From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: RPS runs with multiqueue NIC Date: Tue, 22 Feb 2011 06:46:03 +0100 Message-ID: <1298353563.3360.6.camel@edumazet-laptop> References: <4A6A2125329CFD4D8CC40C9E8ABCAB9F24D3DE6CC5@MILEXCH2.ds.jdsu.net> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: "netdev@vger.kernel.org" To: Jon Zhou Return-path: Received: from mail-bw0-f51.google.com ([209.85.214.51]:33489 "EHLO mail-bw0-f51.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751711Ab1BVFyS (ORCPT ); Tue, 22 Feb 2011 00:54:18 -0500 Received: by bwz10 with SMTP id 10so2871680bwz.10 for ; Mon, 21 Feb 2011 21:54:17 -0800 (PST) In-Reply-To: <4A6A2125329CFD4D8CC40C9E8ABCAB9F24D3DE6CC5@MILEXCH2.ds.jdsu.net> Sender: netdev-owner@vger.kernel.org List-ID: Le lundi 21 f=C3=A9vrier 2011 =C3=A0 19:07 -0800, Jon Zhou a =C3=A9crit= : > Hi >=20 > What will happen if RPS runs on kernel 2.6.36.4 machine where a multi= queue NIC intel x520 installed > Will it decrease or improve performance? It depends on the workload. =46or typical TCP flows, hardware multiqueue is generally better, especially if the NIC supports the ethtool -X command (so that you can tune effective number of cpus dedicated to the softirq processing). Packet are delivered directly to the CPU target, without taking an extr= a IPI. =46or UDP workloads, adding RPS might be a win, because many NIC dont t= ake into account ports numbers in their hash computation, while net/core/dev.c hash function does. If you receive trafic coming from a single remote machine, RPS helps to spread flows between several cpus.