I had sent the following email to Kier and Ian earlier today. I think this issue is important enough to have a wider discussion on this topic. K. Y>>> On Fri, Jun 16, 2006 at 3:23 PM, in message<4492CD25.E57C.0030.0@novell.com>, Ky Srinivasan wrote:> Kier, > > In trying to narrow down the cause for domu''s timeout problems, itlooks> like we have a more serious problem in the xen scheduler. If you haveactive> load on dom0, all other domains get severely starved. A simple way tosee> this is to have a process in an infinite loop on dom0 (for bestresults have> as many instances of this process running as the number of physicalCPUs on> the box). With this load on dom0, all other domains will get littleor no cpu> time. > > Looking at the sedf code, it looks like the default values do give alot of> preference to dom0. What is the implication of setting domu''sscheduling> parameters comparable to what dom0 gets. When I set domu''s period andslice> values to what dom0 has, domu is getting a reasonable share of theCPU.> > Regards, > > K. Y >_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Actually I have noticed that Dom0 gets a default of 20 15 0 0 0 while domu''s gets 100 0 0 1 0. Also even if dom0 is lightly loaded it looks like it does not relinquish the cpu Try giving a fair share of the cpu to dom0 and U by setting the weights equal. That works - Padma -----Original Message----- From: xen-devel-bounces@lists.xensource.com [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Ky Srinivasan Sent: Friday, June 16, 2006 5:43 PM To: xen-devel@lists.xensource.com Subject: [Xen-devel] Fwd: Xen scheduler I had sent the following email to Kier and Ian earlier today. I think this issue is important enough to have a wider discussion on this topic. K. Y>>> On Fri, Jun 16, 2006 at 3:23 PM, in message<4492CD25.E57C.0030.0@novell.com>, Ky Srinivasan wrote:> Kier, > > In trying to narrow down the cause for domu''s timeout problems, itlooks> like we have a more serious problem in the xen scheduler. If you haveactive> load on dom0, all other domains get severely starved. A simple way tosee> this is to have a process in an infinite loop on dom0 (for bestresults have> as many instances of this process running as the number of physicalCPUs on> the box). With this load on dom0, all other domains will get littleor no cpu> time. > > Looking at the sedf code, it looks like the default values do give alot of> preference to dom0. What is the implication of setting domu''sscheduling> parameters comparable to what dom0 gets. When I set domu''s period andslice> values to what dom0 has, domu is getting a reasonable share of theCPU.> > Regards, > > K. Y >_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
I just applied the following patch to xen-unstable to make SEDF a bit saner. With this I can''t get it to hog the CPU 100% like I could before. -- Keir On 17 Jun 2006, at 01:43, Ky Srinivasan wrote:> I had sent the following email to Kier and Ian earlier today. I think > this issue is important enough to have a wider discussion on this > topic. > K. Y > >>>> On Fri, Jun 16, 2006 at 3:23 PM, in message > <4492CD25.E57C.0030.0@novell.com>, > Ky Srinivasan wrote: >> Kier, >> >> In trying to narrow down the cause for domu''s timeout problems, it > looks >> like we have a more serious problem in the xen scheduler. If you have > active >> load on dom0, all other domains get severely starved. A simple way to > see >> this is to have a process in an infinite loop on dom0 (for best > results have >> as many instances of this process running as the number of physical > CPUs on >> the box). With this load on dom0, all other domains will get little > or no cpu >> time. >> >> Looking at the sedf code, it looks like the default values do give a > lot of >> preference to dom0. What is the implication of setting domu''s > scheduling >> parameters comparable to what dom0 gets. When I set domu''s period and > slice >> values to what dom0 has, domu is getting a reasonable share of the > CPU. >> >> Regards, >> >> K. Y >> > > > _______________________________________________ > Xen-devel mailing list > Xen-devel@lists.xensource.com > http://lists.xensource.com/xen-devel_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Kier, We are currently testing with your patch. Your patch gives dom0 greater than 75% of the CPU. This is fine if dom0 is allowed to run only I/O intensive workloads. Is this an implicit assumption here. Regards, K. Y>>> On Sat, Jun 17, 2006 at 6:42 AM, in message<70c2d02deb3e75e333d404e3224ceb2a@cl.cam.ac.uk>, Keir Fraser <Keir.Fraser@cl.cam.ac.uk> wrote:> I just applied the following patch to xen- unstable to make SEDF abit> saner. With this I can''t get it to hog the CPU 100% like I could > before. > > -- Keir_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
On 17 Jun 2006, at 19:36, Ky Srinivasan wrote:> We are currently testing with your patch. Your patch gives dom0 greater > than 75% of the CPU. This is fine if dom0 is allowed to run only I/O > intensive workloads. Is this an implicit assumption here.Our view has always been that dom0 is primarily a service domain. If that isn''t the case then you may want to reduce its guarantee (easily done) but bear in mind that this can have a bad effect on domU I/O performance. Hopefully my patch gives a balance that strongly favours dom0 but not so much that other domains get starved for significant periods. -- Keir _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel