Bug 20939 - REGRESSION(r36446-r36647) Sunspider on windows nightlies 25-33% slower
Summary: REGRESSION(r36446-r36647) Sunspider on windows nightlies 25-33% slower
Status: RESOLVED FIXED
Alias: None
Product: WebKit
Classification: Unclassified
Component: JavaScriptCore (show other bugs)
Version: 528+ (Nightly build)
Hardware: Other Windows XP
: P2 Normal
Assignee: Nobody
URL:
Keywords: InRadar
Depends on:
Blocks:
 
Reported: 2008-09-19 09:11 PDT by nemo
Modified: 2008-09-22 08:04 PDT (History)
4 users (show)

See Also:


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description nemo 2008-09-19 09:11:46 PDT
Running under Windows XP, 2.16GigHz processor, 2 gigs of RAM.  On multiple sunspider runs, the build from the 15th (36446) is 25-33% faster than the build from the 19th (36647).

My runs:
15th
http://x.m8y.org/2o
http://x.m8y.org/2t

19th
http://x.m8y.org/2r
http://x.m8y.org/2s
Comment 1 Adam Roben (:aroben) 2008-09-19 09:14:19 PDT
<rdar://problem/6232388>
Comment 2 Adam Roben (:aroben) 2008-09-19 09:24:18 PDT
Could this be related to using high-res timers on XP (r36626)?
Comment 3 nemo 2008-09-19 09:31:59 PDT
Higher res eh.  Maybe that explains Dromaeo.

http://dromaeo.com/?id=42819,42808
Dromaeo is also slower.  15th is on the left.

Whatever it was, drawing stuff seems faster.
jsMSX is significantly faster on 19th (32 interrupts per second for 5 minutes of letting the goonies emulator do its thing vs 25 interrupts per second on the 15th).

Opera's DHTML performance test also, although not as much.
http://people.opera.com/~pettern/performance-1.html

To get knocked down to 30fps:
19th - 760 dots
15th - 740 dots

So that 25-33% speedup was just due to lower resolution? Hum.
Makes me question benchmarks :(
Comment 4 nemo 2008-09-19 09:57:55 PDT
As was requested, retested with 36620, which is apparently pre-timer resolution increase.
Indeed, I got a slightly higher speed than 15th.  1071ms overall.

Interesting, if it is just "better measurements" then that erases most of the lead over TraceMonkey on SunSpider :)
http://x.m8y.org/2u  -  TraceMonkey - with a result that is smack dab in middle of my 2 WebKit runs.

Comment 5 nemo 2008-09-19 10:11:54 PDT
Interestingly, though, Dromaeo *remains* slower.
http://dromaeo.com/?id=42819,42858,42808
Representing:         r36446,r36620,r36647

Repeated dromaeo and sunspider over two tests.

      Sunspider   Dromaeo
36446  fast         fast
36620  fast         slow
36647  slow         slow

And build 36620 - theoretically pre-high res timer, did not have the 0ms results that the one from the 15th had, in Dromaeo.
Comment 6 nemo 2008-09-19 10:18:52 PDT
As requested, here are detailed HW specs.  From Dell, not sure how to extract details of HW in windows, apart
from memory and cpu speed.
Dell Latitude D820
Intel® Core™ Duo processor
533 or 667 MHz bus
2GiB of 533/667 DDRII SDRAM

http://support.dell.com/support/edocs/systems/latd820/en/ug/specs.htm
Comment 7 Steve Falkenburg 2008-09-19 16:14:10 PDT
Confirmed that rolling out r36626 fixes the SunSpider regression on XP.
Comment 8 nemo 2008-09-22 08:04:59 PDT
As of latest nightly for windows...
      Sunspider   Dromaeo
36446  fast         fast
36620  fast         slow
36647  slow         slow
36712  fast         slow

http://dromaeo.com/?id=42819,42865,42808,43896

What's odd about the Dromaeo results is that it was slow in r36620 (and also had no 0s in the tests) which was supposedly
before the improved timer resolution.
However, the results from 36446 may have simply been buggy.

Since Sunspider is now a speed 1024ms on my machine running r36712 (instead of 1450-1500) going with fixed.