Re: [Xen-users] SMP enabled Dom0 or not?

[Mark Williamson <mark.williamson@xxxxxxxxxxxx>]

> >     I wonder if the performance of a Xen machine can be increased by
> >     disabling SMP in the Linux kernel by default, basically having one
> >     of the 8 processors tied to dom0.

Devoting a (logical) processor to dom0 can improve IO performance for guests, 
it's true.  Note that even just dedicating a hyperthread (if you have them) 
can improve things.

> >     In my scenario I use NFS or iSCSI as file backend. Looking at NFS
> >     there will be a lot of tapdrives, while in the iSCSI scenario there
> >     is fewer overhead in userspace processes.
> >
> >     Could anyone give me a hint on the performance increase or decrease
> >     using SMP vs Uniprocessor?

FYI, XenLinux will automatically optimise itself for UP or SMP operation 
without you having to recompile.  The spinlock operations are patched out if 
the kernel is booted in UP, or patched in for SMP.  Whatsmore, I think this 
is even done at runtime, so a kernel can SMP-ify or de-SMP-ify itself on the 
fly (!).  I think this might have gone into mainline Linux a while back, 
actually.

I wouldn't be surprised if it's actually not possible to run a pure UP dom0 on 
an SMP system but I don't know for sure.

> > In the original "Xen and the Art of Virtualization" they actually
> > disabled SMP and had better IO performance. I don't know if this is
> > still true.
> >
> > Take a look at:
> > http://research.microsoft.com/~tharris/papers/2003-sosp.pdf
> > www.clarkson.edu/class/cs644/xen/files/repeatedxen-usenix04.pdf
> > <http://www.clarkson.edu/class/cs644/xen/files/repeatedxen-usenix04.pdf>
>
> Now I guess in 2003 there was no concept like tapdisk yet. I'll see if I
> can get a clean benchmark. Of 32 VMs doing the same task, SMP vs non-SMP.

Back then dom0 didn't even handle IO for the domains, it was all done in 
Xen ;-)  Things have moved on quite a long way since then!

Worth noting that if the processes in dom0 are just doing IO then they'll be 
blocked most of the time, so the performance may depend less on the number of 
CPUs available to dom0 and more on the regularity of scheduling (i.e. 
deploying a dom0 with dedicated PCPUs is probably the ultimate here).

Cheers,
Mark


-- 
Push Me Pull You - Distributed SCM tool (http://www.cl.cam.ac.uk/~maw48/pmpu/)

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