Github user Ethanlm commented on the issue:
https://github.com/apache/storm/pull/2366
I ran some other performance tests this weekend and here are some results
supporting this patch. Below are average results among all the 27 tests ( 2/4/8
workers * 5000/10000/20000 rates * 3 repetition) for each experiment
(ORIG/SHUFFLE/STORM-2686-H80L20/STORM-2686-H90L40/STORM-2686-H70L30). Thanks
@revans2 . I am going to fix the acker bug now.
|Experiment | rate(tuple/s) | mean(ms) | 99%ile(ms) | 99.9%ile(ms) |
cores | mem(MB)
-- | -- | -- | -- | -- | -- | --
ORIG | 11668.1 | 11.0787 | 18.7359 | 31.2698 | 2.61209 | 351.185
SHUFFLE | 10952.2 | 476.725 | 744.26 | 779.364 | 2.75163 | 511.239
STORM-2686-H80L20 | 11660.9 | 10.3228 | 17.6535 | 32.0146 | 2.62504 |
374.756
STORM-2686-H90L40 | 11659.6 | 10.6373 | 18.5732 | 32.7845 | 2.6264 | 416.69
STORM-2686-H70L30 | 11734.4 | 45.3912 | 504.058 | 571.151 | 2.6467 | 450.725
Note:
SHUFFLE and STORM-2686-H70L30 has extreme high overall average latency
because they have very high average latency in some single experiment
(2vm8spout8splitter8counter20000r.20171014140308 and
2vm8spout8splitter8counter20000r.20171016023641 respectively).
Results are uploaded at
https://github.com/Ethanlm/STORM-2686-perf-study/commit/a790e780e8d3db5f0b68c87320af1aad5a03e0ae
---
