From mboxrd@z Thu Jan 1 00:00:00 1970 From: Rick Jones Subject: Re: large degradation in ip netns add/exec performance in 3.13? Date: Tue, 11 Feb 2014 14:26:46 -0800 Message-ID: <52FAA3A6.9050608@hp.com> References: <52F1461D.8050703@hp.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit To: netdev@vger.kernel.org Return-path: Received: from g4t0016.houston.hp.com ([15.201.24.19]:14462 "EHLO g4t0016.houston.hp.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751379AbaBKW0q (ORCPT ); Tue, 11 Feb 2014 17:26:46 -0500 Received: from g9t2301.houston.hp.com (g9t2301.houston.hp.com [16.216.185.78]) by g4t0016.houston.hp.com (Postfix) with ESMTP id 85CAB14002 for ; Tue, 11 Feb 2014 22:26:46 +0000 (UTC) Received: from [16.103.148.51] (tardy.usa.hp.com [16.103.148.51]) by g9t2301.houston.hp.com (Postfix) with ESMTP id 6096B5C for ; Tue, 11 Feb 2014 22:26:46 +0000 (UTC) In-Reply-To: <52F1461D.8050703@hp.com> Sender: netdev-owner@vger.kernel.org List-ID: On 02/04/2014 11:57 AM, Rick Jones wrote: > Hi - > > I have a dinky little script which creates what I've been calling "fake > routers." It is far from a complete fake router, but it shows what > appears to be a very large degradation in performance in 3.13 compared > to 3.12.9 which itself is slow compared to a 3.5.0-44 kernel canonical > kernel with some upstream commits included: > > > Start/End Average Rate of Creation per Second > "Router" Count 3.5.0-44+ 3.12.9 3.13.0 > ------------------------------------------------------ > 0 to 250 7.58 5.56 2.55 > 250 to 500 7.14 5.81 2.55 > 500 to 750 6.41 5.56 2.55 > 750 to 1000 6.10 4.90 2.55 > 1000 to 1250 5.68 4.39 2.50 > 1250 to 1500 5.21 4.24 2.36 > 1500 to 1750 5.00 3.85 2.23 > 1750 to 2000 4.81 3.62 2.21 > 2000 to 2250 4.55 3.47 2.21 > 2250 to 2500 4.31 3.29 2.14 > 2500 to 2750 4.03 3.09 2.05 > 2750 to 3000 3.73 3.09 2.02 > 3000 to 3250 3.62 2.81 2.02 > 3250 to 3500 3.38 2.72 1.97 > 3500 to 3750 3.21 2.55 1.92 > 3750 to 4000 3.01 2.48 1.87 I see that the 3.13.0 kernel will scale rather well - out to 16 concurrent streams creating these "fake routers" - it drops-off at 32, but this system is only 16 cores, 32 threads (two socket E5-2670) so I suppose that really shouldn't come as a surprise. I did have the system lockup once at 16 concurrent streams on the 3.13.0 kernel. Didn't happen the next two times I tried. happy benchmarking, rick jones Time,3.5.0-44+ 4streams,3.13.0 4 streams,3.13.0 8 streams,3.13.0 16 streams,3.13.0 32 streams 0,0,0,0,0,0 30,518,419,840,1357,1327 60,914,826,1501,1988,1905 90,1242,1189,1995,2465,2332 120,1532,1519,2403,2872,2680 150,1792,1811,2751,3228,2985 180,2060,2089,3076,3537,3278 210,,2347,3359,3830,3542 232,,,,4016, 240,2447,2588,3627,,3778 270,,,3867,, 276,,,,,4032 289,,,4008,, 300,2808,2999,,, 330,,,,, 360,3131,3379,,, 390,,,,, 420,3418,3747,,, 450,,,,, 468,,4004,,, 480,3690,,,, 510,,,,, 540,3945,,,, 556,4004,,,,