Re: Armel userspace performance regression
On Wed, Aug 15, 2007 at 11:25:54AM +0300, Eugene Sanivsky wrote:
> I am experiencing serious performance regression with new armel
> libc/compiler.
> I see up to 45% degradation in Samba throughput for example.
Samba is a very complex piece of software, and you have changed every
single component related to it except the kernel. Do you have any less
complex banchmarks?
> Same Samba versions used for benchmarking (3.0.25b).
Did you compile samba with same options both times?
> Any ideas why?
Since I have not been blessed with gift of prescience, I only have
some guesses. Samba should not be cpu or integer calculation -bound,
but rather io-bound. In fact, samba should just use sendfile() to
page data out. Thus it would seem unlikely a compiler issue. I'd
guess:
- Samba ended up compiled/configured with different options. The
default samba in debian comes out with almost everything enabled by
default.
- Somehow cache ends up being used completly inefficently. If your
platform supports oprofile, it could help tracing this issue.
- There is something wrong with the nptl code in glibc2.5.
> Actions?
Try to find a more isolated test case or a synthetic benchmark
to reproduce the issue.
Reply to: