From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jaspal Subject: Using taskset , powertop and perf with qemu-kvm Date: Wed, 10 Oct 2012 13:22:22 +0530 Message-ID: <50752936.7070703@gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit To: kvm@vger.kernel.org Return-path: Received: from mail-pb0-f46.google.com ([209.85.160.46]:37181 "EHLO mail-pb0-f46.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753458Ab2JJHwZ (ORCPT ); Wed, 10 Oct 2012 03:52:25 -0400 Received: by mail-pb0-f46.google.com with SMTP id rr4so446016pbb.19 for ; Wed, 10 Oct 2012 00:52:25 -0700 (PDT) Sender: kvm-owner@vger.kernel.org List-ID: Hello , I wish to run some cpu intensive tasks on a vm , use only one core specifically #0 of the 6 available ( Xeon 3670 ) and all the while record basic perf events like cpu-cycles , instructions etc. I am using taskset to define the cpu-affinity of the qemu process : taskset -p 0x1 `pgrep qemu` (assuming only 1 vm is running ) I can verify that the vm is now running only on core 0 using top. top also shows that the qemu process is using 99% of it's share of CPU time ( %CPU column in top ) upon running the tasks in vm. However what I fail to see is the increase in time spent of the core 0 in the C0 ( active ) state via powertop. It barely manages to cross 12%. If I run the same tasks ( FFT benchmark ) on the host on a different core ( say #1 ) and monitor it using powertop again , I can see that C0 % for core 1 is above 100% ( 104 ~ 107 ) So why does not the output of top ( qemu is using 99% of it's cpu time , which is spent on core 0 ) co-relate with powertop ( which says core 0 is idle for around 88% of time ) ? Also , powertop reports that core 2 is being used heavily ( 104 % ) and I suspect this is one which is executing the vm operations. But why ? There are no other cpu/memory intensive tasks running on the host. Next , I am using the perf kvm record function to save the stats. If I use all the 7 core counters available along-with -a option , perf generates ~2MB/second . Is there a way to save just the raw count of the events ? Also , if instead of -a , if I give -C 0 , perf just saves around 0.016MB/s as opposed to 1.894MB/s in case of -a option. If the vm is indeed using core 0 , then perf should have recorded the events. On a last note , does perf kvm automatically selects the vm to profile ( in case multiple vms are running ) on the basis of the guest.kallsyms and guest.modules files passed to it ? Thanks , Jaspal