On Fri, Nov 15, 2013 at 12:35:22AM +0400, Alexander Pakhomov wrote: > > dd unenc sp: > sys 17% int 9% > 164 MB/s > > dd_unenc_mp > 30 sys 50 int > 200 MB/s
this roughly shows that a lot of CPUs time is wasted in interrupts in the MP case, probably spinning to grab the kernel_lock held by the other CPU. > dd_enc_sp > 100 sys 0 inter > 54 MB/s > > dd_enc_mp > 60x4 sys 80 int > 42 MB/s > same waste of CPU time, to the point that using two CPUs even decreases the performace. [...] there are use-cases where a single CPU works better than multiple CPUs. I don't know how these block devices work though. thank you for your tests. -- Alexandre
