From mboxrd@z Thu Jan 1 00:00:00 1970 From: Hans de Goede Subject: Re: cat /proc/net/tcp takes 0.5 seconds on x86_64 Date: Wed, 27 Aug 2008 11:14:17 +0200 Message-ID: <48B51AE9.5080709@hhs.nl> References: <200808261549.m7QFnVUN032543@bz-web1.app.phx.redhat.com> <20080826163719.GA25066@redhat.com> <48B44C3F.6020006@cosmosbay.com> <48B452ED.1000308@hhs.nl> <48B469EA.1010807@cosmosbay.com> <48B46E89.4030104@hhs.nl> <48B47543.8080701@cosmosbay.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: Dave Jones , netdev@vger.kernel.org To: Eric Dumazet Return-path: Received: from smtp2.versatel.nl ([62.58.50.89]:33063 "EHLO smtp2.versatel.nl" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752214AbYH0JDH (ORCPT ); Wed, 27 Aug 2008 05:03:07 -0400 In-Reply-To: <48B47543.8080701@cosmosbay.com> Sender: netdev-owner@vger.kernel.org List-ID: Eric Dumazet wrote: > > I dont know, you tell us 50x, but nowhere I saw your numbers on i386, > nor the amount of memory of your test machine. > My machine has 2 Gb of memory, but that is not really relevant as I'm not the reporter of the problem (but I can reproduce it) I'm only in the loop because I maintain the gkrellm package in Fedora which exhibits this problem. Also I gave you no i386 numbers on the same machine because I don't have an i386 install readily available on this machine. Now you can spin this anyway you want, but lets cut to the chase, gkrellmd which is a system monitoring daemon reads /proc/net/tcp and /proc/net/tcp6 3 times a second, with slightly older kernel this used to give a CPU load of 2-3% on x86_64 now it gives a load of 50-70% (on my test machine with a whole 8 tcp connections open). Now you can wing it any way you want, but this is a serious regression. I don't want to go to lkml and start shouting regression REGRESSION regression, but given the treating and enormous amount of help I've received sofar (and remember I'm only the messenger I didn't write nor use gkrellmd) I'm tending towards starting shouting regression on lkml, as atleast there this seems to get some attention. Regards, Hans