From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ben Greear Subject: Socket send-buffer auto-sizing Date: Thu, 07 Jun 2012 10:59:29 -0700 Message-ID: <4FD0EC01.3020300@candelatech.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit To: netdev Return-path: Received: from mail.candelatech.com ([208.74.158.172]:60095 "EHLO ns3.lanforge.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932414Ab2FGR7a (ORCPT ); Thu, 7 Jun 2012 13:59:30 -0400 Received: from [192.168.100.111] (firewall.candelatech.com [70.89.124.249]) (authenticated bits=0) by ns3.lanforge.com (8.14.2/8.14.2) with ESMTP id q57HxTuC032557 (version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO) for ; Thu, 7 Jun 2012 10:59:29 -0700 Sender: netdev-owner@vger.kernel.org List-ID: I'm continuing to test one-way tcp streams in 3.5.0-rc1 on a wifi network. When I do not specify a send buffer size, and thus use the kernel defaults, max speed is about 77Mbps. When I specify 512KB send-buffer, I get speeds up to 185Mbps. When set to 1MB, I get about 198Mbps (and setting higher does not increase the throughput after this). This is without any 'delack' patches applied. My question is: Should the kernel auto-tuner work better? I seem to recall a comments from some years ago that applications should no longer attempt to tune send/recv buffers because the kernel was smart enough to get it at least mostly right. Thanks, Ben -- Ben Greear Candela Technologies Inc http://www.candelatech.com