From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id ; Sat, 29 Sep 2001 13:23:18 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id ; Sat, 29 Sep 2001 13:23:08 -0400 Received: from fenrus.demon.co.uk ([158.152.228.152]:20354 "EHLO fenrus.demon.nl") by vger.kernel.org with ESMTP id ; Sat, 29 Sep 2001 13:22:57 -0400 From: arjan@fenrus.demon.nl To: jalvo@mbay.net (John Alvord) Subject: Re: kernel changes cc: linux-kernel@vger.kernel.org In-Reply-To: X-Newsgroups: fenrus.linux.kernel User-Agent: tin/1.5.8-20010221 ("Blue Water") (UNIX) (Linux/2.4.3-6.0.1 (i586)) Message-Id: Date: Sat, 29 Sep 2001 18:23:15 +0100 Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org In article you wrote: > One aspect that bothers me is the absence of a success criteria. I disagree here. Red Hat uses "must pass the cerberus test" as one of the criteria for kernels. The are other similar criteria, most are obvious (must boot :). All other distributions have similar tests and a few even use the cerberus testsuite as well. Maybe your problem is "absence of tests before Linus releases", well even that isn't fully true as distros run these tests on -pre kernels as well (or -ac kernels, which are mostly in sync with -pre kernels)... > The current competition for best VM is a good example. The fact is that > every operating system will fail with a high enough load. The best you can > hope for is a better degradation then the prior release. There are a few basic creteria here as well, and 2.4.10 fails on some of them so far: 1) Must not kill processes as long as there is plenty of swap or (possibly dirty) cache memory 2) Must not deadlock (as that is a code-bug) 3) Must not livelock without any progress Note that no 2.4 kernel so far really achieves 1) in the presence of highmem; the obvious deadlocks are just pushed further by tuning. > At the moment both 2.4.10 and 2.4.9-ac16 are better then 2.2.19. But > people keep testing under higher and higher loads and (surprise) they both > fail... initiating a search for better degradation logic. 2.4.10 isn't better than 2.2.19 given the criteria above. 2.4.10aa2 might be though... and 2.4.9acX+Rik's patches are solid in testing. Greetings, Arjan van de Ven