[postgis-users] VMWare and PostGIS

[Martin Spott]

Paul Ramsey wrote:

> I have heard of, but never seen, VMWare hosts that attach directly to SAN
> storage over iSCSI. So the host is virtual, but the storage is at full SAN
> speed. A "best of both worlds" situation, but with the added cost of SAN
> infrastructure.

I know this sort of setups as part of my paid work.  According to my
practical experience, at the end of the day there's not too much
difference between handling the iSCSI on the host or in the guest.

Handling iSCSI in the guest provides the benefit of circumventing the
virtualization storage layer (and allows to specifically tune the
initiator for the particular workload).  Anyhow, the bonus is not very
big, because the virtualization guest HBA drivers are pretty efficient
these days.  On the other hand, an iSCSI hardware HBA in the host could
save a few CPU cycles.

What's going to kill your DB performance in both cases is iSCSI in
general.  Even though iSCSI might provide a lot of bandwidth for large
sequential I/O with selected hardware, the performance for small as
well as random I/O is quite bad, simply because it adds a huge amount
of latency.  By running a couple of benchmarks on bare-iron VMware I
was able to prove that access to *local* disks, even through storage
virtualization, outperformed any of the iSCSI setups by at least four ! 
times the numbers of reads/writes as well as bare bandwidth.

Thus, the iSCSI penalty stems from the fact that every I/O block gains
latency by passing two full TCP/IP stacks (plus multiple block device
abstraction layers) and it doesn't matter much wether it's the IP stack
on the host or the one in the guest.

Cheers,
	Martin.
-- 
 Unix _IS_ user friendly - it's just selective about who its friends are !
--------------------------------------------------------------------------