Re: [PATCH] again: Re: Athlon kernel crash (i686 works)

[Dan Maas]

> I've run several LMbench tests against both 686-optimized
> and Athlon-optimized kernels.  The results waver across multiple
> tests, one kernel winning some tests one time and losing the next,
> but the values are all close.

The benefits of the kernel Athlon optimizations are higher memory bandwidth
for bulk copies/clears and less cache pollution. But LMbench isn't going to
show any difference, because its tests use generic x86 mem*() functions, not
Athlon-optimized SSE memory routines like in the Athlon kernel.

*Local* Communication bandwidths in MB/s - bigger is better
Host                OS  Pipe AF    TCP  File   Mmap  Bcopy  Bcopy  Mem   Mem
                             UNIX      reread reread (libc) (hand) read
write
--------- ------------- ---- ---- ---- ------ ------ ------ ------ ---- ----
-
Athlon-1  Linux 2.4.10- 847. 685. 311.  332.4  501.3  176.2  206.2 471.
342.5
Athlon-2  Linux 2.4.10- 882. 586. 187.  331.6  510.2  177.6  207.1 484.
343.5
i686-1    Linux 2.4.10- 863. 586. 299.  320.0  510.2  176.3  206.6 472.
342.6
i686-2    Linux 2.4.10- 874. 318. 199.  319.6  520.2  177.7  206.8 486.
343.5

It should be obvious that LMbench uses sub-optimal memory routines, since
the numbers for "Bcopy" and "Mem read/write" bandwidth are so much lower
than pipe and AF UNIX bandwidths! (the pipe/UNIX tests are basically
equivalent to Bcopy, plus extra user<->kernel transitions and context
switches).

The only cases where I'd expect the Athlon kernel to do better on LMbench
are essentially kernel memcpy() benchmarks - pipe and AF UNIX bandwidths.
I'm not sure if the kernel pipe and UNIX socket code actually uses
Athlon-optimized routines; in any case the small buffer sizes (eg 4KB for
pipes) could be hiding any performance gain.

Regards,
Dan