Re: Slow Xorg performance on dual Opteron + Radeon, Jessie 64-bit

[Chris Edwards <cedwards.debian@cedwards.geek.nz>]

Thanks for the ongoing suggestions.  Some more testing and results:

`grep render /var/log/Xorg.0.log`:
[  8898.907] (II) RADEON(0): Direct rendering enabled

`glxinfo | grep render` gives:
direct rendering: Yes
OpenGL renderer string: Gallium 0.4 on ATI R420

I've tested with an xorg.conf and without.  Without, it comes up in 
1024x768 resolution and the problem seems less severe, but it's still 
noticeable, and Xorg still spends an awful lot of time in 
__memcpy_sse2_unaligned.

While doing some further reading and searching, I discovered that Debian 
only recently switched back to using glibc, after using eglibc since 
about 2009.  Since my first suspicions were around libc and possible 
performance regressions there, I thought I'd take a closer look, using 
the following test program:

<<
#include <string.h>
#include <stdlib.h>

const int SIZE=1048576 * 1024;
const int REPS=20;

int i;

void main() {
        void *chunk_a = malloc(SIZE);
        void *chunk_b = malloc(SIZE);

        for (i = 0; i < REPS; i++)
                memcpy(chunk_b, chunk_a, SIZE);

        free(chunk_a);
        free(chunk_b);
}
>>

I found a .deb of glibc-2.17 for comparison (__memcpy_sse2_unaligned was 
added between glibc-2.17 and glibc-2.18) and used LD_PRELOAD to load 
that version instead of Jessie's native glibc-2.19.

time LD_PRELOAD="/tmp/lib/x86_64-linux-gnu/ld-2.17.so 
/tmp/lib/x86_64-linux-gnu/libc-2.17.so" ./memcpy-benchmark

real    0m12.303s
user    0m11.456s
sys     0m0.844s


And using the system libc:

time ./memcpy-benchmark

real    0m22.509s
user    0m21.504s
sys     0m0.992s


So Jessie's standard libc-2.19 is about half as fast as 2.17 on the same 
hardware and otherwise same OS.  I don't know if this accounts entirely 
for the poor performance I'm seeing, but it probably doesn't help. 
Might this be worth taking up with the GNU libc people?  My guess is it 
might be due to slow SSE2 execution on this old AMD64 CPU, whereas the 
glibc code is probably optimised for much newer hardware.

Thanks again,
Chris