From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from 1wt.eu (1wt.eu [62.212.114.60]) by ozlabs.org (Postfix) with ESMTP id 6A0AC2C00C7 for ; Fri, 12 Apr 2013 01:03:50 +1000 (EST) Date: Thu, 11 Apr 2013 17:03:26 +0200 From: Willy Tarreau To: Sebastian Hesselbarth Subject: Re: [PATCH] net: mv643xx_eth: Add GRO support Message-ID: <20130411150326.GA19978@1wt.eu> References: <1365684023-9967-1-git-send-email-sebastian.hesselbarth@gmail.com> <20130411131333.GD1910@1wt.eu> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: Cc: Andrew Lunn , Jason Cooper , linux-kernel@vger.kernel.org, "David S. Miller" , Soeren Moch , Paul Mackerras , linux-arm-kernel@lists.infradead.org, Dale Farnsworth , netdev@vger.kernel.org, linuxppc-dev@lists.ozlabs.org, Florian Fainelli , Lennert Buytenhek List-Id: Linux on PowerPC Developers Mail List List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Hi Sebastian, On Thu, Apr 11, 2013 at 04:47:49PM +0200, Sebastian Hesselbarth wrote: > I did some simple tests on Dove/Cubox with 'netperf -cCD' and > gso/gro/lro options on > mv643xx_eth. The tests may not be sufficient, as I am not that into > net performance testing. In fact the difference only happens when the NIC has not verified the checksum itself IIRC, which should be for non-IPv4 traffic. I agree that it's not easy to test a bridge with a cubox which has a single port :-) Maybe you'll see a difference in IPv6 traffic or with VLAN traffic, as I seem to remember this chip does not do cksum offloading on VLANs, but I could be wrong. > I tried todays net-next on top of 3.9-rc6 without any gro patch, with > the initial > patch (Soeren) and your proposed patch (Willy). The results show that > both patches > allow a significant increase in throughput compared to > netif_receive_skb (!gro, !lro) > alone. Having gro with lro disabled gives some 2% more throughput > compared to lro only. Indeed this is consistent with my memories, since Eric improved the GRO path, it became faster than LRO on this chip. Regards, Willy From mboxrd@z Thu Jan 1 00:00:00 1970 From: w@1wt.eu (Willy Tarreau) Date: Thu, 11 Apr 2013 17:03:26 +0200 Subject: [PATCH] net: mv643xx_eth: Add GRO support In-Reply-To: References: <1365684023-9967-1-git-send-email-sebastian.hesselbarth@gmail.com> <20130411131333.GD1910@1wt.eu> Message-ID: <20130411150326.GA19978@1wt.eu> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org Hi Sebastian, On Thu, Apr 11, 2013 at 04:47:49PM +0200, Sebastian Hesselbarth wrote: > I did some simple tests on Dove/Cubox with 'netperf -cCD' and > gso/gro/lro options on > mv643xx_eth. The tests may not be sufficient, as I am not that into > net performance testing. In fact the difference only happens when the NIC has not verified the checksum itself IIRC, which should be for non-IPv4 traffic. I agree that it's not easy to test a bridge with a cubox which has a single port :-) Maybe you'll see a difference in IPv6 traffic or with VLAN traffic, as I seem to remember this chip does not do cksum offloading on VLANs, but I could be wrong. > I tried todays net-next on top of 3.9-rc6 without any gro patch, with > the initial > patch (Soeren) and your proposed patch (Willy). The results show that > both patches > allow a significant increase in throughput compared to > netif_receive_skb (!gro, !lro) > alone. Having gro with lro disabled gives some 2% more throughput > compared to lro only. Indeed this is consistent with my memories, since Eric improved the GRO path, it became faster than LRO on this chip. Regards, Willy From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S936151Ab3DKPD7 (ORCPT ); Thu, 11 Apr 2013 11:03:59 -0400 Received: from 1wt.eu ([62.212.114.60]:32812 "EHLO 1wt.eu" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S933306Ab3DKPD6 (ORCPT ); Thu, 11 Apr 2013 11:03:58 -0400 Date: Thu, 11 Apr 2013 17:03:26 +0200 From: Willy Tarreau To: Sebastian Hesselbarth Cc: Andrew Lunn , Jason Cooper , Benjamin Herrenschmidt , linux-kernel@vger.kernel.org, Florian Fainelli , Soeren Moch , Paul Mackerras , Lennert Buytenhek , Dale Farnsworth , netdev@vger.kernel.org, linuxppc-dev@lists.ozlabs.org, "David S. Miller" , linux-arm-kernel@lists.infradead.org Subject: Re: [PATCH] net: mv643xx_eth: Add GRO support Message-ID: <20130411150326.GA19978@1wt.eu> References: <1365684023-9967-1-git-send-email-sebastian.hesselbarth@gmail.com> <20130411131333.GD1910@1wt.eu> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.4.2.3i Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Hi Sebastian, On Thu, Apr 11, 2013 at 04:47:49PM +0200, Sebastian Hesselbarth wrote: > I did some simple tests on Dove/Cubox with 'netperf -cCD' and > gso/gro/lro options on > mv643xx_eth. The tests may not be sufficient, as I am not that into > net performance testing. In fact the difference only happens when the NIC has not verified the checksum itself IIRC, which should be for non-IPv4 traffic. I agree that it's not easy to test a bridge with a cubox which has a single port :-) Maybe you'll see a difference in IPv6 traffic or with VLAN traffic, as I seem to remember this chip does not do cksum offloading on VLANs, but I could be wrong. > I tried todays net-next on top of 3.9-rc6 without any gro patch, with > the initial > patch (Soeren) and your proposed patch (Willy). The results show that > both patches > allow a significant increase in throughput compared to > netif_receive_skb (!gro, !lro) > alone. Having gro with lro disabled gives some 2% more throughput > compared to lro only. Indeed this is consistent with my memories, since Eric improved the GRO path, it became faster than LRO on this chip. Regards, Willy From mboxrd@z Thu Jan 1 00:00:00 1970 From: Willy Tarreau Subject: Re: [PATCH] net: mv643xx_eth: Add GRO support Date: Thu, 11 Apr 2013 17:03:26 +0200 Message-ID: <20130411150326.GA19978@1wt.eu> References: <1365684023-9967-1-git-send-email-sebastian.hesselbarth@gmail.com> <20130411131333.GD1910@1wt.eu> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Cc: Andrew Lunn , Jason Cooper , Benjamin Herrenschmidt , linux-kernel@vger.kernel.org, "David S. Miller" , Soeren Moch , Paul Mackerras , linux-arm-kernel@lists.infradead.org, Dale Farnsworth , netdev@vger.kernel.org, linuxppc-dev@lists.ozlabs.org, Florian Fainelli , Lennert Buytenhek To: Sebastian Hesselbarth Return-path: Content-Disposition: inline In-Reply-To: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: "linux-arm-kernel" Errors-To: linux-arm-kernel-bounces+linux-arm-kernel=m.gmane.org@lists.infradead.org List-Id: netdev.vger.kernel.org Hi Sebastian, On Thu, Apr 11, 2013 at 04:47:49PM +0200, Sebastian Hesselbarth wrote: > I did some simple tests on Dove/Cubox with 'netperf -cCD' and > gso/gro/lro options on > mv643xx_eth. The tests may not be sufficient, as I am not that into > net performance testing. In fact the difference only happens when the NIC has not verified the checksum itself IIRC, which should be for non-IPv4 traffic. I agree that it's not easy to test a bridge with a cubox which has a single port :-) Maybe you'll see a difference in IPv6 traffic or with VLAN traffic, as I seem to remember this chip does not do cksum offloading on VLANs, but I could be wrong. > I tried todays net-next on top of 3.9-rc6 without any gro patch, with > the initial > patch (Soeren) and your proposed patch (Willy). The results show that > both patches > allow a significant increase in throughput compared to > netif_receive_skb (!gro, !lro) > alone. Having gro with lro disabled gives some 2% more throughput > compared to lro only. Indeed this is consistent with my memories, since Eric improved the GRO path, it became faster than LRO on this chip. Regards, Willy