Sam Whittle created BEAM-11707:
----------------------------------
Summary: Optimize WindmillStateCache CPU usage
Key: BEAM-11707
URL: https://issues.apache.org/jira/browse/BEAM-11707
Project: Beam
Issue Type: Bug
Components: runner-dataflow
Reporter: Sam Whittle
Assignee: Sam Whittle
>From profiling nexmark Query11 which has many unique tags per key, I observed
>that the WindmillStateCache cpu usage was 6% of CPU.
The usage appears to be due to the invalidation set maintenance as well as many
reads/inserts.
The invalidation set is maintained so that if a key encounters an error
processing or the cache token changes, we can invalidate all the entries for a
key. Currently this is done by removing all entries for the key from the
cache. Another alternative which appears much more CPU efficient is to instead
leave the entries in the cache but make them unreachable. This can be done by
having a per-key object that uses object equality as part of the cache lookup.
Then to discard entries for the key, we start using a new per-key object.
Cleanup of per-key objects can be done with a weak reference map.
Another cost to the cache is that objects are grouped by window so that they
are kept/evicted all at once. However currently when reading items into the
cache, we fetch the window set and then lookup each tag in it. This could be
cached for the key to avoid multiple cache lookups. Similarly for putting
objects we lookup and insert each tag separately and then update the cache to
update the weight for the per-window set. This could be done once after all
updates for the window have been made.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
