GGraziadei commented on PR #8796:
URL: https://github.com/apache/storm/pull/8796#issuecomment-4760284319

   ## Performance analysis 
   
   The benchmarks were run against the `FileReadWordCountTopo` topology, a 
standard word-count workload from Storm's performance suite, exercising the 
inter-executor transfer path that JCQueue's producer batching governs. The 
topology was deployed on a dockerized dev-cluster, providing a reproducible, 
self-contained Storm environment so that each configuration ran under identical 
resource and parallelism conditions, with the same fixed topology shape (2 
workers, 7 tasks, 7 executors, single spout executor) across all runs so that 
the producer batch-sizing strategy was the only variable. The results 
consistently favored dynamic batching: across batch ceilings of 10, 100, and 
1000 it matched or beat every static configuration on all three metrics, 
delivering throughput gains of up to ~9% and average complete-latency 
reductions of 8–12%, with the largest benefit at small ceilings where static 
batching faces the sharpest latency-versus-throughput trade-off and no measured 
downside at a
 ny setting. An extended 600-second run at ceiling 1000 further showed a brief 
learning phase after which the AIMD controller converges on a stable optimum, 
with latency settling near 376 ms within a ~2 ms standard deviation, confirming 
that the policy discovers a good batch size online and then locks onto it 
rather than relying on a manually tuned fixed value.
   
   Attached raw data and report.
   
   
[jcqueue-dynamic-batch-size.txt](https://github.com/user-attachments/files/29168573/jcqueue-dynamic-batch-size.txt)
   
   [Dynamic Producer Batch Sizing in Apache 
Storm.pdf](https://github.com/user-attachments/files/29168610/Dynamic.Producer.Batch.Sizing.in.Apache.Storm.pdf)
   
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to