From mboxrd@z Thu Jan 1 00:00:00 1970 From: Gordan Bobic Subject: Re: memory performance 20% degradation in DomU -- Sisu Date: Thu, 06 Mar 2014 10:31:56 +0000 Message-ID: <5ff5b79f65bd03837291020cb8173abf@mail.shatteredsilicon.net> References: "\" <20140305173319.GC9528@phenom.dumpdata.com>" " <5317973A.6060502@bobich.net> <20140305222852.GA26966@phenom.dumpdata.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii"; Format="flowed" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20140305222852.GA26966@phenom.dumpdata.com> List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xen.org Errors-To: xen-devel-bounces@lists.xen.org To: xen-devel@lists.xen.org List-Id: xen-devel@lists.xenproject.org On 2014-03-05 22:28, Konrad Rzeszutek Wilk wrote: > On Wed, Mar 05, 2014 at 09:29:30PM +0000, Gordan Bobic wrote: >> Just out of interest, have you tried the same test with HVM DomU? >> The two have different characteristics, and IIRC for some workloads >> PV can be slower than HVM. The recent PVHVM work was intended to >> result in the best aspects of both, but that is more recent than Xen >> 4.3.0. >> >> It is also interesting that your findings are approximately similar >> to mine, albeit with a very different testing methodology: >> >> http://goo.gl/lIUk4y > > Don't know if you used PV drivers (for HVM) and if you used as a > backend a > block device instead of a file. > > But it also helps in using 'fio' to test this sort of thing. I used a dedicated disk which was not altered between the tests. Otherwise I wouldn't have been able to run the same installation on bare metal and virtualized. I don't think disk I/O was particularly relevant in the test - the CPU was always the bottleneck with no iowait time. My impression was that it was the context switching that really crippled virtualized performance, especially in multi-socket or NUMA cases. C2Q I tested on can be considered a dual-socket non-NUMA system in this context since the two dies on it don't share any caches which means higher migration penalties. Throw in the extra Heisenbergism of the domU kernel not having any idea where the hypervisor might schedule the virtual CPU mapping (I didn't pin cores in the test, perhaps I should have) and it is easy to see a case where it gets quite bad when you push the system to saturation. Gordan