Ramine
2/17/2015 1:39:00 AM
Hello,
As you have noticed i have implemented a parallel sort library and
i have done some benchmark on it, and as you have noticed it can
not scale much than 8X on multicore processors, the reason is simple,
cause by nature the parallel sorts algorithms of my parallel sort
library are limited by the serial part of the Amdahl's law that is
the memory transfers from the memory to the L2 caches, so what is
the solution then to make my parallel sort library to scale more on
multicore CPUs ? you have to made it NUMA-aware and to make it scale
on NUMA by parallelizing the memory transfers from the memory to the L2
caches, this way it will scale on NUMA architectue , so i have
decided to make this my next project, i will soon make my
parallel sort library NUMA-aware and i will then make it scale more and
more on NUMA architecture with more and more NUMA nodes, so as you see
i will render my parallel sort library a very powerful parallel sort
library, so, the sorting part and the merging part will be fully
parrallelized and fully scalable on NUMA architecture on my next
parallel sort library , that's a really good news !
Thank you for your time.
Amine Moulay Ramdane.