httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ian Holsman <i...@apache.org>
Subject Benchmark Findings v29 vs v31+jre
Date Fri, 15 Feb 2002 05:22:07 GMT
both of these test have been run with the MTMALLOC library installed
http://webperf.org/a2/v31/2002-02-11-v29/
http://webperf.org/a2/v31/2002-02-12-v31/

without the MTMALLOC library these are the results from v31
http://webperf.org/a2/v31/2002-02-11-v31/

ok.. I've added a breif explanation about what each of the graphs mean.
if there's anything unclear please email me and I'll explain further 
(and add it to the docs for next time)


Some commentry about the test itself.
these tests were run on a 8 CPU E4000 solaris 2.8 machine.
which has it docroot mounted on a NETAppFiler running NFS v2.
There are some custom modules in here, but these pieces have not been
modified between releases.

The test itself involves a random GET of a file. (out of 100,000)
each file has the SAME 10 component includes.

The # of connections/second attempted is increased slowly over time. We 
simulate users coming in over DSL/T1 and 56K modems (40/40/20%)

The test aimed to simulate how one would run the machine in production, 
ie the aim of the test is not to see how many pages we can serve off the 
machine. the aim is to see how the machine copes while increasing the 
load over time. (hence why the machine never reaches 100% cpu utilization)

I ran these for v31+jre patches and not the 32-code base, basically due
to internal time constraints.. these numbers should reflect the 32-code.

For me the most important charts on the page are the response times and 
the CPU utilization.

Ok.. some commentry about the tests results.

1. MTMalloc is a must when running on a SMP solaris machine.
2. On a lightly loaded machine there is no real difference between v29 &
    v31 releases.
3. On a medium load, you can see that the performance improvements in
    v31 come into play. I am presuming most of these are due to the new
    pool allocation code reducing the mutex contention.
4. The variation in response times in v29 have been significantly
    reduced in v31.
5. If you look at the MTMalloc vs non-MTMalloc runs, you should note
    that the malloc mutexes are still a considerable issue, which needs
    to be addressed.


..Ian


Mime
View raw message