Thanks for the information. I have some questions though:
1) Does the P4 you use support VT?
2) Are the VMs you created multi-CPU running SMP OS? If not,
any chance you can repeat your experiment using SMP (by
keeping the total number of virtual CPUs the same? I.e.,
if it was 8 1-CPU VMs, then use 4 2-CPU VMs). I assume
that your POVRay benchmarks utilize pthreads?
3) Have you actually performed the bench marks on multi-core?
Intel? AMD? Specifically I am interested in the data when
running multi VMs on multi-core to see when the shared bus
on Intel will max out. My guess is that AMD should fare
better.
As para-virt vs full-virt, you might want to Google for the
following paper done by VMware guys:
"A Comparison of Software and Hardware Techniques for x86
Virtualization" by Keith Adams and Ole Agesen, ASPLOS'06
Regards,
Shing-Shong
>
> Mostly true but not completely. para-virt is faster overall but full
> virt is *supposed* to be faster for CPU processing. I put stars around
> supposed because I've never gotten full virt to work and so I haven't
> done my own testing. I've had to go on what other people say.
>
> When I did my testing of paravirtualization prior to the IU LinuxFest
> presentation I did, I found that Xen domUs where about 5% slower than
> the native CPU speed. Which honestly is pretty fast compared to other
> emulators and things like VMWare. I did this test by creating a special
> VM that all it did was render the standard benchmarking frame in POVRay
> (3d raytracer) and compare it with how long it took in a native non-Xen
> kernel. Then I'd copy this VM and run X amount of them at the same time
> to test the results with Xen. This was done on a single Pentium 4
> 2.4GHz with 2GB of RAM. Each VM only needed something like 96MB so I was
> able to run around 16 at once.
>
> Interestingly, the CPU was most efficient when I ran 8 VMs at once. If
> I ran 4, 6, 10 or 12 it was less efficient. This is in terms of work done in
> the same amount of time.
>
> I've used this 8 VM results as the policy at Suso for how many VMs we
> allow per core on our Xen service. 8 per core seems to be a good
> balance between getting good usage out of the CPU and not making
> everybodys VM seem slow. Then of course multiple cores help when there
> are deviations from the norm.
>
> I've been pretty impressed with the VMs responsiveness even under
> significant load on the host machine or other VMs. They did a good job
> with the scheduler.
>
> One thing interesting is that the consultant from RH that was at Cook
> last month, went to go work for Dreamworks Animation after he helped us
> and he said he was going to setup a renderfarm for them that utilized
> Xen. That struck me as interesting since I would think that you'd want
> all the raw processing power you could get for rendering and what is the
> use of segregating it. Oh well, I don't know anything other than that
> about their setup.
>
>
_______________________________________________
BLUG mailing list
BLUG@linuxfan.com
http://mailman.cs.indiana.edu/mailman/listinfo/blug
No comments:
Post a Comment