C/P/N/Q experiment: one-shot with 10ms think time

Aleksey Shipilev aleksey.shipilev at oracle.com
Mon Oct 15 03:41:34 PDT 2012


This is the update on C/P/N/Q experiment. This time we had fixed a few
of methodological glitches, and also shifted our tests to one-shot modes
with 10ms think time in between the invocations. This allows FJP to
obtain quiscence, and is better in assessing the "cold start" performance.

The results for decomposition benchmark on 2x8x2 Xeon E5-2680
(SandyBridge) running Solaris 11, and 20120925 lambda nightly with -d64
-XX:-TieredCompilation -XX:+UseParallelOldGC -XX:+UseNUMA
-XX:-UseBiasedLocking -XX:+UseCondCardMark, is here [1].

This time it was generally focused on C=1, P=32 plane. You might find it
convenient to look into high-level charts [2]. Raw data is here [3].

  - the break-even chart is very similar to one we had before
  - there is a vague additional break-even isoline above the general one

All in all, while this experiment is supposed to be more accurate than
previous, the same empirical conclusions hold.


[1] http://shipilev.net/pub/jdk/lambda/20121003-breakeven/
[3] http://shipilev.net/pub/jdk/lambda/20121003-breakeven/score.data

More information about the lambda-dev mailing list