benchmark results of page granurality patch
The kernel patch http://shimizu-lab.et.u-tokai.ac.jp/~nshimizu makes
HPC applications faster. I ran some benchmark with pached kernel.
The benchmark results on EB164:
- A 1000x1000 matrix transpose: 41% faster than the normal
- 1000x1000 matrix to matrix multiply 22% faster than the normal
(This is my code not the ATLAS nor CXML, they uses sub-block copy
to reduce the translation misses and the effectiveness will be
little [with CXML it was only 2% faster with my patch])
- 131072 point FFT 97% faster than the normal
Fortran users should be care, this patch will not accelerate the
performance for the "common" area. Because many fortran compilers
set the protection information of the area as read_only at the
initial allocation, and then at the access exception it will turn
on the write_enable of the ONE page. It is difficult for OS to
detect that the block of the pages can be turned on the flag.
--------------
BTW, why 2.4.0-test? kernel crashes on the mount?
Is there any solution on the mount problem?
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
Reply to: