Extremely poor performance crunching random numbers under PIV-FC5

[Andy Green]

BankHacker wrote:
>> Ahaha, so if I understand you the strace output is huge for the dynamic
>> one and smaller for the static one?
> 
> Yes, but it is no huge, just a little bit bigger. Only several lines
> greater, those I wrote in the previous message.
> 
> 
>> Can you post an example of one of the "10 million" syscalls that is
>> present in the strace output for the dynamic case (if I understood you
>> correctly)?
> 
> There is no trace for individual syscalls (supposing rand() function
> is really a syscall ...) in the trace output.

I think the news is that rand() does not generate a syscall, or we 
should expect to see it in the output of strace.  Jakub says it is just 
a call to a usermode function in libc too.

>> echo "0" >/proc/sys/kernel/randomize_va_space
>> echo "0" >/proc/sys/kernel/exec-shield
> I am afraid the problem is still present.

Oh well just a guess since the difference seems to hinge on whether the 
library is loaded in or not.

>     gcc test-cpu-2.c -o libm-test-cpu-2 /usr/lib/libm.a -lm -O3

>     10 M de rand() en 44.800 sec (example.: 3392997)
>     10 M de sqrt(i) en 0.180 sec (example.: 3162)
>     10 M de log(i) en 0.920 sec (example.: 16)
>     10 M de log10(i) en 0.950 sec (example.: 6)
>     ...
> 
> So identical results in rand() but worst results in other mathematical
> functions.

It seems that maybe we can just say the whole problem comes down to that 
rand() from dynamically linked libc.so.6 is slower to execute than what 
is supposed to be the same code statically linked.

> Doing top meanwhile Fedora is executing
> (static/dynamic/libm)_test-cpu-2 gives always a load of 1.05 more or
> less. So the CPU is working at full throttle in every case.

If you can't get random_r() to work (see eg

http://refspecs.freestandards.org/LSB_1.2.0/gLSB/baselib-random-r-3.html

) I think what I would do here is open /dev/urandom and read stuff from 
it, see how that performs.  Maybe you can read 4K bytes for example from 
it into a buffer to limit syscall overhead, and move an int * along the 
buffer.

-Andy


-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/x-pkcs7-signature
Size: 4492 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://listman.redhat.com/archives/fedora-list/attachments/20060518/319d1703/attachment-0001.bin>