[ConnClark]

Hi all,

I thought I would let you all know about an upcoming linux kernel feature. Its called Transparent Hugepage Support. It improves performance by using larger virtual memory pages supported by processors. This improves performance by reducing TLB misses and table walks. It has been recently incorporated into the linux-mm kernel and at some point soon make it into the main branch.

http://lwn.net/Articles/419933/

current code
http://git.kernel.org/gitweb.cgi?p=l....git;a=summary

I did a simple benchmark with it on the combustion/engineFoam/kivaTest tutorial with a single core on my Turion x2 processor in my laptop. I get a consistent 1% speed improvement over multiple tests.

time without it enabled

real 16m43.317s
user 16m41.611s
sys 0m1.460s


time with it enabled

real 16m32.794s
user 16m30.629s
sys 0m1.633s


Yes it only saved 10.5 seconds of run time but that was done without any changes to the application executable. I acknowledge that 1% isn't a lot but I would expect this to provide more of a boost on multicore test runs because there would be less memory bus contention. I would also expect better performance on more advanced processors since they have more huge page TLBs. Other cases also might show more of an improvement. (Note some software shows a 20% improvement)

Since some cases can take several hours to run (my record so far is 100 hours) I figure a 1% speed bump would be welcome.

I would be interested to see others performance test results on real systems.