Siddhant Sangwan created HDDS-14345:
---------------------------------------
Summary: Performance optimisation for random pipeline choose policy
Key: HDDS-14345
URL: https://issues.apache.org/jira/browse/HDDS-14345
Project: Apache Ozone
Issue Type: Improvement
Components: SCM
Reporter: Siddhant Sangwan
Assignee: Siddhant Sangwan
Using ThreadLocalRandom instead of Math.random() gives significant performance
improvement in my micro benchmark. Small improvement for the overall system,
but still worth it since it's a simple change and this policy is used for each
container allocation.
Without ThreadLocalRandom (current master branch):
{code}
---- S6 pipelines=500 healthy=100% threads=8 ----
BEGIN_MEASURE method=choosePipelineIndex policy=Random pipelines=500 threads=8
itersPerThread=200000
END_MEASURE method=choosePipelineIndex policy=Random pipelines=500 threads=8
itersPerThread=200000
choosePipelineIndex Random n=500 thr=8 295.03 ns/op 3389471
ops/s sink=11761373
{code}
So for selecting one out of 500 pipelines and 8 threads doing this, the latency
is 295.03 nanoseconds per operation.
With ThreadLocalRandom:
{code}
---- S6 pipelines=500 healthy=100% threads=8 ----
BEGIN_MEASURE method=choosePipelineIndex policy=Random pipelines=500 threads=8
itersPerThread=200000
END_MEASURE method=choosePipelineIndex policy=Random pipelines=500 threads=8
itersPerThread=200000
choosePipelineIndex Random n=500 thr=8 6.25 ns/op 160065355
ops/s sink=12453620
{code}
Latency = 6.25 nanoseconds per operation.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
