On Sun, Apr 10, 2016 at 12:04:21PM +0200, Mike Galbraith wrote: > On Sat, 2016-04-09 at 15:05 -0400, Chris Mason wrote: > > > This does preserve the existing logic to prefer idle cores over idle > > CPU threads, and includes some tests to try and avoid the idle scan when > > we're > > actually better off sharing a non-idle CPU with someone else. > > My box says the "oh nevermind" checks aren't selective enough, tbench > dropped 4% at clients=cores, and 2% at clients=threads.
I knew this part would need more experimentation, so I kept v1 as simple as possible. On my box, tbench clients=cores is 5% faster, clients=threads is 4% faster. bounce_to_target() is a small version of task_hot(), I did get more accurate decisions by using the full task_hot(), so I can try that again. I'm testing on one of these: Intel(R) Xeon(R) CPU E5-2660 v2 @ 2.20GHz, which has two sockets and 10 cores per socket. What are you testing with? If it's two sockets or less I may be able to find one to reproduce with. > > > Benchmarks in production show overall capacity going up between 2-5% > > depending on the metric. > > Latency rules all loads certainly exist, and clearly want some love, > but the bigger the socket, and the more threads/core, the more that > traverse is gonna hurt the others, so seems either we need a better > filter, or a (yeah yeah, yet another damn) tweakable. > > Oh, and bounce_to_target() seems an odd way to say full_traverse. Sure, I can rename it. -chris
