From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756890Ab1JCQmm (ORCPT ); Mon, 3 Oct 2011 12:42:42 -0400 Received: from mx.binnacle.cx ([74.95.187.105]:59503 "EHLO mx.binnacle.cx" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751640Ab1JCQmd (ORCPT ); Mon, 3 Oct 2011 12:42:33 -0400 Message-Id: <6.2.5.6.2.20111003121925.03a83a28@binnacle.cx> X-Mailer: QUALCOMM Windows Eudora Version 6.2.5.6 Date: Mon, 03 Oct 2011 12:28:13 -0400 To: Eric Dumazet From: starlight@binnacle.cx Subject: Re: big picture UDP/IP performance question re 2.6.18 -> 2.6.32 Cc: linux-kernel@vger.kernel.org, netdev , Willy Tarreau , Peter Zijlstra , Stephen Hemminger In-Reply-To: <1317658588.2442.5.camel@edumazet-HP-Compaq-6005-Pro -SFF-PC> References: <6.2.5.6.2.20111003112108.03a83a28@binnacle.cx> <1317658588.2442.5.camel@edumazet-HP-Compaq-6005-Pro-SFF-PC> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org At 06:16 PM 10/3/2011 +0200, Eric Dumazet wrote: >> >Please send full "dmesg" output >> >> Attached. >> > >Minor note : > >It seems you use 4096 2MB hugepages, and your >machine has 1GB pages availabe, you could >try to use them. > >(boot cmd : hugepagesz=1GB hugepages=8 ) Interesting. I'll give it a try for sure, but since hugepages only gives a 5% improvement taking the page size up probably will not dent it by more than an additional percent or so. BTW I goofed on the kernel version number in my last post. It should be 2.6.39.4. That .27 was my mind transposing the 2.6.32 sub-version onto 2.6.39. Have one closing thought, which is to wonder if the 9.98% swapper [kernel.kallsyms] [k] native_safe_halt CPU is included in the 2.6.18(rhel) CPU accounting or not. If it's excluded that would make 2.6.39.4 look much better--possibly put it at near parity though I'm too lazy to do the math just now. My guess is that this number is something of an artifact of the clock-tick CPU histogram logic. Perhaps it's seeing the IP register parked in the idle-loop HALT instruction some of the time and counting it as non-idle time. I'd be happier if it either counted all idle time and showed the 45% that the poll=idle run shows, or none of the idle time at all--if it is in fact idle time. Unfortunately 'perf' is not available for 2.6.18(rhel) so the value cannot be compared.