On Fri, Mar 18, 2016 at 6:02 PM, Michal Mazurek <akf...@jasminek.net> wrote: > BFS has one shared queue for all CPUs, maybe there is a very good reason > for that, we'll see.
Michal, first of all congrats to optimistic results in interactive workloads. Honestly I'm a little bit worried about your attempts since I think if doing any scheduler modernization then it shall also consider to be NUMA-aware or at least take NUMA into consideration. The problem is that NUMA is no longer domain of high-cpu sockets boxes, but it's also available or possibly to be available in 1 socket Haswell boxes when configured right. Intel calls this "cluster-on-die" for your information. Anyway, although worried, I'm still curious about your hacking in scheduler, so I keep my fingers crossed! Thanks, Karel