[
https://issues.apache.org/jira/browse/IGNITE-22835?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17911413#comment-17911413
]
Alexey Scherbakov commented on IGNITE-22835:
--------------------------------------------
Implemented shared distruptor (single queue for all events) and benchmarked
this after merging recent main (rev b2edb816b85e057abb8bb688af6549b309f8a1cb)
32 writer threads
Results are:
# shared=false logit=true
Benchmark (batch) (fsync) (partitionCount) Mode Cnt
Score Error Units
UpsertKvBenchmark.upsert 1 false 32 thrpt 20
212379.795 ± 28700.643 ops/s
# shared=true logit=false
Benchmark (batch) (fsync) (partitionCount) Mode Cnt
Score Error Units
UpsertKvBenchmark.upsert 1 false 32 thrpt 20
123629.541 ± 11835.884 ops/s
# shared=true logit=true
Benchmark (batch) (fsync) (partitionCount) Mode Cnt
Score Error Units
UpsertKvBenchmark.upsert 1 false 32 thrpt 20
262326.805 ± 35990.099 ops/s
# shared=false logit=false
Benchmark (batch) (fsync) (partitionCount) Mode Cnt
Score Error Units
UpsertKvBenchmark.upsert 1 false 32 thrpt 20
167142.261 ± 15218.456 ops/s
Summary: this optimization shines only with logit log storage, so we need to
productize it first before merging this.
> Latency penalty for using disruptor threads
> -------------------------------------------
>
> Key: IGNITE-22835
> URL: https://issues.apache.org/jira/browse/IGNITE-22835
> Project: Ignite
> Issue Type: Improvement
> Reporter: Vladislav Pyatkov
> Assignee: Alexey Scherbakov
> Priority: Major
> Labels: ignite-3
> Time Spent: 10m
> Remaining Estimate: 0h
>
> h3. Motivation
> All three disruptors are part of the RAFT implementation. Each time a command
> has to be replicated, all the disruptors are used. This leads to a bad impact
> on the entire operation's latency.
> {noformat}
> onBeforeApplyCmd:%node_3344%JRaft-Request-Processor-3 0.2 5415400 5415600
> Here is hidden 6.1 us
> RaftNodeOnEvent:%node_3344%JRaft-NodeImpl-Disruptor_stripe_2-0 0.0 5421700
> 5421700
> --
> RaftNodeOnEvent:%node_3344%JRaft-NodeImpl-Disruptor_stripe_2-0 0.0 5421700
> 5421700
> Here is hidden 2.8 us
> LogManagerOnEvent:%node_3344%JRaft-LogManager-Disruptor_stripe_2-0 0.1
> 5424500 5424600
> --
> flushLog:%node_3344%JRaft-LogManager-Disruptor_stripe_2-0 18.7 5426400 5445100
> Here is hidden 2.9 us
> FSMCallerOnEvent:%node_3344%JRaft-FSMCaller-Disruptor_stripe_1-0 0.1 5448000
> 5448100
> {noformat}
> h3. Definition of done
> * We can try to use the disruptor wait policy, depending on the case.
> * Maybe we can reduce a number of the dusruptor queues.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
