Use of large memory pages possible with newer linux kernels

[Dresdenboy]

http://kerneltrap.org/node/view/418

That sounds very interesting! And it offers new possibilities for efficient FFT implementations.

[Dresdenboy]

Ok, some more details about the use of these large pages:

Oracle managed to get a 8% speedup by using the large pages. Although I have little experience in this area I think for FFTs the speedup will be much larger, because:

• Even if data is already in the L1 cache the accessing time can increase if the memory addresses of these data are actually spread over many memory pages.
• The limited amount of TLB entries requires fine tuning of FFT algorithms to avoid TLB thrashing as much as possible - but this avoidance could cause less efficient algorithms.
• why is it so hard for large size FFTs to come at least close to the FFT MFLOPS for FFTs running completely inside L1 (or L2) cache in times of memory prefetching?
• I need at least 2 mem-read/write passes to do a large size FFT - but todays max transfer rates for P4/Opteron/AFX systems (6.4GB/s = reading up to 750 times the 1024k FFT data set per second) is hardly reachable because it drops significantly for large strides.

I roughly estimate that at least a speedup of 10-30% could be possible.

Older discussions regarding this topic can be found here:
http://www.mail-archive.com/mersenne.../msg07117.html

[Dresdenboy]

Here is a paper which gives some numbers for achieved speedups by using "superpages": http://portal.acm.org/citation.cfm?i...l=GUIDE&dl=ACM

(I think the pdf can only be seen by people which have a subscription)

Here some results of memory intensive benchmarks with some impressive examples (together with the average speedups because in most cases the speed increase is <10%):

SPEC benchmark    Speedup by using superpages

vpr                      38.3%
mcf                      67.6%
vortex                   11.2%
bzip2                    14.0%
average for SPECint      11.2%

galgel                   28.9%
art                      12.2%
lucas                    28.0% :cool:
apsi                     82.7%
average for SPECfp       11.0%


and some non-SPEC benchmarks:
FFTW                     54.9% :grin:
Matrix                  654.6% :shock:

They implemented it in FreeBSD on the Alpha CPU.

I think these numbers tell enough to justify any GIMPS-related research in this area - especially as it is available with Linux kernel versions 2.5.36+ and 2.6. Actually (for first tests) it just need a change of the memory allocating and releasing system calls.

[Dresdenboy]

Here the paper (and other stuff) is accessible for everyone:

http://www.cs.rice.edu/~jnavarro/superpages/