[Xen-users] Poor pgbench results in RAM-centric select test

[Szymon Juraszczyk <szymon@xxxxxxxxxxxxxx>]

  Hi,

  I've recently deployed a Xen 4.1.3 server based on Debian Wheezy for
Dom0. The hardware is a dual E5-2630 Xeon machine. I deployed a PVHVM
webserver there and I am quite satisfied with it's performance. I have
two SSD drives there and powerful Adaptec controller there and the plan
was to place a DB server on another VM and use these SSD drives for that.

  I wanted to do some baseline benchmarking before the dbserver
deployment and then some strange pgbench results came out of that. To
cut a long story short, select-only benchmarks on ~300 MB databases
(fully fitting in RAM) give results even 10 times worse than expected (!?).

  I tested PVHVM domU with various 3.2 and 2.6.32 kernels, various core
counts and RAM sizes. What's wierd, I got the same poor results when I
run the benchmark on dom0 (in desperation I installed PostgreSQL there),
again with various RAM and core count settings. At most I got up to 10k
select transactions per second, while on the same hardware using the
same dom0 kernel booted up without Xen hypervisor the results were
around 100k tps (!?!).

  Please note that the database fits compeletely in RAM even on a 512 MB
RAM dom0/domU (that is one of the variants I tested). There is obviously
some bottleneck with cached I/O or maybe more generic problem with RAM
access. Hard drive access is completely irrelevant here as the database
gets cached in kernel buffers right at the start of the benchmark (which
runs for 5 minutes to get more accurate results). There was no point for
me in testing the SSD drives in action having observed this strange
behaviour.

  Has anyone observed something like that? Any idea what may be wrong?

-- 
Best regards,
Szymon Juraszczyk

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