[
https://issues.apache.org/jira/browse/CASSANDRA-15388?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16993993#comment-16993993
]
Blake Eggleston commented on CASSANDRA-15388:
---------------------------------------------
Looking at the code for GCProfiler, it looks like the ThreadMXBean is used to
measure allocation. ThreadMXBean was one of the first things I looked at, and
it wasn’t accurate enough, although it is available as an alternate measurement
method in these tests.
AFAIK, ThreadMXBean uses TLAB overflows to approximate allocations with minimum
overhead, and the java allocation instrumenter actually invokes the measurement
method anytime anything is allocated. So it’s much slower, but super accurate,
which is perfect for this use case.
> Add compaction allocation measurement test to support compaction gc
> optimization.
> ----------------------------------------------------------------------------------
>
> Key: CASSANDRA-15388
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15388
> Project: Cassandra
> Issue Type: Sub-task
> Components: Local/Compaction
> Reporter: Blake Eggleston
> Assignee: Blake Eggleston
> Priority: Normal
> Fix For: 4.0
>
>
> This adds a test that is able to quickly and accurately measure the effect of
> potential gc optimizations against a wide range of (synthetic) compaction
> workloads. This test accurately measures allocation rates from 16 workloads
> in less that 2 minutes.
> This test uses google’s {{java-allocation-instrumenter}} agent to measure the
> workloads. Measurements using this agent are very accurate and pretty
> repeatable from run to run, with most variance being negligible (1-2 bytes
> per partition), although workloads with larger but fewer partitions vary a
> bit more (still less that 0.03%).
> The thinking behind this patch is that with compaction, we’re generally
> interested in the memory allocated per partition, since garbage scales more
> or less linearly with the number of partitions compacted. So measuring
> allocation from a small number of partitions that otherwise represent real
> world use cases is a good enough approximation.
> In addition to helping with compaction optimizations, this test could be used
> as a template for future optimization work. This pattern could also be used
> to set allocation limits on workloads/operations and fail CI if the
> allocation behavior changes past some threshold.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
