Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751897AbWJMUyU (ORCPT ); Fri, 13 Oct 2006 16:54:20 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751902AbWJMUyT (ORCPT ); Fri, 13 Oct 2006 16:54:19 -0400 Received: from 1wt.eu ([62.212.114.60]:36868 "EHLO 1wt.eu") by vger.kernel.org with ESMTP id S1751897AbWJMUyT (ORCPT ); Fri, 13 Oct 2006 16:54:19 -0400 Date: Fri, 13 Oct 2006 22:52:47 +0200 From: Willy Tarreau To: "H. Peter Anvin" Cc: Alistair John Strachan , Kay Sievers , Linus Torvalds , Linux Kernel Mailing List Subject: Re: Linux 2.6.19-rc2 Message-ID: <20061013205247.GG5050@1wt.eu> References: <200610131840.28411.s0348365@sms.ed.ac.uk> <452FD305.6070902@zytor.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <452FD305.6070902@zytor.com> User-Agent: Mutt/1.5.11 Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 1877 Lines: 44 On Fri, Oct 13, 2006 at 10:55:17AM -0700, H. Peter Anvin wrote: > Alistair John Strachan wrote: > >On Friday 13 October 2006 17:49, Linus Torvalds wrote: > >>Ok, it's a week since -rc1, so -rc2 is out there. > > > >Does anybody know what's up with the git server? Hopefully it's just my > >connection... > > > >[alistair] 18:38 [~/linux-git] git pull > >fatal: unexpected EOF > >Fetch failure: > >git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git > > > > No, this is the result of a serious problem with gitweb. > > We run gitweb behind a cache (otherwise it would be unacceptably > expensive), but when httpd starts timing out on gitweb, it spawns gitweb > over and over and over again, and the load on the machine skyrockets, > throttling other services. > > This happens every time we're on one server instead of two (one server > is down right now for network rewiring.) > > I think for now I'm just going to put a loadavg cap on running gitweb... I encountered a similar problem (to a far lower scale) when putting gitweb on my poor parisc server behind my ADSL line. The machine used to OOM several times a week. So I've set up an haproxy instance in front of it with a limit on the number of concurrent connections per backend. All excess connections are queued and served as soon as one slot frees up. The machine has never crashed since. Would you be interested in some help in trying to set it up ? Assuming that epoll is available, I have absolutely no worries about the load. As an added benefit, it could also provide HA and LB but I understand it's not the main concern right now. Willy - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/