From mboxrd@z Thu Jan 1 00:00:00 1970 From: Avi Kivity Subject: Re: KVM with hugepages generate huge load with two guests Date: Sun, 21 Nov 2010 13:28:12 +0200 Message-ID: <4CE9024C.3060709@redhat.com> References: <20101001223048.GA31596@amt.cnet> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: kvm@vger.kernel.org, Marcelo Tosatti To: Dmitry Golubev Return-path: Received: from mx1.redhat.com ([209.132.183.28]:51855 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751435Ab0KUL2R (ORCPT ); Sun, 21 Nov 2010 06:28:17 -0500 In-Reply-To: Sender: kvm-owner@vger.kernel.org List-ID: On 11/21/2010 02:24 AM, Dmitry Golubev wrote: > Hi, > > Seems that nobody is interested in this bug :( > It's because the information is somewhat confused. There's a way to prepare bug reports that gets developers competing to see who solves it first. > Anyway I wanted to add a bit more to this investigation. > > Once I put "nohz=off highres=off clocksource=acpi_pm" in guest kernel > options, the guests started to behave better - they do not stay in the > slow state, but rather get there for some seconds (usually up to > minute, but sometimes 2-3 minutes) and then get out of it (this cycle > repeats once in a while - every approx 3-6 minutes). Once the > situation became stable, so that I am able to leave the guests without > very much worries, I also noticed that sometimes the predicted > swapping occurs, although rarely (I waited about half an hour to catch > the first swapping on the host). Here is a fragment of vmstat. Note > that when the first column shows 8-9 - the slowness and huge load > happens. You can also see how is appears and disappears (with nohz and > kvm-clock it did not go out of slowness period, but with tsc clock the > probability of getting out is significantly lower): > Are you sure it is hugepages related? Can you post kvm_stat output while slowness is happening? 'perf top' on the host? and on the guest? -- error compiling committee.c: too many arguments to function