[Fedora-xen] Benchmarking Report (Xen Virtualiztion , Para vs Fully)

[Daniel P. Berrange]

On Sat, Apr 14, 2007 at 01:03:42PM -0400, Kiranjeet Kaur Beniwal wrote:
> Hi everyone ,
> 
> We are the students of seneca college toronto . We have developed a project 
> which performs the benchmarking tests over Xen machine and compare Para and 
> fully virtualiztions . Here is our final report of results of benchmarking 
> Xen tests. We used three different tools to perform the tests and posted the
> data and graphs in our report.
> 
> The html format of our project is 
> http://tux.senecac.on.ca/~mshaver/xenbench/xenbench.xhtml

The pure-compute results look reasonable, but I don't believe those results for
Disk I/O in fully-virt for a second - Section 6, Disk I/O Unixbench is the exact
opposite of expectation - were results flipped?.  Based knowledge of the architecture
of paravirt/ fullyvirt I/O subsystems, there is no way in the world fullyvirt
will ever outperform paravirt for network or disk. In previous benchmarks I've
typically seen Paravirt have x2 the disk performance of fullyvirt, and have
x10 the network performance.

I'm not familiar with UnixBench enough to say whether it is good benchmark
to use, or whether these were just runtime anomolies. I'd recommend doing
some disk I/O tests using IOZone which is the pretty common standard for
disk benchmarking.

I'd also recommend doing some scaling tests using multiple concurrent guests,
since people rarely use virtualization for just a single guest. When adding
additional concurrently executing guests I'd expect to see the gap between
paravirt and full-virt increase even further. 

Regards,
Dan.
--
|=- Red Hat, Engineering, Emerging Technologies, Boston.  +1 978 392 2496 -=|
|=-           Perl modules: http://search.cpan.org/~danberr/              -=|
|=-               Projects: http://freshmeat.net/~danielpb/               -=|
|=-  GnuPG: 7D3B9505   F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505  -=|