to set a hard minimum for this case.
*>>>* I've been trying to get some insight by diffing the reports gathered
*>* from the meminfo servlet
*>* https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy <https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy>
*>>>* Here is such an example of a diff:
*>* https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example-txt <https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example-txt>
*>>* meminfo has pmap output included to get the report of the memory map
*>* of the process. I have just noticed that most of the memory has
*>* already been mmap:ed from the OS and it's just growing in RSS size.
*>* For example:
*>* < 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]
*>* > 00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]
*>>* The pmap output from lucid64 didn't include the RSS size, so you have
*>* to use cflinuxfs2 for this. It's also better because of other reasons.
*>* The glibc in lucid64 is old and has some bugs around the MALLOC_ARENA_MAX.
*>>* I was manually able to estimate the maximum size of the RSS size of
*>* what the Java process will consume by simply picking the large
*>* anon-blocks from the pmap report and calculating those blocks by the
*>* allocated virtual size (VSS).
*>* Based on this calculation, I picked the minimum of 330M for "native"
*>* in open_jdk_jre.yml as I mentioned before.
*>>* It looks like these rows are for the Heap size:
*>* < 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]
*>* > 00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]
*>>* It looks like the JVM doesn't fully allocate that block in RSS
*>* initially and most of the growth of RSS size comes from that in my
*>* case. In your case, it might be something different.
*>>* I also added a servlet for getting glibc malloc_info statistics in XML
*>* format (). I haven't really analysed that information because of time
*>* constraints and because I don't have a pressing problem any more. btw.
*>* The malloc_info XML report is missing some key elements, that has been
*>* added in later glibc versions
*>* (https://github.com/bminor/glibc/commit/4d653a59ffeae0f46f76a40230e2cfa9587b7e7e <https://github.com/bminor/glibc/commit/4d653a59ffeae0f46f76a40230e2cfa9587b7e7e>).
*>>* If killjava.sh never fires and the app crashed with Warden out of
*>* memory errors, then I believe it's the kernel's cgroups OOM killer
*>* that has killed the container processes. I have found this location
*>* where Warden oom notifier gets the OOM notification event:
*>* https://github.com/cloudfoundry/warden/blob/ad18bff/warden/lib/warden/container/features/mem_limit.rb#L70 <https://github.com/cloudfoundry/warden/blob/ad18bff/warden/lib/warden/container/features/mem_limit.rb#L70>
*>* This is the oom.c source code:
*>* https://github.com/cloudfoundry/warden/blob/ad18bff7dc56acbc55ff10bcc6045ebdf0b20c97/warden/src/oom/oom.c <https://github.com/cloudfoundry/warden/blob/ad18bff7dc56acbc55ff10bcc6045ebdf0b20c97/warden/src/oom/oom.c>
*>* . It reads the cgroups control files and receives events from the
*>* kernel that way.
*>>* I'd suggest that you use pmap for the Java process after it has
*>* started and calculate the maximum RSS size by calculating the VSS size
*>* of the large anon blocks instead of RSS for the blocks that the Java
*>* process has reserved for it's different memory areas (I think you
*>* shouldn't . You should discard adding VSS for the
*>* CompressedClassSpaceSize block.
*>* After this calculation, add enough memory to the "native" parameter in
*>* JBP until the RSS size calculated this way stays under the limit.
*>* That's the only "method" I have come up by now.
*>>* It might be required to have some RSS space allocated for any zip/jar
*>* files read by the Java process. I think that Java uses mmap files for
*>* zip file reading by default and that might go on top of all other limits.
*>* To test this theory, I'd suggest testing by adding
*>* -Dsun.zip.disableMemoryMapping=true system property setting to
*>* JAVA_OPTS. That disables the native mmap for zip/jar file reading. I
*>* haven't had time to test this assumption.
*>>* I guess the only way to understand how Java allocates memory is to
*>* look at the source code.
*>* from http://openjdk.java.net/projects/jdk8u/ <http://openjdk.java.net/projects/jdk8u/> , the instructions to get
*>* the source code of JDK 8:
*>* hg clone http://hg.openjdk.java.net/jdk8u/jdk8u;cd <http://hg.openjdk.java.net/jdk8u/jdk8u;cd> jdk8u;sh get_source.sh
*>* This tool is really good for grepping and searching the source code:
*>* http://geoff.greer.fm/ag/ <http://geoff.greer.fm/ag/> <http://geoff.greer.fm/ag/ <http://geoff.greer.fm/ag/>>
*>* On Ubuntu it's in silversearcher-ag package, "apt-get install
*>* silversearcher-ag" and on MacOSX brew it's "brew install
*>* the_silver_searcher".
*>* This alias is pretty useful:
*>* alias codegrep='ag --color --group --pager less -C 5'
*>* Then you just search for the correct location in code by starting with
*>* the tokens you know about:
*>* codegrep MaxMetaspaceSize
*>* this gives pretty good starting points in looking how the JDK
*>* allocates memory.
*>>* So the JDK source code is only a few commands away.
*>>* It would be interesting to hear more about this if someone has the
*>* time to dig in to this. This is about how far I got and I hope sharing
*>* this information helps someone continue. :)
*>>>* Lari
*>* github/twitter: lhotari
*>

Truncated



--
--
Christopher Frost - GoPivotal UK

Hi Lari,

Thanks again for your input. Have you seen this problem with versions of
Tomcat before 8.0.20?

David and I think we've narrowed down the issue to a change from using
Tomcat 8.0.18 to 8.0.21. We're running more tests and collaborating with
Pivotal support. We also noticed that non-prod versions of our apps were
taking longer to crash, so it would seem to be activity-related at least.

Do you know how Tomcat's APR/NIO memory gets allocated? Is there a way of
telling from pmap whether pages are being used for NIO buffers or by the
APR?

I wonder if the other folks that have reported CF out of memory errors with
later versions of Tomcat are seeing slow creeps in native memory
consumption?

--
Regards,

Daniel Jones
EngineerBetter.com

fyi. Tomcat 8.0.20 might be consuming more memory than 8.0.18:
https://github.com/cloudfoundry/java-buildpack/issues/166#issuecomment-94517568

Other things we’ve tried:

- We set verbose garbage collection to verify there was no
memory size issues within the JVM. There wasn’t.

- We tried setting minimum memory for native, it had no
effect. The container still gets killed

- We tried adjusting the ‘memory heuristics’ so that they
added up to 80 rather than 100. This had the effect of causing a delay
in the container being killed. However it still was killed.

I think adjusting memory heuristics so that they add up to 80 doesn't
make a difference because the values aren't percentages.
The values are proportional weighting values used in the memory
calculation:
https://github.com/grails-samples/java-buildpack/blob/b4abf89/docs/jre-oracle_jre.md#memory-calculation

I found out that the only way to reserve "unused" memory is to set a
high value for the native memory lower bound in the memory_sizes.native
setting of config/open_jdk_jre.yml .
Example:
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25

This seems like classic memory leak behaviour to me.

In my case it wasn't a classical Java memory leak, since the Java
application wasn't leaking memory. I was able to confirm this by getting
some heap dumps with the HeapDumpServlet
(https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy)
and analyzing them.

In my case the JVM's RSS memory size is slowly growing. It probably is
some kind of memory leak since one process I've been monitoring now is
very close to the memory limit. The uptime is now almost 3 weeks.

Here is the latest diff of the meminfo report.
https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example2-txt

From a Java perspective this isn't classical. The JVM heap isn't filling
up. The problem is that RSS size is slowly growing and will eventually
cause the Java process to cross the memory boundary so that the process
gets kill by the Linux kernel cgroups OOM killer.

RSS size might be growing because of many reasons. I have been able to
slow down the growth by doing the various MALLOC_ and JVM parameter
tuning (-XX:MinMetaspaceExpansion=1M -XX:CodeCacheExpansionSize=1M). I'm
able to get a longer uptime, but the problem isn't solved.

Lari


On 15-05-11 06:41 AM, Head-Rapson, David wrote:

Thanks for the continued advice.



We’ve hit on a key discovery after yet another a soak test this weekend.

- When we deploy using Tomcat 8.0.18 we don’t see the issue

- When we deploy using Tomcat 8.0.20 (same app version, same
CF space, same services bound, same JBP code version, same JRE
version, running at the same time), we see the crashes occurring after
just a couple of hours.



Ideally we’d go ahead with the memory calculations you mentioned
however we’re stuck on lucid64 because we’re using Pivotal CF 1.3.x &
we’re having upgrade issues to 1.4.x.

So we’re not able to adjust MALLOC_ARENA_MAX, nor are we able to view
RSS in pmap as you describe



Other things we’ve tried:

- We set verbose garbage collection to verify there was no
memory size issues within the JVM. There wasn’t.

- We tried setting minimum memory for native, it had no
effect. The container still gets killed

- We tried adjusting the ‘memory heuristics’ so that they
added up to 80 rather than 100. This had the effect of causing a delay
in the container being killed. However it still was killed.



This seems like classic memory leak behaviour to me.



*From:*Lari Hotari [mailto:lari.hotari(a)sagire.fi] *On Behalf Of *Lari
Hotari
*Sent:* 08 May 2015 16:25
*To:* Daniel Jones; Head-Rapson, David
*Cc:* cf-dev(a)lists.cloudfoundry.org
*Subject:* Re: [Cf-dev] [vcap-dev] Java OOM debugging




For my case, it turned out to be essential to reserve enough memory
for "native" in the JBP. For the 2GB total memory, I set the minimum
to 330M. With that setting I have been able to get over 2 weeks up
time by now.

I mentioned this in my previous email:

The workaround for that in my case was to add a native key under
memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is
for a 2GB total memory).
see example
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the memory
bounds. I'm sure there is now also a way to get the keys without
forking the buildpack. I could have also adjusted the percentage
portions, but I wanted to set a hard minimum for this case.


I've been trying to get some insight by diffing the reports gathered
from the meminfo servlet
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy


Here is such an example of a diff:
https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example-txt

meminfo has pmap output included to get the report of the memory map
of the process. I have just noticed that most of the memory has
already been mmap:ed from the OS and it's just growing in RSS size.
For example:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

The pmap output from lucid64 didn't include the RSS size, so you have
to use cflinuxfs2 for this. It's also better because of other reasons.
The glibc in lucid64 is old and has some bugs around the MALLOC_ARENA_MAX.

I was manually able to estimate the maximum size of the RSS size of
what the Java process will consume by simply picking the large
anon-blocks from the pmap report and calculating those blocks by the
allocated virtual size (VSS).
Based on this calculation, I picked the minimum of 330M for "native"
in open_jdk_jre.yml as I mentioned before.

It looks like these rows are for the Heap size:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

It looks like the JVM doesn't fully allocate that block in RSS
initially and most of the growth of RSS size comes from that in my
case. In your case, it might be something different.

I also added a servlet for getting glibc malloc_info statistics in XML
format (). I haven't really analysed that information because of time
constraints and because I don't have a pressing problem any more. btw.
The malloc_info XML report is missing some key elements, that has been
added in later glibc versions
(https://github.com/bminor/glibc/commit/4d653a59ffeae0f46f76a40230e2cfa9587b7e7e).

If killjava.sh never fires and the app crashed with Warden out of
memory errors, then I believe it's the kernel's cgroups OOM killer
that has killed the container processes. I have found this location
where Warden oom notifier gets the OOM notification event:
https://github.com/cloudfoundry/warden/blob/ad18bff/warden/lib/warden/container/features/mem_limit.rb#L70
This is the oom.c source code:
https://github.com/cloudfoundry/warden/blob/ad18bff7dc56acbc55ff10bcc6045ebdf0b20c97/warden/src/oom/oom.c
. It reads the cgroups control files and receives events from the
kernel that way.

I'd suggest that you use pmap for the Java process after it has
started and calculate the maximum RSS size by calculating the VSS size
of the large anon blocks instead of RSS for the blocks that the Java
process has reserved for it's different memory areas (I think you
shouldn't . You should discard adding VSS for the
CompressedClassSpaceSize block.
After this calculation, add enough memory to the "native" parameter in
JBP until the RSS size calculated this way stays under the limit.
That's the only "method" I have come up by now.

It might be required to have some RSS space allocated for any zip/jar
files read by the Java process. I think that Java uses mmap files for
zip file reading by default and that might go on top of all other limits.
To test this theory, I'd suggest testing by adding
-Dsun.zip.disableMemoryMapping=true system property setting to
JAVA_OPTS. That disables the native mmap for zip/jar file reading. I
haven't had time to test this assumption.

I guess the only way to understand how Java allocates memory is to
look at the source code.
from http://openjdk.java.net/projects/jdk8u/ , the instructions to get
the source code of JDK 8:
hg clone http://hg.openjdk.java.net/jdk8u/jdk8u;cd jdk8u;sh get_source.sh
This tool is really good for grepping and searching the source code:
http://geoff.greer.fm/ag/ <http://geoff.greer.fm/ag/>
On Ubuntu it's in silversearcher-ag package, "apt-get install
silversearcher-ag" and on MacOSX brew it's "brew install
the_silver_searcher".
This alias is pretty useful:
alias codegrep='ag --color --group --pager less -C 5'
Then you just search for the correct location in code by starting with
the tokens you know about:
codegrep MaxMetaspaceSize
this gives pretty good starting points in looking how the JDK
allocates memory.

So the JDK source code is only a few commands away.

It would be interesting to hear more about this if someone has the
time to dig in to this. This is about how far I got and I hope sharing
this information helps someone continue. :)


Lari
github/twitter: lhotari

On 15-05-08 10:02 AM, Daniel Jones wrote:

Hi Lari et al,



Thanks for your help Lari.



David and I are pairing on this issue, and we're yet to resolve
it. We're in the process of creating a repeatable test case (our
most crashy app makes calls to external services that need
mocking), but in the meantime, here's what we've seen.



Between Java Buildpack commit e89e546 and 17162df, we see apps
crashing with Warden out of memory errors. killjava.sh never
fires, and this has led us to believe that the kernel is shooting
a cgroup process in the head after the cgroup oversteps its memory
limit. We cannot find any evidence of the OOM killer firing in any
logs, but we may not be looking in the right place.



The JBP is setting heap to be 70%, metaspace to be 15% (with max
set to the same as initial), 5% for "stack", 5% for "normalised
stack" and 10% for "native". We do not understand why this adds up
to 105%, but haven't looked into the JBP algorithm yet. Any
pointers on what "normalised stack" is would be much appreciated,
as this doesn't appear in the list of heuristics supplied via app env.



Other team members tried applying the same settings that you
suggested - thanks for this. Apps still crash with these settings,
albeit less frequently.



After reading the blog you linked to
(http://java.dzone.com/articles/java-8-permgen-metaspace) we
wondered whether the increased /reserved /metaspace claimed after
metaspace GC might be causing a problem; however we reused the
test code to create a metaspace leak in a CF app and saw metaspace
GCs occur correctly, and memory usage never grow over
MaxMetaspaceSize. This figures, as the committed metaspace is
still less than MaxMetaspaceSize, and the reserved appears to be
whatever RAM is free across the whole DEA.



We noted that an Oracle blog
(https://blogs.oracle.com/poonam/entry/about_g1_garbage_collector_permanent)
mentions that the metaspace size parameters are approximate. We're
currently wondering if native allocations by Tomcat (APR, NIO) are
taking up more container memory, and so when the metaspace fills,
it's creeping slightly over the limit and triggering the kernel's
OOM killer.



Any suggestions would be much appreciated. We've tried to resist
tweaking heuristics blindly, but are running out of options as
we're struggling to figure out how the Java process is using
/committed/ memory. pmap seems to show virtual memory, and so it's
hard to see if things like the metaspace or NIO ByteBuffers are
nabbing too much and trigger the kernel's OOM killer.



Thanks for all your help,



Daniel Jones & David Head-Rapson



On Wed, Apr 29, 2015 at 8:07 PM, Lari Hotari <Lari(a)hotari.net
<mailto:Lari(a)hotari.net>> wrote:

Hi,

I created a few tools to debug OOM problems since the application
I was responsible for running on CF was failing constantly because
of OOM problems. The problems I had, turned out not to be actual
memory leaks in the Java application.

In the "cf events appname" log I would get entries like this:
2015-xx-xxTxx:xx:xx.00-0400 app.crash appname
index: 1, reason: CRASHED, exit_description: out of memory,
exit_status: 255

These type of entries are produced when the container goes over
it's memory resource limits. It doesn't mean that there is a
memory leak in the Java application. The container gets killed by
the Linux kernel oom killer
(https://github.com/cloudfoundry/warden/blob/master/warden/README.md#limit-handle-mem-value)
based on the resource limits set to the warden container.

The memory limit is specified in number of bytes. It is enforced
using the control group associated with the container. When a
container exceeds this limit, one or more of its processes will be
killed by the kernel. Additionally, the Warden will be notified
that an OOM happened and it subsequently tears down the container.

In my case it never got killed by the killjava.sh script that gets
called in the java-buildpack when an OOM happens in Java.

This is the tool I built to debug the problems:
https://github.com/lhotari/java-buildpack-diagnostics-app
I deployed that app as part of the forked buildpack I'm using.
Please read the readme about what it's limitations are. It worked
for me, but it might not work for you. It's opensource and you can
fork it. :)

There is a solution in my toolcase for creating a heapdump and
uploading that to S3:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy
The readme explains how to setup Amazon S3 keys for this:
https://github.com/lhotari/java-buildpack-diagnostics-app#amazon-s3-setup
Once you get a dump, you can then analyse the dump in a java
profiler tool like YourKit.

I also have a solution that forks the java-buildpack modifies
killjava.sh and adds a script that uploads the heapdump to S3 in
the case of OOM:
https://github.com/lhotari/java-buildpack/commit/2d654b80f3bf1a0e0f1bae4f29cb85f56f5f8c46

In java-buildpack-diagnostics-app I have also other tools for
getting Linux operation system specific memory information, for
example:

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemorySmapServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MallocInfoServlet.groovy

These tools are handy for looking at details of the Java process
RSS memory usage growth.

There is also a solution for getting ssh shell access inside your
application with tmate.io <http://tmate.io>:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/TmateSshServlet.groovy
(this version is only compatible with the new "cflinuxfs2" stack)

It looks like there are serious problems on CloudFoundry with the
memory sizing calculation. An application that doesn't have a OOM
problem will get killed by the oom killer because the Java process
will go over the memory limits.
I filed this issue:
https://github.com/cloudfoundry/java-buildpack/issues/157 , but
that might not cover everything.

The workaround for that in my case was to add a native key under
memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that
is for a 2GB total memory).
see example
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the
memory bounds. I'm sure there is now also a way to get the keys
without forking the buildpack. I could have also adjusted the
percentage portions, but I wanted to set a hard minimum for this case.

It was also required to do some other tuning.

I added this to JAVA_OPTS:
-XX:CompressedClassSpaceSize=256M -XX:InitialCodeCacheSize=64M
-XX:CodeCacheExpansionSize=1M -XX:CodeCacheMinimumFreeSpace=1M
-XX:ReservedCodeCacheSize=200M -XX:MinMetaspaceExpansion=1M
-XX:MaxMetaspaceExpansion=8M -XX:MaxDirectMemorySize=96M
while trying to keep the Java process from growing in RSS memory size.

The memory overhead of a 64 bit Java process on Linux can be
reduced by specifying these environment variables:

stack: cflinuxfs2
.
.
.
env:
MALLOC_ARENA_MAX: 2
MALLOC_MMAP_THRESHOLD_: 131072
MALLOC_TRIM_THRESHOLD_: 131072
MALLOC_TOP_PAD_: 131072
MALLOC_MMAP_MAX_: 65536

MALLOC_ARENA_MAX works only on cflinuxfs2 stack (the lucid64 stack
has a buggy version of glibc).

explanation about MALLOC_ARENA_MAX from Heroku:
https://devcenter.heroku.com/articles/tuning-glibc-memory-behavior
some measurement data how it reduces memory consumption:
https://devcenter.heroku.com/articles/testing-cedar-14-memory-use

I have created a PR to add this to CF java-buildpack:
https://github.com/cloudfoundry/java-buildpack/pull/160

I also created an issues
https://github.com/cloudfoundry/java-buildpack/issues/163 and
https://github.com/cloudfoundry/java-buildpack/pull/159 .

I hope this information helps others struggling with OOM problems
in CF.
I'm not saying that this is a ready made solution just for you.
YMMV. It worked for me.

-Lari




On 15-04-29 10:53 AM, Head-Rapson, David wrote:

Hi,

I’m after some guidance on how to get profile Java apps in CF,
in order to get to the bottom of memory issues.

We have an app that’s crashing every few hours with OOM error,
most likely it’s a memory leak.

I’d like to profile the JVM and work out what’s eating memory,
however tools like yourkit require connectivity INTO the JVM
server (i.e. the warden container), either via host / port or
via SSH.

Since warden containers cannot be connected to on ports other
than for HTTP and cannot be SSHd to, neither of these works
for me.



I tried installed a standalone JDK onto the warden container,
however as soon as I ran ‘jmap’ to invoke the dump, warden
cleaned up the container – most likely for memory
over-consumption.



I had previously found a hack in the Weblogic buildpack
(https://github.com/pivotal-cf/weblogic-buildpack/blob/master/docs/container-wls-monitoring.md)
for modifying the start script which, when used with
–XX:HeapDumpOnOutOfMemoryError, should copy any heapdump files
to a file share somewhere. I have my own custom buildpack so
I could use something similar.

Has anyone got a better solution than this?



We would love to use newrelic / app dynamics for this however
we’re not allowed. And I’m not 100% certain they could help
with this either.



Dave



The information transmitted is intended for the person or
entity to which it is addressed and may contain confidential,
privileged or copyrighted material. If you receive this in
error, please contact the sender and delete the material from
any computer. Fidelity only gives information on products and
services and does not give investment advice to retail clients
based on individual circumstances. Any comments or statements
made are not necessarily those of Fidelity. All e-mails may be
monitored. FIL Investments International (Reg. No.1448245),
FIL Investment Services (UK) Limited (Reg. No. 2016555), FIL
Pensions Management (Reg. No. 2015142) and Financial
Administration Services Limited (Reg. No. 1629709) are
authorised and regulated in the UK by the Financial Conduct
Authority. FIL Life Insurance Limited (Reg No. 3406905) is
authorised in the UK by the Prudential Regulation Authority
and regulated in the UK by the Financial Conduct Authority and
the Prudential Regulation Authority. Registered offices at
Oakhill House, 130 Tonbridge Road, Hildenborough, Tonbridge,
Kent TN11 9DZ.

--
You received this message because you are subscribed to the
Google Groups "Cloud Foundry Developers" group.
To view this discussion on the web visit
https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com
<https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com?utm_medium=email&utm_source=footer>.
To unsubscribe from this group and stop receiving emails from
it, send an email to vcap-dev+unsubscribe(a)cloudfoundry.org
<mailto:vcap-dev+unsubscribe(a)cloudfoundry.org>.




_______________________________________________
Cf-dev mailing list
Cf-dev(a)lists.cloudfoundry.org <mailto:Cf-dev(a)lists.cloudfoundry.org>
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev





--

Regards,



Daniel Jones

EngineerBetter.com

Thanks for the continued advice.

We’ve hit on a key discovery after yet another a soak test this weekend.

- When we deploy using Tomcat 8.0.18 we don’t see the issue

- When we deploy using Tomcat 8.0.20 (same app version, same CF space, same services bound, same JBP code version, same JRE version, running at the same time), we see the crashes occurring after just a couple of hours.

Ideally we’d go ahead with the memory calculations you mentioned however we’re stuck on lucid64 because we’re using Pivotal CF 1.3.x & we’re having upgrade issues to 1.4.x.
So we’re not able to adjust MALLOC_ARENA_MAX, nor are we able to view RSS in pmap as you describe

Other things we’ve tried:

- We set verbose garbage collection to verify there was no memory size issues within the JVM. There wasn’t.

- We tried setting minimum memory for native, it had no effect. The container still gets killed

- We tried adjusting the ‘memory heuristics’ so that they added up to 80 rather than 100. This had the effect of causing a delay in the container being killed. However it still was killed.

This seems like classic memory leak behaviour to me.

From: Lari Hotari [mailto:lari.hotari(a)sagire.fi] On Behalf Of Lari Hotari
Sent: 08 May 2015 16:25
To: Daniel Jones; Head-Rapson, David
Cc: cf-dev(a)lists.cloudfoundry.org
Subject: Re: [Cf-dev] [vcap-dev] Java OOM debugging


For my case, it turned out to be essential to reserve enough memory for "native" in the JBP. For the 2GB total memory, I set the minimum to 330M. With that setting I have been able to get over 2 weeks up time by now.

I mentioned this in my previous email:

The workaround for that in my case was to add a native key under memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is for a 2GB total memory).
see example https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the memory bounds. I'm sure there is now also a way to get the keys without forking the buildpack. I could have also adjusted the percentage portions, but I wanted to set a hard minimum for this case.

I've been trying to get some insight by diffing the reports gathered from the meminfo servlet https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy

Here is such an example of a diff:
https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example-txt

meminfo has pmap output included to get the report of the memory map of the process. I have just noticed that most of the memory has already been mmap:ed from the OS and it's just growing in RSS size. For example:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

The pmap output from lucid64 didn't include the RSS size, so you have to use cflinuxfs2 for this. It's also better because of other reasons. The glibc in lucid64 is old and has some bugs around the MALLOC_ARENA_MAX.

I was manually able to estimate the maximum size of the RSS size of what the Java process will consume by simply picking the large anon-blocks from the pmap report and calculating those blocks by the allocated virtual size (VSS).
Based on this calculation, I picked the minimum of 330M for "native" in open_jdk_jre.yml as I mentioned before.

It looks like these rows are for the Heap size:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

It looks like the JVM doesn't fully allocate that block in RSS initially and most of the growth of RSS size comes from that in my case. In your case, it might be something different.

I also added a servlet for getting glibc malloc_info statistics in XML format (). I haven't really analysed that information because of time constraints and because I don't have a pressing problem any more. btw. The malloc_info XML report is missing some key elements, that has been added in later glibc versions (https://github.com/bminor/glibc/commit/4d653a59ffeae0f46f76a40230e2cfa9587b7e7e).

If killjava.sh never fires and the app crashed with Warden out of memory errors, then I believe it's the kernel's cgroups OOM killer that has killed the container processes. I have found this location where Warden oom notifier gets the OOM notification event:
https://github.com/cloudfoundry/warden/blob/ad18bff/warden/lib/warden/container/features/mem_limit.rb#L70
This is the oom.c source code: https://github.com/cloudfoundry/warden/blob/ad18bff7dc56acbc55ff10bcc6045ebdf0b20c97/warden/src/oom/oom.c . It reads the cgroups control files and receives events from the kernel that way.

I'd suggest that you use pmap for the Java process after it has started and calculate the maximum RSS size by calculating the VSS size of the large anon blocks instead of RSS for the blocks that the Java process has reserved for it's different memory areas (I think you shouldn't . You should discard adding VSS for the CompressedClassSpaceSize block.
After this calculation, add enough memory to the "native" parameter in JBP until the RSS size calculated this way stays under the limit.
That's the only "method" I have come up by now.

It might be required to have some RSS space allocated for any zip/jar files read by the Java process. I think that Java uses mmap files for zip file reading by default and that might go on top of all other limits.
To test this theory, I'd suggest testing by adding -Dsun.zip.disableMemoryMapping=true system property setting to JAVA_OPTS. That disables the native mmap for zip/jar file reading. I haven't had time to test this assumption.

I guess the only way to understand how Java allocates memory is to look at the source code.
from http://openjdk.java.net/projects/jdk8u/ , the instructions to get the source code of JDK 8:
hg clone http://hg.openjdk.java.net/jdk8u/jdk8u;cd jdk8u;sh get_source.sh
This tool is really good for grepping and searching the source code: http://geoff.greer.fm/ag/
On Ubuntu it's in silversearcher-ag package, "apt-get install silversearcher-ag" and on MacOSX brew it's "brew install the_silver_searcher".
This alias is pretty useful:
alias codegrep='ag --color --group --pager less -C 5'
Then you just search for the correct location in code by starting with the tokens you know about:
codegrep MaxMetaspaceSize
this gives pretty good starting points in looking how the JDK allocates memory.

So the JDK source code is only a few commands away.

It would be interesting to hear more about this if someone has the time to dig in to this. This is about how far I got and I hope sharing this information helps someone continue. :)


Lari
github/twitter: lhotari
On 15-05-08 10:02 AM, Daniel Jones wrote:
Hi Lari et al,

Thanks for your help Lari.

David and I are pairing on this issue, and we're yet to resolve it. We're in the process of creating a repeatable test case (our most crashy app makes calls to external services that need mocking), but in the meantime, here's what we've seen.

Between Java Buildpack commit e89e546 and 17162df, we see apps crashing with Warden out of memory errors. killjava.sh never fires, and this has led us to believe that the kernel is shooting a cgroup process in the head after the cgroup oversteps its memory limit. We cannot find any evidence of the OOM killer firing in any logs, but we may not be looking in the right place.

The JBP is setting heap to be 70%, metaspace to be 15% (with max set to the same as initial), 5% for "stack", 5% for "normalised stack" and 10% for "native". We do not understand why this adds up to 105%, but haven't looked into the JBP algorithm yet. Any pointers on what "normalised stack" is would be much appreciated, as this doesn't appear in the list of heuristics supplied via app env.

Other team members tried applying the same settings that you suggested - thanks for this. Apps still crash with these settings, albeit less frequently.

After reading the blog you linked to (http://java.dzone.com/articles/java-8-permgen-metaspace) we wondered whether the increased reserved metaspace claimed after metaspace GC might be causing a problem; however we reused the test code to create a metaspace leak in a CF app and saw metaspace GCs occur correctly, and memory usage never grow over MaxMetaspaceSize. This figures, as the committed metaspace is still less than MaxMetaspaceSize, and the reserved appears to be whatever RAM is free across the whole DEA.

We noted that an Oracle blog (https://blogs.oracle.com/poonam/entry/about_g1_garbage_collector_permanent) mentions that the metaspace size parameters are approximate. We're currently wondering if native allocations by Tomcat (APR, NIO) are taking up more container memory, and so when the metaspace fills, it's creeping slightly over the limit and triggering the kernel's OOM killer.

Any suggestions would be much appreciated. We've tried to resist tweaking heuristics blindly, but are running out of options as we're struggling to figure out how the Java process is using committed memory. pmap seems to show virtual memory, and so it's hard to see if things like the metaspace or NIO ByteBuffers are nabbing too much and trigger the kernel's OOM killer.

Thanks for all your help,

Daniel Jones & David Head-Rapson

On Wed, Apr 29, 2015 at 8:07 PM, Lari Hotari <Lari(a)hotari.net<mailto:Lari(a)hotari.net>> wrote:
Hi,

I created a few tools to debug OOM problems since the application I was responsible for running on CF was failing constantly because of OOM problems. The problems I had, turned out not to be actual memory leaks in the Java application.

In the "cf events appname" log I would get entries like this:
2015-xx-xxTxx:xx:xx.00-0400 app.crash appname index: 1, reason: CRASHED, exit_description: out of memory, exit_status: 255

These type of entries are produced when the container goes over it's memory resource limits. It doesn't mean that there is a memory leak in the Java application. The container gets killed by the Linux kernel oom killer (https://github.com/cloudfoundry/warden/blob/master/warden/README.md#limit-handle-mem-value) based on the resource limits set to the warden container.

The memory limit is specified in number of bytes. It is enforced using the control group associated with the container. When a container exceeds this limit, one or more of its processes will be killed by the kernel. Additionally, the Warden will be notified that an OOM happened and it subsequently tears down the container.
In my case it never got killed by the killjava.sh script that gets called in the java-buildpack when an OOM happens in Java.

This is the tool I built to debug the problems:
https://github.com/lhotari/java-buildpack-diagnostics-app
I deployed that app as part of the forked buildpack I'm using.
Please read the readme about what it's limitations are. It worked for me, but it might not work for you. It's opensource and you can fork it. :)

There is a solution in my toolcase for creating a heapdump and uploading that to S3:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy
The readme explains how to setup Amazon S3 keys for this: https://github.com/lhotari/java-buildpack-diagnostics-app#amazon-s3-setup
Once you get a dump, you can then analyse the dump in a java profiler tool like YourKit.

I also have a solution that forks the java-buildpack modifies killjava.sh and adds a script that uploads the heapdump to S3 in the case of OOM:
https://github.com/lhotari/java-buildpack/commit/2d654b80f3bf1a0e0f1bae4f29cb85f56f5f8c46

In java-buildpack-diagnostics-app I have also other tools for getting Linux operation system specific memory information, for example:

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemorySmapServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MallocInfoServlet.groovy

These tools are handy for looking at details of the Java process RSS memory usage growth.

There is also a solution for getting ssh shell access inside your application with tmate.io<http://tmate.io>:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/TmateSshServlet.groovy (this version is only compatible with the new "cflinuxfs2" stack)

It looks like there are serious problems on CloudFoundry with the memory sizing calculation. An application that doesn't have a OOM problem will get killed by the oom killer because the Java process will go over the memory limits.
I filed this issue: https://github.com/cloudfoundry/java-buildpack/issues/157 , but that might not cover everything.

The workaround for that in my case was to add a native key under memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is for a 2GB total memory).
see example https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the memory bounds. I'm sure there is now also a way to get the keys without forking the buildpack. I could have also adjusted the percentage portions, but I wanted to set a hard minimum for this case.

It was also required to do some other tuning.

I added this to JAVA_OPTS:
-XX:CompressedClassSpaceSize=256M -XX:InitialCodeCacheSize=64M -XX:CodeCacheExpansionSize=1M -XX:CodeCacheMinimumFreeSpace=1M -XX:ReservedCodeCacheSize=200M -XX:MinMetaspaceExpansion=1M -XX:MaxMetaspaceExpansion=8M -XX:MaxDirectMemorySize=96M
while trying to keep the Java process from growing in RSS memory size.

The memory overhead of a 64 bit Java process on Linux can be reduced by specifying these environment variables:

stack: cflinuxfs2
.
.
.
env:
MALLOC_ARENA_MAX: 2
MALLOC_MMAP_THRESHOLD_: 131072
MALLOC_TRIM_THRESHOLD_: 131072
MALLOC_TOP_PAD_: 131072
MALLOC_MMAP_MAX_: 65536

MALLOC_ARENA_MAX works only on cflinuxfs2 stack (the lucid64 stack has a buggy version of glibc).

explanation about MALLOC_ARENA_MAX from Heroku:
https://devcenter.heroku.com/articles/tuning-glibc-memory-behavior
some measurement data how it reduces memory consumption: https://devcenter.heroku.com/articles/testing-cedar-14-memory-use

I have created a PR to add this to CF java-buildpack:
https://github.com/cloudfoundry/java-buildpack/pull/160

I also created an issues https://github.com/cloudfoundry/java-buildpack/issues/163 and https://github.com/cloudfoundry/java-buildpack/pull/159 .

I hope this information helps others struggling with OOM problems in CF.
I'm not saying that this is a ready made solution just for you. YMMV. It worked for me.

-Lari



On 15-04-29 10:53 AM, Head-Rapson, David wrote:
Hi,
I’m after some guidance on how to get profile Java apps in CF, in order to get to the bottom of memory issues.
We have an app that’s crashing every few hours with OOM error, most likely it’s a memory leak.
I’d like to profile the JVM and work out what’s eating memory, however tools like yourkit require connectivity INTO the JVM server (i.e. the warden container), either via host / port or via SSH.
Since warden containers cannot be connected to on ports other than for HTTP and cannot be SSHd to, neither of these works for me.

I tried installed a standalone JDK onto the warden container, however as soon as I ran ‘jmap’ to invoke the dump, warden cleaned up the container – most likely for memory over-consumption.

I had previously found a hack in the Weblogic buildpack (https://github.com/pivotal-cf/weblogic-buildpack/blob/master/docs/container-wls-monitoring.md) for modifying the start script which, when used with –XX:HeapDumpOnOutOfMemoryError, should copy any heapdump files to a file share somewhere. I have my own custom buildpack so I could use something similar.
Has anyone got a better solution than this?

We would love to use newrelic / app dynamics for this however we’re not allowed. And I’m not 100% certain they could help with this either.

Dave

The information transmitted is intended for the person or entity to which it is addressed and may contain confidential, privileged or copyrighted material. If you receive this in error, please contact the sender and delete the material from any computer. Fidelity only gives information on products and services and does not give investment advice to retail clients based on individual circumstances. Any comments or statements made are not necessarily those of Fidelity. All e-mails may be monitored. FIL Investments International (Reg. No.1448245), FIL Investment Services (UK) Limited (Reg. No. 2016555), FIL Pensions Management (Reg. No. 2015142) and Financial Administration Services Limited (Reg. No. 1629709) are authorised and regulated in the UK by the Financial Conduct Authority. FIL Life Insurance Limited (Reg No. 3406905) is authorised in the UK by the Prudential Regulation Authority and regulated in the UK by the Financial Conduct Authority and the Prudential Regulation Authority. Registered offices at Oakhill House, 130 Tonbridge Road, Hildenborough, Tonbridge, Kent TN11 9DZ.
--
You received this message because you are subscribed to the Google Groups "Cloud Foundry Developers" group.
To view this discussion on the web visit https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com<https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com?utm_medium=email&utm_source=footer>.
To unsubscribe from this group and stop receiving emails from it, send an email to vcap-dev+unsubscribe(a)cloudfoundry.org<mailto:vcap-dev+unsubscribe(a)cloudfoundry.org>.


_______________________________________________
Cf-dev mailing list
Cf-dev(a)lists.cloudfoundry.org<mailto:Cf-dev(a)lists.cloudfoundry.org>
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev



--
Regards,

Daniel Jones
EngineerBetter.com

For my case, it turned out to be essential to reserve enough memory for
"native" in the JBP. For the 2GB total memory, I set the minimum to
330M. With that setting I have been able to get over 2 weeks up time by
now.

I mentioned this in my previous email:

The workaround for that in my case was to add a native key under
memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is
for a 2GB total memory).
see example
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the memory
bounds. I'm sure there is now also a way to get the keys without
forking the buildpack. I could have also adjusted the percentage
portions, but I wanted to set a hard minimum for this case.

I've been trying to get some insight by diffing the reports gathered
from the meminfo servlet
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy


Here is such an example of a diff:
https://gist.github.com/lhotari/ee77decc2585f56cf3ad#file-meminfo_diff_example-txt

meminfo has pmap output included to get the report of the memory map of
the process. I have just noticed that most of the memory has already
been mmap:ed from the OS and it's just growing in RSS size. For example:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

The pmap output from lucid64 didn't include the RSS size, so you have to
use cflinuxfs2 for this. It's also better because of other reasons. The
glibc in lucid64 is old and has some bugs around the MALLOC_ARENA_MAX.

I was manually able to estimate the maximum size of the RSS size of what
the Java process will consume by simply picking the large anon-blocks
from the pmap report and calculating those blocks by the allocated
virtual size (VSS).
Based on this calculation, I picked the minimum of 330M for "native" in
open_jdk_jre.yml as I mentioned before.

It looks like these rows are for the Heap size:
< 00000000a7600000 1471488 1469556 1469556 rw--- [ anon ]

00000000a7600000 1471744 1470444 1470444 rw--- [ anon ]

It looks like the JVM doesn't fully allocate that block in RSS initially
and most of the growth of RSS size comes from that in my case. In your
case, it might be something different.

I also added a servlet for getting glibc malloc_info statistics in XML
format (). I haven't really analysed that information because of time
constraints and because I don't have a pressing problem any more. btw.
The malloc_info XML report is missing some key elements, that has been
added in later glibc versions
(https://github.com/bminor/glibc/commit/4d653a59ffeae0f46f76a40230e2cfa9587b7e7e).

If killjava.sh never fires and the app crashed with Warden out of memory
errors, then I believe it's the kernel's cgroups OOM killer that has
killed the container processes. I have found this location where Warden
oom notifier gets the OOM notification event:
https://github.com/cloudfoundry/warden/blob/ad18bff/warden/lib/warden/container/features/mem_limit.rb#L70
This is the oom.c source code:
https://github.com/cloudfoundry/warden/blob/ad18bff7dc56acbc55ff10bcc6045ebdf0b20c97/warden/src/oom/oom.c
. It reads the cgroups control files and receives events from the kernel
that way.

I'd suggest that you use pmap for the Java process after it has started
and calculate the maximum RSS size by calculating the VSS size of the
large anon blocks instead of RSS for the blocks that the Java process
has reserved for it's different memory areas (I think you shouldn't .
You should discard adding VSS for the CompressedClassSpaceSize block.
After this calculation, add enough memory to the "native" parameter in
JBP until the RSS size calculated this way stays under the limit.
That's the only "method" I have come up by now.

It might be required to have some RSS space allocated for any zip/jar
files read by the Java process. I think that Java uses mmap files for
zip file reading by default and that might go on top of all other limits.
To test this theory, I'd suggest testing by adding
-Dsun.zip.disableMemoryMapping=true system property setting to
JAVA_OPTS. That disables the native mmap for zip/jar file reading. I
haven't had time to test this assumption.

I guess the only way to understand how Java allocates memory is to look
at the source code.
from http://openjdk.java.net/projects/jdk8u/ , the instructions to get
the source code of JDK 8:
hg clone http://hg.openjdk.java.net/jdk8u/jdk8u;cd jdk8u;sh get_source.sh
This tool is really good for grepping and searching the source code:
http://geoff.greer.fm/ag/
On Ubuntu it's in silversearcher-ag package, "apt-get install
silversearcher-ag" and on MacOSX brew it's "brew install
the_silver_searcher".
This alias is pretty useful:
alias codegrep='ag --color --group --pager less -C 5'
Then you just search for the correct location in code by starting with
the tokens you know about:
codegrep MaxMetaspaceSize
this gives pretty good starting points in looking how the JDK allocates
memory.

So the JDK source code is only a few commands away.

It would be interesting to hear more about this if someone has the time
to dig in to this. This is about how far I got and I hope sharing this
information helps someone continue. :)


Lari
github/twitter: lhotari

On 15-05-08 10:02 AM, Daniel Jones wrote:

Hi Lari et al,

Thanks for your help Lari.

David and I are pairing on this issue, and we're yet to resolve it.
We're in the process of creating a repeatable test case (our most
crashy app makes calls to external services that need mocking), but in
the meantime, here's what we've seen.

Between Java Buildpack commit e89e546 and 17162df, we see apps
crashing with Warden out of memory errors. killjava.sh never fires,
and this has led us to believe that the kernel is shooting a cgroup
process in the head after the cgroup oversteps its memory limit. We
cannot find any evidence of the OOM killer firing in any logs, but we
may not be looking in the right place.

The JBP is setting heap to be 70%, metaspace to be 15% (with max set
to the same as initial), 5% for "stack", 5% for "normalised stack" and
10% for "native". We do not understand why this adds up to 105%, but
haven't looked into the JBP algorithm yet. Any pointers on what
"normalised stack" is would be much appreciated, as this doesn't
appear in the list of heuristics supplied via app env.

Other team members tried applying the same settings that you suggested
- thanks for this. Apps still crash with these settings, albeit less
frequently.

After reading the blog you linked to
(http://java.dzone.com/articles/java-8-permgen-metaspace) we wondered
whether the increased /reserved /metaspace claimed after metaspace GC
might be causing a problem; however we reused the test code to create
a metaspace leak in a CF app and saw metaspace GCs occur correctly,
and memory usage never grow over MaxMetaspaceSize. This figures, as
the committed metaspace is still less than MaxMetaspaceSize, and the
reserved appears to be whatever RAM is free across the whole DEA.

We noted that an Oracle blog
(https://blogs.oracle.com/poonam/entry/about_g1_garbage_collector_permanent)
mentions that the metaspace size parameters are approximate. We're
currently wondering if native allocations by Tomcat (APR, NIO) are
taking up more container memory, and so when the metaspace fills, it's
creeping slightly over the limit and triggering the kernel's OOM killer.

Any suggestions would be much appreciated. We've tried to resist
tweaking heuristics blindly, but are running out of options as we're
struggling to figure out how the Java process is using
/committed/ memory. pmap seems to show virtual memory, and so it's
hard to see if things like the metaspace or NIO ByteBuffers are
nabbing too much and trigger the kernel's OOM killer.

Thanks for all your help,

Daniel Jones & David Head-Rapson

On Wed, Apr 29, 2015 at 8:07 PM, Lari Hotari <Lari(a)hotari.net
<mailto:Lari(a)hotari.net>> wrote:

Hi,

I created a few tools to debug OOM problems since the application
I was responsible for running on CF was failing constantly because
of OOM problems. The problems I had, turned out not to be actual
memory leaks in the Java application.

In the "cf events appname" log I would get entries like this:
2015-xx-xxTxx:xx:xx.00-0400 app.crash appname
index: 1, reason: CRASHED, exit_description: out of memory,
exit_status: 255

These type of entries are produced when the container goes over
it's memory resource limits. It doesn't mean that there is a
memory leak in the Java application. The container gets killed by
the Linux kernel oom killer
(https://github.com/cloudfoundry/warden/blob/master/warden/README.md#limit-handle-mem-value)
based on the resource limits set to the warden container.

The memory limit is specified in number of bytes. It is enforced
using the control group associated with the container. When a
container exceeds this limit, one or more of its processes will
be killed by the kernel. Additionally, the Warden will be
notified that an OOM happened and it subsequently tears down the
container.

In my case it never got killed by the killjava.sh script that gets
called in the java-buildpack when an OOM happens in Java.

This is the tool I built to debug the problems:
https://github.com/lhotari/java-buildpack-diagnostics-app
I deployed that app as part of the forked buildpack I'm using.
Please read the readme about what it's limitations are. It worked
for me, but it might not work for you. It's opensource and you can
fork it. :)

There is a solution in my toolcase for creating a heapdump and
uploading that to S3:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy
The readme explains how to setup Amazon S3 keys for this:
https://github.com/lhotari/java-buildpack-diagnostics-app#amazon-s3-setup
Once you get a dump, you can then analyse the dump in a java
profiler tool like YourKit.

I also have a solution that forks the java-buildpack modifies
killjava.sh and adds a script that uploads the heapdump to S3 in
the case of OOM:
https://github.com/lhotari/java-buildpack/commit/2d654b80f3bf1a0e0f1bae4f29cb85f56f5f8c46

In java-buildpack-diagnostics-app I have also other tools for
getting Linux operation system specific memory information, for
example:

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemorySmapServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MallocInfoServlet.groovy

These tools are handy for looking at details of the Java process
RSS memory usage growth.

There is also a solution for getting ssh shell access inside your
application with tmate.io <http://tmate.io>:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/TmateSshServlet.groovy
(this version is only compatible with the new "cflinuxfs2" stack)

It looks like there are serious problems on CloudFoundry with the
memory sizing calculation. An application that doesn't have a OOM
problem will get killed by the oom killer because the Java process
will go over the memory limits.
I filed this issue:
https://github.com/cloudfoundry/java-buildpack/issues/157 , but
that might not cover everything.

The workaround for that in my case was to add a native key under
memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that
is for a 2GB total memory).
see example
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the
memory bounds. I'm sure there is now also a way to get the keys
without forking the buildpack. I could have also adjusted the
percentage portions, but I wanted to set a hard minimum for this case.

It was also required to do some other tuning.

I added this to JAVA_OPTS:
-XX:CompressedClassSpaceSize=256M -XX:InitialCodeCacheSize=64M
-XX:CodeCacheExpansionSize=1M -XX:CodeCacheMinimumFreeSpace=1M
-XX:ReservedCodeCacheSize=200M -XX:MinMetaspaceExpansion=1M
-XX:MaxMetaspaceExpansion=8M -XX:MaxDirectMemorySize=96M
while trying to keep the Java process from growing in RSS memory size.

The memory overhead of a 64 bit Java process on Linux can be
reduced by specifying these environment variables:

stack: cflinuxfs2
.
.
.
env:
MALLOC_ARENA_MAX: 2
MALLOC_MMAP_THRESHOLD_: 131072
MALLOC_TRIM_THRESHOLD_: 131072
MALLOC_TOP_PAD_: 131072
MALLOC_MMAP_MAX_: 65536

MALLOC_ARENA_MAX works only on cflinuxfs2 stack (the lucid64 stack
has a buggy version of glibc).

explanation about MALLOC_ARENA_MAX from Heroku:
https://devcenter.heroku.com/articles/tuning-glibc-memory-behavior
some measurement data how it reduces memory consumption:
https://devcenter.heroku.com/articles/testing-cedar-14-memory-use

I have created a PR to add this to CF java-buildpack:
https://github.com/cloudfoundry/java-buildpack/pull/160

I also created an issues
https://github.com/cloudfoundry/java-buildpack/issues/163 and
https://github.com/cloudfoundry/java-buildpack/pull/159 .

I hope this information helps others struggling with OOM problems
in CF.
I'm not saying that this is a ready made solution just for you.
YMMV. It worked for me.

-Lari




On 15-04-29 10:53 AM, Head-Rapson, David wrote:

Hi,

I’m after some guidance on how to get profile Java apps in CF, in
order to get to the bottom of memory issues.

We have an app that’s crashing every few hours with OOM error,
most likely it’s a memory leak.

I’d like to profile the JVM and work out what’s eating memory,
however tools like yourkit require connectivity INTO the JVM
server (i.e. the warden container), either via host / port or via
SSH.

Since warden containers cannot be connected to on ports other
than for HTTP and cannot be SSHd to, neither of these works for me.



I tried installed a standalone JDK onto the warden container,
however as soon as I ran ‘jmap’ to invoke the dump, warden
cleaned up the container – most likely for memory over-consumption.



I had previously found a hack in the Weblogic buildpack
(https://github.com/pivotal-cf/weblogic-buildpack/blob/master/docs/container-wls-monitoring.md)
for modifying the start script which, when used with
–XX:HeapDumpOnOutOfMemoryError, should copy any heapdump files to
a file share somewhere. I have my own custom buildpack so I
could use something similar.

Has anyone got a better solution than this?



We would love to use newrelic / app dynamics for this however
we’re not allowed. And I’m not 100% certain they could help with
this either.



Dave



The information transmitted is intended for the person or entity
to which it is addressed and may contain confidential, privileged
or copyrighted material. If you receive this in error, please
contact the sender and delete the material from any computer.
Fidelity only gives information on products and services and does
not give investment advice to retail clients based on individual
circumstances. Any comments or statements made are not
necessarily those of Fidelity. All e-mails may be monitored. FIL
Investments International (Reg. No.1448245), FIL Investment
Services (UK) Limited (Reg. No. 2016555), FIL Pensions
Management (Reg. No. 2015142) and Financial Administration
Services Limited (Reg. No. 1629709) are authorised and regulated
in the UK by the Financial Conduct Authority. FIL Life Insurance
Limited (Reg No. 3406905) is authorised in the UK by the
Prudential Regulation Authority and regulated in the UK by the
Financial Conduct Authority and the Prudential Regulation
Authority. Registered offices at Oakhill House, 130 Tonbridge
Road, Hildenborough, Tonbridge, Kent TN11 9DZ.

--
You received this message because you are subscribed to the
Google Groups "Cloud Foundry Developers" group.
To view this discussion on the web visit
https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com
<https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com?utm_medium=email&utm_source=footer>.
To unsubscribe from this group and stop receiving emails from it,
send an email to vcap-dev+unsubscribe(a)cloudfoundry.org
<mailto:vcap-dev+unsubscribe(a)cloudfoundry.org>.

_______________________________________________
Cf-dev mailing list
Cf-dev(a)lists.cloudfoundry.org <mailto:Cf-dev(a)lists.cloudfoundry.org>
https://lists.cloudfoundry.org/mailman/listinfo/cf-dev




--
Regards,

Daniel Jones
EngineerBetter.com

Hi Lari et al,

Thanks for your help Lari.

David and I are pairing on this issue, and we're yet to resolve it. We're
in the process of creating a repeatable test case (our most crashy app
makes calls to external services that need mocking), but in the meantime,
here's what we've seen.

Between Java Buildpack commit e89e546 and 17162df, we see apps crashing
with Warden out of memory errors. killjava.sh never fires, and this has led
us to believe that the kernel is shooting a cgroup process in the head
after the cgroup oversteps its memory limit. We cannot find any evidence of
the OOM killer firing in any logs, but we may not be looking in the right
place.

The JBP is setting heap to be 70%, metaspace to be 15% (with max set to the
same as initial), 5% for "stack", 5% for "normalised stack" and 10% for
"native". We do not understand why this adds up to 105%, but haven't looked
into the JBP algorithm yet. Any pointers on what "normalised stack" is
would be much appreciated, as this doesn't appear in the list of heuristics
supplied via app env.

Other team members tried applying the same settings that you suggested -
thanks for this. Apps still crash with these settings, albeit less
frequently.

After reading the blog you linked to (
http://java.dzone.com/articles/java-8-permgen-metaspace) we wondered
whether the increased *reserved *metaspace claimed after metaspace GC might
be causing a problem; however we reused the test code to create a metaspace
leak in a CF app and saw metaspace GCs occur correctly, and memory usage
never grow over MaxMetaspaceSize. This figures, as the committed metaspace
is still less than MaxMetaspaceSize, and the reserved appears to be
whatever RAM is free across the whole DEA.

We noted that an Oracle blog (
https://blogs.oracle.com/poonam/entry/about_g1_garbage_collector_permanent)
mentions that the metaspace size parameters are approximate. We're
currently wondering if native allocations by Tomcat (APR, NIO) are taking
up more container memory, and so when the metaspace fills, it's creeping
slightly over the limit and triggering the kernel's OOM killer.

Any suggestions would be much appreciated. We've tried to resist tweaking
heuristics blindly, but are running out of options as we're struggling to
figure out how the Java process is using *committed* memory. pmap seems to
show virtual memory, and so it's hard to see if things like the metaspace
or NIO ByteBuffers are nabbing too much and trigger the kernel's OOM killer.

Thanks for all your help,

Daniel Jones & David Head-Rapson

--
Regards,

Daniel Jones
EngineerBetter.com

Hi,

I created a few tools to debug OOM problems since the application I was
responsible for running on CF was failing constantly because of OOM
problems. The problems I had, turned out not to be actual memory leaks
in the Java application.

In the "cf events appname" log I would get entries like this:
2015-xx-xxTxx:xx:xx.00-0400 app.crash appname index:
1, reason: CRASHED, exit_description: out of memory, exit_status: 255

These type of entries are produced when the container goes over it's
memory resource limits. It doesn't mean that there is a memory leak in
the Java application. The container gets killed by the Linux kernel oom
killer
(https://github.com/cloudfoundry/warden/blob/master/warden/README.md#limit-handle-mem-value)
based on the resource limits set to the warden container.

The memory limit is specified in number of bytes. It is enforced using
the control group associated with the container. When a container
exceeds this limit, one or more of its processes will be killed by the
kernel. Additionally, the Warden will be notified that an OOM happened
and it subsequently tears down the container.

In my case it never got killed by the killjava.sh script that gets
called in the java-buildpack when an OOM happens in Java.

This is the tool I built to debug the problems:
https://github.com/lhotari/java-buildpack-diagnostics-app
I deployed that app as part of the forked buildpack I'm using.
Please read the readme about what it's limitations are. It worked for
me, but it might not work for you. It's opensource and you can fork it. :)

There is a solution in my toolcase for creating a heapdump and uploading
that to S3:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/HeapDumpServlet.groovy
The readme explains how to setup Amazon S3 keys for this:
https://github.com/lhotari/java-buildpack-diagnostics-app#amazon-s3-setup
Once you get a dump, you can then analyse the dump in a java profiler
tool like YourKit.

I also have a solution that forks the java-buildpack modifies
killjava.sh and adds a script that uploads the heapdump to S3 in the
case of OOM:
https://github.com/lhotari/java-buildpack/commit/2d654b80f3bf1a0e0f1bae4f29cb85f56f5f8c46

In java-buildpack-diagnostics-app I have also other tools for getting
Linux operation system specific memory information, for example:

https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemoryInfoServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MemorySmapServlet.groovy
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/MallocInfoServlet.groovy

These tools are handy for looking at details of the Java process RSS
memory usage growth.

There is also a solution for getting ssh shell access inside your
application with tmate.io:
https://github.com/lhotari/java-buildpack-diagnostics-app/blob/master/src/main/groovy/io/github/lhotari/jbpdiagnostics/TmateSshServlet.groovy
(this version is only compatible with the new "cflinuxfs2" stack)

It looks like there are serious problems on CloudFoundry with the memory
sizing calculation. An application that doesn't have a OOM problem will
get killed by the oom killer because the Java process will go over the
memory limits.
I filed this issue:
https://github.com/cloudfoundry/java-buildpack/issues/157 , but that
might not cover everything.

The workaround for that in my case was to add a native key under
memory_sizes in open_jdk_jre.yml and set the minimum to 330M (that is
for a 2GB total memory).
see example
https://github.com/grails-samples/java-buildpack/blob/22e0f6a/config/open_jdk_jre.yml#L25
that was how I got the app I'm running on CF to stay within the memory
bounds. I'm sure there is now also a way to get the keys without forking
the buildpack. I could have also adjusted the percentage portions, but I
wanted to set a hard minimum for this case.

It was also required to do some other tuning.

I added this to JAVA_OPTS:
-XX:CompressedClassSpaceSize=256M -XX:InitialCodeCacheSize=64M
-XX:CodeCacheExpansionSize=1M -XX:CodeCacheMinimumFreeSpace=1M
-XX:ReservedCodeCacheSize=200M -XX:MinMetaspaceExpansion=1M
-XX:MaxMetaspaceExpansion=8M -XX:MaxDirectMemorySize=96M
while trying to keep the Java process from growing in RSS memory size.

The memory overhead of a 64 bit Java process on Linux can be reduced by
specifying these environment variables:

stack: cflinuxfs2
.
.
.
env:
MALLOC_ARENA_MAX: 2
MALLOC_MMAP_THRESHOLD_: 131072
MALLOC_TRIM_THRESHOLD_: 131072
MALLOC_TOP_PAD_: 131072
MALLOC_MMAP_MAX_: 65536

MALLOC_ARENA_MAX works only on cflinuxfs2 stack (the lucid64 stack has a
buggy version of glibc).

explanation about MALLOC_ARENA_MAX from Heroku:
https://devcenter.heroku.com/articles/tuning-glibc-memory-behavior
some measurement data how it reduces memory consumption:
https://devcenter.heroku.com/articles/testing-cedar-14-memory-use

I have created a PR to add this to CF java-buildpack:
https://github.com/cloudfoundry/java-buildpack/pull/160

I also created an issues
https://github.com/cloudfoundry/java-buildpack/issues/163 and
https://github.com/cloudfoundry/java-buildpack/pull/159 .

I hope this information helps others struggling with OOM problems in CF.
I'm not saying that this is a ready made solution just for you. YMMV. It
worked for me.

-Lari



On 15-04-29 10:53 AM, Head-Rapson, David wrote:

Hi,

I’m after some guidance on how to get profile Java apps in CF, in
order to get to the bottom of memory issues.

We have an app that’s crashing every few hours with OOM error, most
likely it’s a memory leak.

I’d like to profile the JVM and work out what’s eating memory, however
tools like yourkit require connectivity INTO the JVM server (i.e. the
warden container), either via host / port or via SSH.

Since warden containers cannot be connected to on ports other than for
HTTP and cannot be SSHd to, neither of these works for me.



I tried installed a standalone JDK onto the warden container, however
as soon as I ran ‘jmap’ to invoke the dump, warden cleaned up the
container – most likely for memory over-consumption.



I had previously found a hack in the Weblogic buildpack
(https://github.com/pivotal-cf/weblogic-buildpack/blob/master/docs/container-wls-monitoring.md)
for modifying the start script which, when used with
–XX:HeapDumpOnOutOfMemoryError, should copy any heapdump files to a
file share somewhere. I have my own custom buildpack so I could use
something similar.

Has anyone got a better solution than this?



We would love to use newrelic / app dynamics for this however we’re
not allowed. And I’m not 100% certain they could help with this either.



Dave



The information transmitted is intended for the person or entity to
which it is addressed and may contain confidential, privileged or
copyrighted material. If you receive this in error, please contact the
sender and delete the material from any computer. Fidelity only gives
information on products and services and does not give investment
advice to retail clients based on individual circumstances. Any
comments or statements made are not necessarily those of Fidelity. All
e-mails may be monitored. FIL Investments International (Reg.
No.1448245), FIL Investment Services (UK) Limited (Reg. No. 2016555),
FIL Pensions Management (Reg. No. 2015142) and Financial
Administration Services Limited (Reg. No. 1629709) are authorised and
regulated in the UK by the Financial Conduct Authority. FIL Life
Insurance Limited (Reg No. 3406905) is authorised in the UK by the
Prudential Regulation Authority and regulated in the UK by the
Financial Conduct Authority and the Prudential Regulation Authority.
Registered offices at Oakhill House, 130 Tonbridge Road,
Hildenborough, Tonbridge, Kent TN11 9DZ.

--
You received this message because you are subscribed to the Google
Groups "Cloud Foundry Developers" group.
To view this discussion on the web visit
https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com
<https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com?utm_medium=email&utm_source=footer>.
To unsubscribe from this group and stop receiving emails from it, send
an email to vcap-dev+unsubscribe(a)cloudfoundry.org
<mailto:vcap-dev+unsubscribe(a)cloudfoundry.org>.

Dave,

I've found using SSH tunnels can make debugging the JVM easier. I haven't
tried using Yourkit via a tunnel, but I suspect it would work. Maybe worth
a try. Some more details on setting up the tunnel here.

http://mikusa.blogspot.com/2014/08/debugging-java-applications-on.html

Dan

On Wed, Apr 29, 2015 at 10:53 AM, Head-Rapson, David <
David.Head-Rapson(a)fil.com> wrote:

Hi,

I’m after some guidance on how to get profile Java apps in CF, in order to
get to the bottom of memory issues.

We have an app that’s crashing every few hours with OOM error, most likely
it’s a memory leak.

I’d like to profile the JVM and work out what’s eating memory, however
tools like yourkit require connectivity INTO the JVM server (i.e. the
warden container), either via host / port or via SSH.

Since warden containers cannot be connected to on ports other than for
HTTP and cannot be SSHd to, neither of these works for me.



I tried installed a standalone JDK onto the warden container, however as
soon as I ran ‘jmap’ to invoke the dump, warden cleaned up the container –
most likely for memory over-consumption.



I had previously found a hack in the Weblogic buildpack (
https://github.com/pivotal-cf/weblogic-buildpack/blob/master/docs/container-wls-monitoring.md)
for modifying the start script which, when used with
–XX:HeapDumpOnOutOfMemoryError, should copy any heapdump files to a file
share somewhere. I have my own custom buildpack so I could use something
similar.

Has anyone got a better solution than this?



We would love to use newrelic / app dynamics for this however we’re not
allowed. And I’m not 100% certain they could help with this either.



Dave



The information transmitted is intended for the person or entity to which
it is addressed and may contain confidential, privileged or copyrighted
material. If you receive this in error, please contact the sender and
delete the material from any computer. Fidelity only gives information on
products and services and does not give investment advice to retail clients
based on individual circumstances. Any comments or statements made are not
necessarily those of Fidelity. All e-mails may be monitored. FIL
Investments International (Reg. No.1448245), FIL Investment Services (UK)
Limited (Reg. No. 2016555), FIL Pensions Management (Reg. No. 2015142) and
Financial Administration Services Limited (Reg. No. 1629709) are authorised
and regulated in the UK by the Financial Conduct Authority. FIL Life
Insurance Limited (Reg No. 3406905) is authorised in the UK by the
Prudential Regulation Authority and regulated in the UK by the Financial
Conduct Authority and the Prudential Regulation Authority. Registered
offices at Oakhill House, 130 Tonbridge Road, Hildenborough, Tonbridge,
Kent TN11 9DZ.

--
You received this message because you are subscribed to the Google Groups
"Cloud Foundry Developers" group.
To view this discussion on the web visit
https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com
<https://groups.google.com/a/cloudfoundry.org/d/msgid/vcap-dev/DFFA4ADB9F3BC34194429921AB329336408CAB04%40UKFIL7006WIN.intl.intlroot.fid-intl.com?utm_medium=email&utm_source=footer>
.

To unsubscribe from this group and stop receiving emails from it, send an
email to vcap-dev+unsubscribe(a)cloudfoundry.org.