From mboxrd@z Thu Jan 1 00:00:00 1970 From: James Bottomley Subject: Re: next-20081030: voyager compile busted Date: Fri, 31 Oct 2008 12:42:14 -0500 Message-ID: <1225474934.3264.4.camel@localhost.localdomain> References: <20081030173030.d8847df6.sfr@canb.auug.org.au> <20081030180323.GA13399@x200.localdomain> <1225398287.19324.1.camel@localhost.localdomain> <20081030205230.GO27407@elte.hu> <1225400283.19324.6.camel@localhost.localdomain> <20081030210704.GA25665@elte.hu> <1225401708.19324.35.camel@localhost.localdomain> <490A2A30.5040905@zytor.com> <1225403266.19324.42.camel@localhost.localdomain> <490A2C3D.4020202@zytor.com> <20081030224057.GB20507@elte.hu> Mime-Version: 1.0 Content-Type: text/plain Content-Transfer-Encoding: 7bit Return-path: Received: from accolon.hansenpartnership.com ([76.243.235.52]:37905 "EHLO accolon.hansenpartnership.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752634AbYJaWBp (ORCPT ); Fri, 31 Oct 2008 18:01:45 -0400 In-Reply-To: <20081030224057.GB20507@elte.hu> Sender: linux-next-owner@vger.kernel.org List-ID: To: Ingo Molnar Cc: "H. Peter Anvin" , Alexey Dobriyan , Stephen Rothwell , linux-next@vger.kernel.org, linux-kernel@vger.kernel.org, Thomas Gleixner , Yinghai Lu On Thu, 2008-10-30 at 23:40 +0100, Ingo Molnar wrote: > * H. Peter Anvin wrote: > > > James Bottomley wrote: > > > > > > Yes ... been having IRC conversations about that. We'd need to use > > > runtime patching to fix the performance regressions virtualisation has > > > been causing us first ... but then we could use it for voyager. > > > > I thought we already were, at least to some degree (the call sites > > are way too big and way bigger than they need to be, so we end up > > with a lot of NOPs. I proposed a solution to Jeremy at Kernel > > Summit, but he basically said "I don't want to maintain that, I > > don't care about hardware performance", which is understandable but > > highly unfortunate.) > > in practice the function pointer overhead is almost unmeasurable for > things like IPIs which are rather expensive to issue. In fact it's > probably more expensive on Voyager as it does not utilize the local > APIC for SMP messaging (which is pretty much the only > performance-sensitive thing here). It uses its own glue logic it > appears, which is almost certainly behind the system bus. Heh ... don't be so sure: The VIC uses an array of up to 8 standard 8255 IRQ lines, which is slow. The QIC uses CPI trigger via cache line interference, which is far faster than APIC bus based systems (but is fairly similar to the way SAPIC systems work ... almost a decade after the original QIC system). > James, how many cycles do typical SMP ops take on Voyager? To be honest, I'm not too bothered about that. What bothers me is that if a normal system performance goes down because of the function pointer replacement (remember all x86 SMP systems will see an increase in function pointer usage because of this approach), then it probably wasn't worth it. However, the only way to be sure is to try it and get the benchmarks. James