On Jan 26, 7:22 am, parisse <bernard.pari...@ujf-grenoble.fr> wrote: > > For dense problems the answer is tentatively yes, however you > > can also shrink the size of the heap. See the "chaining" section > > in http://www.cecm.sfu.ca/~rpearcea/sdmp/sdmp_div.pdf > > The details of what may be faster or not will depend on your > > implementation. > > I'm already using chaining following your previous paper, and I guess > you also do in parallel sdmp, but you get a noticeable superlinear > improvement, hence the natural question: in this situation how much > threads should you launch if you have n processors? (it should be more > than n)
I launch n threads for n cores. Otherwise context switches will occur which drain performance. Also, you can't synchronize frequently when you have more threads than cpus because then you may have to wait more than one timeslice. The paper describes the strategy, which is to manually split up the global work (merging the result of the threads). The threads cooperate to perform this work. At the same time, it is very important to allow the threads to be as independent as possible. --~--~---------~--~----~------------~-------~--~----~ To post to this group, send email to sage-devel@googlegroups.com To unsubscribe from this group, send email to sage-devel-unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/sage-devel URLs: http://www.sagemath.org -~----------~----~----~----~------~----~------~--~---
