morrySnow commented on PR #66951:
URL: https://github.com/apache/doris/pull/66951#issuecomment-5342655494

   ### Benchmark status
   
   I ran the primary-expression JMH benchmark before claiming any performance 
result.
   
   Configuration:
   - Host: shared `dev-hk` (Intel Xeon Platinum 8457C), OpenJDK 17.0.18, JMH 
1.37
   - Baseline: `910b741ba329efb8f406fd13e87520e9bed95501`, jar SHA-256 
`13875a066f7261efc2809f6ea032bbf78aaec17972693925f31539ad28dae3c7`
   - Candidate: `1bfe0afbd4997ec48709694486a94b0d2a8791b2`, jar SHA-256 
`65e2cad21d2592523302d05becc913a8605650afd5e692a720486d4ad801dae6`
   - 1 thread, 5 forks, 8 x 1 s warmup, 10 x 1 s measurement, `-prof gc`; 
parser-only, standalone end-to-end, controls, and late-error cases
   - Affinity and telemetry: logical CPU 30, monitored with its SMT sibling 
126; steal was zero. Synchronous C2 compilation (`-XX:-TieredCompilation 
-Xbatch`) was used after a pilot showed tiered-compilation fork drift.
   
   Result: **the performance gate is not passed, so this PR currently makes no 
performance claim.**
   
   The shared host produced runtime contention that `taskset` cannot exclude. A 
full baseline round had 16/24 cases above the pre-set 5% fork-CV limit (maximum 
39.88%). A smaller baseline pilot was stable (four selected cases at 
1.01%-3.26% CV), but candidate C1 failed on one end-to-end case at 6.20% CV. 
Repeating C1 moved the noise to different parser-only cases (20.01% and 26.87% 
CV), which confirms environmental interference rather than a reproducible 
candidate regression. Rejected rounds and per-CPU telemetry were retained 
instead of deleting slow forks.
   
   Before merge, the benchmark should be rerun on a dedicated host or exclusive 
cpuset using the same artifacts and `B1-C1-C2-B2` order. The existing 
correctness tests/build remain valid, but they do not substitute for this 
performance gate.


-- 
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]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to