[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] RE: [XENOPROFILE] and "xm vcpu-pin" doesn't work as expected



Ray,

That is normal behavior. In Xen when --separate=cpu is used, samples are
separated by "virtual" cpu. That means, CPU0 samples will have all
samples for vcpu=0 for dom0 and vcpu=0 for the guest. This seems to be
more usefull to users profiling user and kernel code which may not be
aware of Xen scheduling decisions which may move vcpus around. It would
be nice to add the capability of separating samples by physical cpu in
Xen, but I think no one is working on this. Patches would be greatly
appreciated, though. :)

Regards

Renato
 

> -----Original Message-----
> From: Ray Bryant [mailto:raybry@xxxxxxxxxxxxxxxxx] 
> Sent: Friday, January 19, 2007 1:07 PM
> To: Santos, Jose Renato G
> Cc: xen-devel@xxxxxxxxxxxxxxxxxxx; Woller, Thomas; Thomas Friebel
> Subject: Re: [XENOPROFILE] and "xm vcpu-pin" doesn't work as expected
> 
> Renato,
> 
> I've been trying to profile an HVM guest and its 
> corresponding Xen usage so I used the "xm vcpu-pin" command 
> to pin the HVM guest to CPU 1 and dom0 to CPU 0 (this is a 
> single socket, dual core setup, but that probably doesn't 
> matter here).
> 
> I then used the opsetup --separate=cpu option to split out 
> the profile samples 
> by CPU, fully expecting to see no "domain1-xen" samples on 
> CPU 0.   However, 
> what I actually see is 90% (more or less) of the samples on 
> CPU 0 and the remainder on CPU 1.
> 
> This problem occurs on both AMD V and VT.
> 
> So, I guess the question is:
> 
> (1)   Is xenoprof reporting the correct CPU, or
> (2)   Does "xm vcpu-pin" really do what it claims it does?
> 
> Any ideas on this?   For the moment, I'm ignoring the issue, 
> but I would like 
> it better if I could isolate the VCPU's to a dedicated CPU in 
> order to reduce 
> measurement overhead in the guest.   (e. g. by forcing the 
> trace daemon to 
> run on CPU 0 and the measured guest to run on CPU 1.)
> 
> BTW, this is on changeset 13062.
> 
> --
> Ray Bryant
> AMD Performance Labs                   Austin, Tx
> 512-602-0038 (o)                 512-507-7807 (c)
> 
> 
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.