Re: [Xen-users] Xen VMs and Unixbench: single vs multiple cpu behaviour

[Dario Faggioli <dario.faggioli@xxxxxxxxxx>]

On Sun, 2015-11-15 at 08:03 +0100, Marko ÄukiÄ wrote:
> Hello!
> I am doing a comparison of different hypervisors using the Unixbench
> benchmark software.Â
> The testing is done on a machine with an Intel Core i5 cpu (4 cores).
> 
> I am running two series of tests:Â
> - test A: 10 repetitions of the unixbench test in a VM with 1 cpu
> - test B: 10 repetitions of the same test in a VM with 4 cpus
> 
> Test B has two sets of results because in a multicpu machine
> Unixbench runs a
> Âsingle test first (let's call it B-1) and then runs 4 tests in
> parallel - one for each cpu (test B-2).
> 
> Compared to other hypervisors, in Xen VMs the results of tests A and
> B-2 are very good.
> Results of B-1, however, are much worse at tests with system calls.
> For example:
> 
> Xen VM:
> test A create process test average score: 1871,7
> test B-1 create process test average score: 286,7 (!)
> test B-2 create process test average score:3259,2
>
So, when a single instance of UnixBench is run inside a VM with 4
vCPUs, on an host with 4 pCPUs, performance drops, is this description
accurate?

If yes, it is a bit weird, as I'd expect results from A and B-1 to be
similar.

The 4 host pCPUs are coming from 2 physical cores, each one with
hyperthreading, I'm guessing?

> I noticed the same behaviour using Xen 4.4.2 and 4.5.1 (both
> installed with default
> options in ubuntu server 14.04.2 and 15.10 respectively).
> 
I'm familiar with UnixBench, so I'll try to reproduce. I don't have any
similar hardware, so I'd have to somehow "make it up", by tweaking the
configuration of bigger boxes, which might not be that accurate...

Anyway, can we in the meantime see the following:
Â- the output of `xl vcpu-list`, issued while the benchmark isÂ
 Ârunning, both for A and B-1 cases
Â- the log produced by
 Â`xl debug-key r ; xl dmesg > xl-debug-keyr.log`, issued while theÂ
 Âbenchmark is running, for both A and B-1 cases
Â- the benchmark result (no need to run it 10 times, in this case) inÂ
 Âcase you pin the VM's vCPUs like this (issue the command after theÂ
 ÂVM has started, but before running the benchmark inside it):
  xl vcpu-pin 0 all 0-3 ; xl vcpu-pin <vmid> all 4
Â- the same, but with pinning configured like this:
  for i in `seq 0 4`;do xl vcpu-pin <vmid> $i $i; done

Finally, UnixBench's score comes from the combination of the results of
various benchmarks. Can you post the full results (i.e., the results of
all the benchmarks UnixBench performs) of (just one run of) case A and
case B-1?

Regards,
Dario

-- 
<<This happens because I choose it to happen!>> (Raistlin Majere)
-----------------------------------------------------------------
Dario Faggioli, Ph.D, http://about.me/dario.faggioli
Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxx
http://lists.xen.org/xen-users