From mboxrd@z Thu Jan 1 00:00:00 1970 From: Alexander Duyck Subject: Re: [net-next 01/11] ixgbe: Do not use DCA to prefetch the entire packet into the cache Date: Tue, 13 Nov 2012 09:41:38 -0800 Message-ID: <50A28652.6020505@intel.com> References: <1352815405-751-1-git-send-email-jeffrey.t.kirsher@intel.com> <1352815405-751-2-git-send-email-jeffrey.t.kirsher@intel.com> <1352816011.6185.33.camel@edumazet-glaptop> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit Cc: Jeff Kirsher , davem@davemloft.net, netdev@vger.kernel.org, gospo@redhat.com, sassmann@redhat.com To: Eric Dumazet Return-path: Received: from mga11.intel.com ([192.55.52.93]:63447 "EHLO mga11.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752108Ab2KMRlj (ORCPT ); Tue, 13 Nov 2012 12:41:39 -0500 In-Reply-To: <1352816011.6185.33.camel@edumazet-glaptop> Sender: netdev-owner@vger.kernel.org List-ID: On 11/13/2012 06:13 AM, Eric Dumazet wrote: > On Tue, 2012-11-13 at 06:03 -0800, Jeff Kirsher wrote: >> From: Alexander Duyck >> >> The way the code was previously written it was causing DCA to prefetch the >> entire packet into the cache when it was enabled. That is excessive as we >> only really need the headers. >> >> We are now prefetching the headers via software so doing this from DCA would >> be redundant anyway. So clear the bit that was causing us to prefetch the >> packet data and instead only use DCA for the descriptor rings. >> >> Signed-off-by: Alexander Duyck >> Tested-by: Phil Schmitt >> Signed-off-by: Jeff Kirsher >> --- > Excellent ! > > My own ixgbe cards are moving so I cant test this, do you guys have some > numbers to share ? > > Thanks In my tests I saw no real change because of the DCA changes. I kind of suspected that would be the case as mentioned in the patch the prefetch and DCA were both doing the same thing so by dropping the extra DCA prefetch I am just polluting the case less. I have a similar set of changes for the ixgbe transmit path that are similar to the changes I made to igb in this patch set. Those are a bit more interesting as they actually decreased the ixgbe_xmit_frame_ring function overhead for some of my tests by something like 15%. Thanks, Alex