Comment #3 on issue 2053 by [email protected]: Unexpected performance
varaiations on physics simulation
http://code.google.com/p/v8/issues/detail?id=2053
I have started getting incrementally slower results benchmarking with d8
(25% slower) -- and am mystified about the reason.
I previously collected and reported these results showing a performance
average of just over 2000 model-steps/s.
$ out/native/d8 md2d.js
initializing to temp 100
*** MD warmup: 1000 steps
*** MD benchmark: 5 runs, 1000 steps
*** MD run 1: 2028
*** MD run 2: 2009.9999999999998
*** MD run 3: 1993
*** MD run 4: 2072
*** MD run 5: 2015.0000000000002
*** MD ave: 2023.6
Now I am averaging a bit under 1500.
I have NOT re-compiled D8.
After the slowdown started occurring I then re-started my mac, holding the
*shift* key down after login to
prevent any startup items from running and having almost no other programs
running.
The following data were collected after the restart:
$ out/native/d8 md2d.js
initializing to temp 100
*** MD warmup: 1000 steps
*** MD benchmark: 5 runs, 1000 steps
*** MD run 1: 1497
*** MD run 2: 1465
*** MD run 3: 1483
*** MD run 4: 1493
*** MD run 5: 1478
*** MD ave: 1483.2
I would like to understand better what could cause this variation -- I'd
like to have a stable way of generating
performance benchmarks so that increases or decreases in performance can be
attributed to changes we make to the modeling engine
The data collected above were done after re-starting my mac, holding the
*shift* key down after login to
prevent any startup items from running and having almost no other programs
running.
Since making the change to md2d.js to collect and average five benchmarking
runs I have made only trivial changes to md2d.js.
Testing the earliest version of md2d.js that includes the multiple runs
feature produces the same result:
$ curl
https://raw.github.com/gist/2289269/8460935d9219d516f7c89de9cb12887bcb33d90f/md2d.js
-o md2d-84609.js
$ out/native/d8 md2d-84609.js
initializing to temp 100
*** MD warmup: 1000 steps
*** MD benchmark: 5 runs, 1000 steps
*** MD run 1: 1457
*** MD run 2: 1439
*** MD run 3: 1449
*** MD run 4: 1448
*** MD run 5: 1446
*** MD ave: 1447.8
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
