From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751700Ab3KRSFG (ORCPT ); Mon, 18 Nov 2013 13:05:06 -0500 Received: from userp1040.oracle.com ([156.151.31.81]:47168 "EHLO userp1040.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751418Ab3KRSFC (ORCPT ); Mon, 18 Nov 2013 13:05:02 -0500 Message-ID: <528A56A7.3020301@oracle.com> Date: Mon, 18 Nov 2013 11:04:23 -0700 From: Khalid Aziz Organization: Oracle Corp User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:24.0) Gecko/20100101 Thunderbird/24.1.0 MIME-Version: 1.0 To: Andrea Arcangeli , Andrew Morton CC: linux-mm@kvack.org, linux-kernel@vger.kernel.org, Pravin Shelar , Greg Kroah-Hartman , Ben Hutchings , Christoph Lameter , Johannes Weiner , Mel Gorman , Rik van Riel , Andi Kleen , Minchan Kim , Linus Torvalds Subject: Re: [PATCH 0/3] mm: hugetlbfs: fix hugetlbfs optimization v2 References: <1384537668-10283-1-git-send-email-aarcange@redhat.com> In-Reply-To: <1384537668-10283-1-git-send-email-aarcange@redhat.com> Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit X-Source-IP: acsinet21.oracle.com [141.146.126.237] Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 11/15/2013 10:47 AM, Andrea Arcangeli wrote: > Hi, > > 1/3 is a bugfix so it should be applied more urgently. 1/3 is not as > fast as the current upstream code in the hugetlbfs + directio extreme > 8GB/sec benchmark (but 3/3 should fill the gap later). The code is > identical to the one I posted in v1 just rebased on upstream and was > developed in collaboration with Khalid who already tested it. > > 2/3 and 3/3 had very little testing yet, and they're incremental > optimization. 2/3 is minor and most certainly worth applying later. > > 3/3 instead complicates things a bit and adds more branches to the THP > fast paths, so it should only be applied if the benchmarks of > hugetlbfs + directio show that it is very worthwhile (that has not > been verified yet). If it's not worthwhile 3/3 should be dropped (and > the gap should be filled in some other way if the gap is not caused by > the _mapcount mangling as I guessed). Ideally this should bring even > more performance than current upstream code, as current upstream code > still increased the _mapcount in gup_fast by mistake, while this > eliminates the locked op on the tail page cacheline in gup_fast too > (which is required for correctness too). Hi Andrea, I ran directio benchmark and here are the performance numbers (MBytes/sec): Block size 3.12 3.12+patch 1 3.12+patch 1,2,3 ---------- ---- ------------ ---------------- 1M 8467 8114 7648 64K 4049 4043 4175 Performance numbers with 64K reads look good but there is further deterioration with 1M reads. -- Khalid