From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ben Greear Subject: Re: [Patch net-next] tcp: add a global sysctl to control TCP delayed ack Date: Thu, 04 Apr 2013 16:39:04 -0700 Message-ID: <515E0F18.8080705@candelatech.com> References: <1365070560-11544-1-git-send-email-amwang@redhat.com> <20130404224810.GE23056@order.stressinduktion.org> <20130404.192539.1157364077083658871.davem@davemloft.net> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: hannes@stressinduktion.org, amwang@redhat.com, netdev@vger.kernel.org, eric.dumazet@gmail.com, rick.jones2@hp.com, shemminger@vyatta.com, tgraf@suug.ch, David.Laight@ACULAB.COM To: David Miller Return-path: Received: from mail.candelatech.com ([208.74.158.172]:34668 "EHLO ns3.lanforge.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1763641Ab3DDXjp (ORCPT ); Thu, 4 Apr 2013 19:39:45 -0400 In-Reply-To: <20130404.192539.1157364077083658871.davem@davemloft.net> Sender: netdev-owner@vger.kernel.org List-ID: On 04/04/2013 04:25 PM, David Miller wrote: > From: Hannes Frederic Sowa > Date: Fri, 5 Apr 2013 00:48:10 +0200 > >> I totally understand the objections that were given regarding this >> patch. But for defense of this patch we also provide a knob to disable >> slow start after idle, which from my point of view is as "evil" as >> this change. > > I completely disagree, slow start after idle is way too aggressively > throwing past history away, so turning that off is much safer. I've been carrying a per-socket way to control delayed ack in my tree for a while, and it is a big performance gain on wifi networks where the network is basically half-duplex. If I recall correctly, it's worth 50+Mbps throughput improvement in some cases. If it matters, I can post more detailed numbers. There are drawbacks when you set delayed ack too high: The ramp-up time for TCP takes a good bit longer. But, some applications may want to trade slower startup time for better bulk transport, and at moderate delayed-ack values, the ramp up time is not noticeably impaired. Thanks, Ben -- Ben Greear Candela Technologies Inc http://www.candelatech.com