[
https://issues.apache.org/jira/browse/HADOOP-6166?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12735908#action_12735908
]
Scott Carey commented on HADOOP-6166:
-------------------------------------
For the 32 bit results, try passing -server on the command line. It behaves
quite differently with loop unrolling and certain low level optimizations in
the JIT versus -client (which is only default on 32 bit windows, and anyone who
would run Hadoop there and wanted better performance would pass -server to
speed it up).
Are you specifying a -Xmx memory value? What about -Xms? On windows with
-client, the VM has unusual default memory and GC values, I've found that
setting its NewRatio more like the other platforms helps a lot: -XX:NewRatio=4
or something like that may make your results more consistent across the
platforms (and faster on 32 bit windows).
On my environment, on the previous set of tests, changing from _10 to _12 to
_14 on JDK6 did not seem to do much. But I was manually setting -Xmx512m for
all of my tests. I can try again later, but there is something odd about the
results slowing down so much on the 1.6.0_14 version.
It is also curious that the PureJavaCrc32New -- which only changes the loop
style --also slows down but not as much as the older PureJavaCrc32 and goes
from always about 15% slower to a little bit faster. My guess is something
configuration related has changed with respect to some default JVM settings.
I think there may be some improvement possible in the 8_8 case in how the 9
XORs at the end are done. Perhaps all in one line? or in 3 sets of 3? Or
more likely the compiler is smart enough to do the register optimization
itself? Perhaps not, Intel's C code even avoids a single line with more than 4
XORs at once for some reason.
> Improve PureJavaCrc32
> ---------------------
>
> Key: HADOOP-6166
> URL: https://issues.apache.org/jira/browse/HADOOP-6166
> Project: Hadoop Common
> Issue Type: Improvement
> Components: util
> Reporter: Tsz Wo (Nicholas), SZE
> Assignee: Tsz Wo (Nicholas), SZE
> Attachments: c6166_20090722.patch, c6166_20090722_benchmark_32VM.txt,
> c6166_20090722_benchmark_64VM.txt, c6166_20090727.patch
>
>
> Got some ideas to improve CRC32 calculation.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
