Date: Sun, 04 May 1997 20:43:33 -0500 From: Alex Nash <nash@mcs.com> To: "Eric J. Chet" <ejc@bazzle.com> Cc: freebsd-smp@FreeBSD.ORG, "Dan O'Brien" <dob@bazzle.com> Subject: Re: SMP performance (what am I doing wrong) Message-ID: <336D3B45.15FB7483@mcs.com> References: <Pine.BSF.3.96.970504202029.19017A-100000@gargoyle.bazzle.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Eric J. Chet wrote: > I'm not seeing any difference between a -current-up kernel > and a -current-mp kernel with rc5. On a current-up kernel I was > seeing 122k keys a second. On a -current-mp box with 2 rc5 > processes running I was seeing 2x122k keys a second. I was > running the rc5 clients with a nice value of 20. [ rc5 output ] > Try running the two rc5 clients if have a dual smp box. What might have > happend was your single rc5 client was bouncing between cpus and thats > where your performance lose came from. The current schedular doesn't > have any code to help process cpu affinity. I'm using a dual pentium-133 > box BTW. I did try two simultaneous rc5 clients (see my last mail for the differing results I got by running the same test twice). However, you just pointed me in the right direction: You ran the client in key cracking mode, as opposed to the benchmarking mode I had run it in. Mine ran for about tens of seconds, yours for tens of minutes... I ran the client the same way you did and watched the output in top. The cpuidle processes started out >50% CPU and then slowly decayed. After about 45 seconds (or thereabouts) they were consuming a few percent. Another 15 seconds later they dropped to 0%, and have stayed there ever since. I don't doubt I will get 2X throughput now :) Perhaps you could try this and see if you get similar results: Run 'top -S -I' and then start the two clients. Do you see a similar decay in idle proc CPU%? Alex
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?336D3B45.15FB7483>