From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S933194AbdKGIYv (ORCPT ); Tue, 7 Nov 2017 03:24:51 -0500 Received: from mx2.suse.de ([195.135.220.15]:43392 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S933049AbdKGIYu (ORCPT ); Tue, 7 Nov 2017 03:24:50 -0500 Subject: Re: [PATCH] locking/pvqspinlock: Hybrid PV queued/unfair locks To: Waiman Long , Peter Zijlstra , Ingo Molnar Cc: linux-kernel@vger.kernel.org, Paolo Bonzini , =?UTF-8?B?UmFkaW0gS3LEjW3DocWZ?= , Boris Ostrovsky References: <1509723331-10469-1-git-send-email-longman@redhat.com> From: Juergen Gross Message-ID: <22362e36-e69c-6bb9-9679-997eb615fe56@suse.com> Date: Tue, 7 Nov 2017 09:24:48 +0100 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.4.0 MIME-Version: 1.0 In-Reply-To: <1509723331-10469-1-git-send-email-longman@redhat.com> Content-Type: text/plain; charset=utf-8 Content-Language: de-DE Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 03/11/17 16:35, Waiman Long wrote: > Currently, all the lock waiters entering the slowpath will do one > lock stealing attempt to acquire the lock. That helps performance, > especially in VMs with over-committed vCPUs. However, the current > pvqspinlocks still don't perform as good as unfair locks in many cases. > On the other hands, unfair locks do have the problem of lock starvation > that pvqspinlocks don't have. > > This patch combines the best attributes of an unfair lock and a > pvqspinlock into a hybrid lock with 2 modes - queued mode & unfair > mode. A lock waiter goes into the unfair mode when there are waiters > in the wait queue but the pending bit isn't set. Otherwise, it will > go into the queued mode waiting in the queue for its turn. > > On a 2-socket 36-core E5-2699 v3 system (HT off), a kernel build > (make -j) was done in a VM with unpinned vCPUs 3 times with the > best time selected and is the number of vCPUs available. The build > times of the original pvqspinlock, hybrid pvqspinlock and unfair lock > with various number of vCPUs are as follows: > > vCPUs pvqlock hybrid pvqlock unfair lock > ----- ------- -------------- ----------- > 30 342.1s 329.1s 329.1s > 36 314.1s 305.3s 307.3s > 45 345.0s 302.1s 306.6s > 54 365.4s 308.6s 307.8s > 72 358.9s 293.6s 303.9s > 108 343.0s 285.9s 304.2s > > The hybrid pvqspinlock performs better or comparable to the unfair > lock. > > By turning on QUEUED_LOCK_STAT, the table below showed the number > of lock acquisitions in unfair mode and queue mode after a kernel > build with various number of vCPUs. > > vCPUs queued mode unfair mode > ----- ----------- ----------- > 30 9,130,518 294,954 > 36 10,856,614 386,809 > 45 8,467,264 11,475,373 > 54 6,409,987 19,670,855 > 72 4,782,063 25,712,180 > > It can be seen that as the VM became more and more over-committed, > the ratio of locks acquired in unfair mode increases. This is all > done automatically to get the best overall performance as possible. > > The table below shows the kernel build times on a smaller 2-socket > 16-core 32-thread E5-2620 v4 system. > > vCPUs pvqlock hybrid pvqlock unfair lock > ----- ------- -------------- ----------- > 16 436.8s 433.4s 435.6s > 36 366.2s 364.8s 364.5s > 48 423.6s 376.3s 370.2s > 64 433.1s 376.6s 376.8s > > Again, the performance of the hybrid pvqspinlock was comparable to > that of the unfair lock. > > Signed-off-by: Waiman Long Reviewed-by: Juergen Gross Juergen