From owner-freebsd-smp Tue Feb 26 12:13:30 2002 Delivered-To: freebsd-smp@freebsd.org Received: from smtp.whitebarn.com (Spin.whitebarn.com [216.0.13.113]) by hub.freebsd.org (Postfix) with ESMTP id 1E86137B428 for ; Tue, 26 Feb 2002 12:12:46 -0800 (PST) Received: from NewStorm.whitebarn.com (NewStorm.whitebarn.com [216.0.13.77]) by smtp.whitebarn.com (8.9.3/8.9.3) with ESMTP id OAA44394; Tue, 26 Feb 2002 14:12:38 -0600 (CST) (envelope-from Bob@Talarian.Com) Subject: Re: Performance vs. Stable From: Bob Van Valzah To: Garance A Drosihn Cc: smp@FreeBSD.ORG In-Reply-To: References: <1014749196.231.48.camel@NewStorm.WhiteBarn.Com> Content-Type: text/plain Content-Transfer-Encoding: 7bit X-Mailer: Evolution/1.0.2 Date: 26 Feb 2002 14:12:38 -0600 Message-Id: <1014754359.231.77.camel@NewStorm.WhiteBarn.Com> Mime-Version: 1.0 Sender: owner-freebsd-smp@FreeBSD.ORG Precedence: bulk List-ID: List-Archive: (Web Archive) List-Help: (List Instructions) List-Subscribe: List-Unsubscribe: X-Loop: FreeBSD.org Yes, good point, sorry. I shouldn't've glossed over that. The short answer is that I used GENERIC kernels with as few changes as possible. So on -CURRENT, the kernel I tested with was running INVARIANTS, INVARIANT_SUPPORT, WITNESS, and WITNESS_SKIPSPIN. These were not set on my test -STABLE kernel since they're not set in GENERIC there. If these settings are causing large differences in timings, I think it'll show up when I compare kgmon/gprof results between kernels. I can also just compile a kernel without these settings and see how it differs. I'm still very new to this so I'm trying to map the benchmarking space and see where there are interesting questions. You've opened my eyes to an area I'd missed but should've seen. Are there others? I can also see interesting questions around the impact of options SMP and APIC when running workloads that don't benefit from MP hardware. I'd also like to get some numbers for "macro" benchmarks that might benefit from MP hardware. Finally, I'd like to be able to track progress over time. Bob On Tue, 2002-02-26 at 13:40, Garance A Drosihn wrote: > At 12:46 PM -0600 2/26/02, Bob Van Valzah wrote: > >I have run lmbench on 4.5-STABLE and 5.0-CURRENT (as of late > >last week). I believe this might give some insight into the > >relative performance of these kernels on uniprocessor micro > >benchmarks. > > The first thing you need to include with any kind of benchmark > is the kernel settings you are running with. I'm not saying > that you have anything set wrong, but it is important to list > the settings just so everyone is aware that the settings make > a big difference on performance. > > In particular, the settings for: > INVARIANTS > INVARIANT_SUPPORT > WITNESS > WITNESS_SKIPSPIN > in both systems. There's also the question of malloc options, > which I must admit I don't quite understand how the settings > are different between -stable and -release, or how to change > them. But they do effect almost any kind of serious benchmark > attempts. > > -- > Garance Alistair Drosehn = gad@eclipse.acs.rpi.edu > Senior Systems Programmer or gad@freebsd.org > Rensselaer Polytechnic Institute or drosih@rpi.edu To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-smp" in the body of the message