Skip to contentSkip to conversation

again: Re: Athlon kernel crash (i686 works)

PATCH Dan Maas 2 participants 2 messages Oct 10, 2001 1 message not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Dan Maas

author
#1 lore raw
Quoted 4 lines

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

#2 lore raw
Quoted 4 lines
Quoted 4 lines

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.

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

There are no SSE optimizations (yet).
There are prefetch/movntq tricks.
Optimized fast_clear_page() is 3x faster than normal one, optimized fast_copy_page() is 1.5x faster than normal one. (roughly, it depends on your mem and CPU MHz)

I can mail a test program to you if you are curious.

Signature · 2 lines
Best regards, VDA
mailto:[email protected]