From mboxrd@z Thu Jan 1 00:00:00 1970 From: Gleb Natapov Subject: Re: Improving scheduler for KVM Date: Fri, 1 Nov 2013 13:03:52 +0200 Message-ID: <20131101110351.GF20205@redhat.com> References: <20131029075143.GU15657@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: kvm@vger.kernel.org To: R <19890121wr@gmail.com> Return-path: Received: from mx1.redhat.com ([209.132.183.28]:16795 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751408Ab3KALDy (ORCPT ); Fri, 1 Nov 2013 07:03:54 -0400 Content-Disposition: inline In-Reply-To: Sender: kvm-owner@vger.kernel.org List-ID: On Fri, Nov 01, 2013 at 02:51:28PM +0800, R wrote: > Hi, > > Another quick question. > > Is there any benchmark that is used to measure the performance of a VM > scheduler? > I am not sure what do you mean by "performance of a VM scheduler". Since you are the one who wants to improve scheduler I would have assumed that you know what aspects of it you are going to improve and know how to show current inefficiency. > > 2013/10/30 R <19890121wr@gmail.com>: > > Hi > > Not only lock waiter preemption, but the scheduler may affect VMs' IO > > throughput. > > Experiments must be taken to figure out the reason. > > > > I am trying to use preempt_notifier and vmexit handler to build a more > > efficient scheduler. > > Something like priority boosting. > > > > 2013/10/29 Gleb Natapov : > >> On Tue, Oct 29, 2013 at 10:20:37AM +0800, R wrote: > >>> Hi, everyone > >>> > >>> I am a graduate student. And now I have some spare time. > >>> I notice that KVM uses kernel scheduler to schedule VCPUs. > >>> But there exists many problem beyond the capability of current > >>> scheduler. (e.g. Lock Waiter Preemption problem) > >>> > >>> And I don't want to reinvent the wheel. So I want to implement a > >>> module which can be used by the scheduler to schedule VCPUs more > >>> efficient. > >>> > >>> Is there any documentation about any problem that I should pay attention to? > >>> Any comment is welcome. > >>> > >> If you are thinking about gang scheduler it was done before, but it will > >> never be accepted upstream. And IIRC pvticketlock result was close if > >> not better than gang scheduling. If you are thinking about something > >> else then implement it in Linux scheduler directly. Linux scheduler is > >> not pluggable, so you cannot change it from a module. > >> > >> -- > >> Gleb. > > > > > > > > -- > > Thanks > > Rui Wu > > > > -- > Thanks > Rui Wu -- Gleb.