Hi!
Some of you may recall Ian Smith's post about low benchmark scores for
openSuse 10.3 [1]. Now that the Suse team has fixed one of the main
issues, it is time to benchmark again. This time, openSuse 10.3 scores
_much_ better than the last time.
Further investigation of the bad performance revealed that the main
cause for the extremely low performance was very high overhead for
system calls (syscalls), so I filed bug #333739. Tony Jones had a look
at it, fixing both kernel and auditd. The bug is closed now, new
packages were released a while ago.
Another source of bad benchmark results is grep. The version of grep
included in Suse 10.3 is awfully slow when processing UTF-8 data, see
bug #308698 [3]. On my system, grep on Suse 10.3 is ~30 times slower
than on 10.0. Exact results may vary depending on setup, but it is
slooow. Grep was also a source of inconsistent benchmark results with
unixbench 5.1, because that version of unixbench did not set the LANG
environment variable consistently. Unixbench 5.1.1 fixes that.
I decided to run another benchmark, comparing openSuse 10.3 with new
kernel+auditd and a fast grep against one with the old kernel+auditd and
the default grep (as originally benchmarked by Ian Smith). As a fast
grep, I simply took the rpm for Suse 10.2. As a reference, I have
included results for Suse Linux 10.0. All tests were run in runlevel 1
with unixbench 5.1.1 (identical unixbench binaries compiled under Suse
10.0) on a Pentium M 1.3Ghz.
10.3 new 10.3 old 10.0
========== ======== ==========
Dhrystone 247 247 242
Whetstone 177 177 177
Execl 642 603 714
File Copy 1024 560 484 538
File Copy 256 432 366 454
File Copy 4096 809 742 597
Pipe Throughput 374 295 479
Context Switch 473 394 569
Process Creat 821 793 1032
Shell Scripts1 524 358 556
Shell Scripts8 522 349 537
System Call 772 336 882
------ ------ ------
Index Score: 485 393 513
The current version of openSuse 10.3 easily outscores the GM version.
The modified version of grep is responsible for higher scores in the
shell script tests since they use grep as one of their commands. It
should not affect the other test results at all. Maybe Suse should
consider to downgrade grep if the performance problems cannot be
resolved, because the new grep is slow to a point where it is simply
unusable.
Fixes in kernel+auditd significantly helped all tests that interact a
lot with the kernel (basically all but the first two). Syscall
performance is measured directly in the last test, where the score has
more than doubled!
However, comparing results to 10.0 shows that overall, the performance
of the base system is still not as good as it could be, at least in this
benchmark. Suse 10.0 still scores higher in most areas. Since the
syscall test scores are ~12% lower than in 10.0, it seems logical that
the slowdown is due to the kernel. It would be interesting to see if the
vanilla kernel suffers from the same slowdown or if this is specific to
Suse.
Regards
nordi
--
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
