5.2.2 server takes over all CPU - becomes unresponsive

[dsudar]

Hi,
I upgraded my Ubuntu 14.04 machine to 5.2.2 on Monday (from 5.1.4) and large imports of ScanR plates (384 wells, 4 channels, 4 fields) since then cause the OMERO server to take over all CPU on the machine which makes the machine unresponsive for a long time. In fact, OMERO.web session crashed because of that (see QA case 17124). This appears to happen immediately after the image import has concluded and it's probably creating thumbnails or something like that.

Here's a screen shot of htop while this is happening:

htop_OMERO.GIF (116.69 KiB) Viewed 2924 times

In master.err, I get a number of exceptions related to GC overhead limit exceeded:

Exception in thread "bitronix-task-scheduler" Exception in thread "OMERO.scheduler_QuartzSchedulerThread" java.lang.OutOfMemoryError: GC overhead limit exceeded
at bitronix.tm.timer.TaskScheduler.executeElapsedTasks(TaskScheduler.java:258)
at bitronix.tm.timer.TaskScheduler.run(TaskScheduler.java:244)
Exception in thread "Blitz-0-Ice.ThreadPool.Server-8" java.lang.OutOfMemoryError: GC overhead limit exceeded
Exception in thread "Blitz-0-Ice.ThreadPool.Client-6" Exception in thread "Timer-0" java.lang.OutOfMemoryError: GC overhead limit exceeded

Exception: java.lang.OutOfMemoryError thrown from the UncaughtExceptionHandler in thread "OMERO.scheduler_QuartzSchedulerThread"

Exception: java.lang.OutOfMemoryError thrown from the UncaughtExceptionHandler in thread "pool-5-thread-7"
Exception in thread "pool-5-thread-7" java.lang.OutOfMemoryError: GC overhead limit exceeded
java.lang.OutOfMemoryError: GC overhead limit exceeded
Exception in thread "Blitz-0-Ice.ThreadPool.Server-16"
Exception: java.lang.OutOfMemoryError thrown from the UncaughtExceptionHandler in thread "Blitz-0-Ice.ThreadPool.Server-16"
Exception in thread "Blitz-0-Ice.ThreadPool.Server-19" java.lang.OutOfMemoryError: GC overhead limit exceeded
Exception in thread "Ice.Timer" java.lang.OutOfMemoryError: GC overhead limit exceeded
Exception in thread "perf4j-async-stats-appender-sink-CoalescingStatistics" java.lang.OutOfMemoryError: GC overhead limit exceeded
Exception in thread "Blitz-0-Ice.ThreadPool.Client-8" java.lang.OutOfMemoryError: GC overhead limit exceeded

None of this happened under 5.1.4 and outside the 5.2.2 upgrade nothing else changed on the machine.

Thanks,
- Damir

[dsudar]

Quick follow-up: I tried importing the exact same data set with another OMERO 5.2.2 instance on a Centos6.7 system (similarly powerful system as the Ubuntu unit). It imported okay but I did notice that immediately after the command line import concluded, the CPU use spiked a number of times extremely high but then came down to a normal low CPU use within 1 minute.
I have that data set ready for your testing tar'ed and bzip2'ed up (5.5GB) in case that helps.
Thanks,
- Damir

[atarkowska]

Hi Damir,

Could you try to increase the heap size? For more details please look at https://www.openmicroscopy.org/site/sup ... mance.html

Ola

[dsudar]

Hi Ola,
Thanks for the quick response.

I did not yet try a larger heap size and had been using the percent method which results in:

omero_user@omero:~$ omero admin jvmcfg
JVM Settings:
============
blitz=-Xmx5053m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions
indexer=-Xmx3368m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions
pixeldata=-Xmx5053m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions
repository=-Xmx3368m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions

The server has 32GB physical memory (I'm waiting for more memory that is on back-order). What would you suggest I try for the new settings?

And do you think that allocating more memory will mitigate the extreme CPU usage issue? I had the same settings with 5.1.4 and this issue never occurred.

Thanks,
- Damir

[dsudar]

Hi Ola,

On my Ubuntu 14.04 server I upped the allocations to all the 4 components and re-ran the import
omero admin jvmcfg is now:

omero_user@omero:~$ omero admin jvmcfg
JVM Settings:
============
blitz=-Xmx10106m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions # Settings({'percent': '30'})
indexer=-Xmx5053m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions # Settings({'percent': '15'})
pixeldata=-Xmx10106m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions # Settings({'percent': '30'})
repository=-Xmx5053m -XX:MaxPermSize=1g -XX:+IgnoreUnrecognizedVMOptions # Settings({'percent': '15'})

As before the import itself completes and the extreme CPU usage spikes are there for about 2 minutes and then it quiets down to normal levels. So indeed, it appears that allocating more memory helps. But I'm still wondering, what is causing those extreme CPU usage spikes immediately after the import itself completes? Is that thumbnail creation or something?

Thanks,
- Damir

[atarkowska]

Hi Damir,

Could you confirm if you tried exactly the same import on 5.1?

Could you try

Code: Select all

bin/omero import ... --debug

(see https://www.openmicroscopy.org/site/sup ... ort--debug) or

Code: Select all

$ importer-cli ... --debug ALL

, then pipe that to a file and also send us server logs to help say what is happening

Did you try to turn off thumbnail generation on import

Code: Select all

--skip thumbnails

(see https://www.openmicroscopy.org/site/sup ... port--skip) to see if it makes any difference with CPU usage.

Ola