Trial patch for page granuality hint support to 2.4.0-test1

Trial patch for page granuality hint support to 2.4.0-test1

Post by Naohiko Shimi » Sat, 24 Jun 2000 04:00:00



Hi,
I made a patch for support the Alpha's page granuality hint bits.
I think it will help many HPC applications.

This patch is only for performance testing, and I am working on
real kernel patch for general use.

But before I go in depth, I want to evaluate the performance
of this patch.  Then I need testers for the patch.

Because my first patch is for performance evaluation, there are some
restrictions.

 1) It only works when the memory is locked with mclockall(MCL_FUTURE).
 2) It only works when the heap area is allocated with brk system call.
 3) The kernel seems not to release the locked memory even when the
    program terminated normally, then you must reboot to release the memory.
    Is it kernel bug, or something wrong with my patch? Anyway the new patch
    for general use will not require the memory lock.

Who can run 2.4.0-test1 kernel, please help me with the benchmark.
I only tested with my EB164-300MHz, and got the following result for
the memory thoughput of a large matrix transpose.

     No GH area         GH area
        45MB/S          55MB/S

The patch and benchmark will be in the following URL:

http://shimizu-lab.et.u-tokai.ac.jp/~nshimizu

Thank you in advance,

Naohiko Shimizu
Dept. Communication Engr./Univ. TOKAI
1117 Kitakaname Hiratsuka 259-12 Japan
TEL.+81-463-58-1211(ext. 4084) FAX.+81-463-58-8320
http://shimizu-lab.et.u-tokai.ac.jp/

 
 
 

Trial patch for page granuality hint support to 2.4.0-test1

Post by Tony Hammit » Sat, 24 Jun 2000 04:00:00


Cool!

Thanks for the effort.  This should be pretty effective.

Keep up the good work!

Regards,

Tony Hammitt


> Hi,
> I made a patch for support the Alpha's page granuality hint bits.
> I think it will help many HPC applications.

> This patch is only for performance testing, and I am working on
> real kernel patch for general use.

> But before I go in depth, I want to evaluate the performance
> of this patch.  Then I need testers for the patch.

> Because my first patch is for performance evaluation, there are some
> restrictions.

>  1) It only works when the memory is locked with mclockall(MCL_FUTURE).
>  2) It only works when the heap area is allocated with brk system call.
>  3) The kernel seems not to release the locked memory even when the
>     program terminated normally, then you must reboot to release the memory.
>     Is it kernel bug, or something wrong with my patch? Anyway the new patch
>     for general use will not require the memory lock.

> Who can run 2.4.0-test1 kernel, please help me with the benchmark.
> I only tested with my EB164-300MHz, and got the following result for
> the memory thoughput of a large matrix transpose.

>      No GH area         GH area
>         45MB/S          55MB/S

> The patch and benchmark will be in the following URL:

> http://shimizu-lab.et.u-tokai.ac.jp/~nshimizu

> Thank you in advance,

> Naohiko Shimizu
> Dept. Communication Engr./Univ. TOKAI
> 1117 Kitakaname Hiratsuka 259-12 Japan
> TEL.+81-463-58-1211(ext. 4084) FAX.+81-463-58-8320
> http://shimizu-lab.et.u-tokai.ac.jp/