Operating LanguageTool: Resources and Configuration

Development
1

Hey Community!

This is a follow-up to this topic: Fragen zur Optimierung des LanguageTool

We have forked languageTool and build our own image that we run it in our kubernetes cluster. We face issues with the memory consumption that we were not able to fix although we made quite a few experiments.

The current setup

resources:
  requests:
    cpu: 2
    memory: 3584Mi
  limits:
    cpu: 4
    memory: 4096Mi
    ephemeral-storage: 1Gi

Observation

What we observe is that languageTool gets OOM killed by the kernel (meaning that it exceeds the 4GB limit) after getting about 15 requests shortly after one another (each approx. 50kB, 100.000 characters total, 75.000 text, 25.000 annotations, 80 matches). Apart from this, there was no other load.

What we tried

  • we removed fasttext to make sure it doesn’t not add to the memory - no change
  • we tried to set the Xmx of the java process to 2GB or 3GB - no change
  • we tried to set the maxCheckThreads to 4 (equal to the max number of CPUs) and the maxWorkQueueSize size to 50 - no change

To me, it seems suspicious that this amout of requests exceed 4GB of RAM.

What we considered trying

  • giving it more memory - we try to avoid that :cross_mark:
  • change or configure the garbage collector, e.g. setting the ratio of younf and old items with XX:NewRatio or trying ZGC or Shenandoah
  • configure metaspace e.g. -XX:MaxMetaspaceSize=256m
  • splitting the requests to LanguageTool to achieve a smaller size
  • up to now, I haven’t tried the combination of maxCheckThreads and setting Xmx 2GB/3GB but I expect no substancial change

What do you suggest? Is there anything else you would try?

1 Like