ArcSDE and VM Ware

[runegullstrom]

Hi all,

I'm having a physical server which is twice as fast as a virtualized server (VM Ware).
I'm measuring it with sdeanalyze from the SE_TOOLKIT. The servers have different specs but I've set all the oracle init params such that they are identical. Could this be due to VM Ware? I've read
somewhere that there are known issues with VM Ware as compared to Hyper V.

specs are:
VM Machine
Intel Xeon CPU @ 2.27 GHz (3 processors)
10 GB RAM
Windows Server 2008 R2 Standard
ARCSDE 10.0
ARCSDE 10.0 Query analysis Tool
Oracle 11G

Physical server
Intel Xeon CPU @ 3.33 GHz (8 processors)
31.9 GB RAM
Windows Server 2008 R2 Standard
ARCSDE 10.0
ARCSDE 10.0 Query analysis Tool
Oracle 11G

I've set the oracle init params and memory params such that the physical server uses the same amount of memory and cpu's. Only GHz would be different.

The difference must be the VM Ware or?

cheers

[VinceAngelo]

THREE processors on the VM? Is it 2, 4, or 8?  Is the database in the same VM too?
Was your benchmark on a single large table, or were you exercising multiple tables
simultaneously?

The *best* performance I have ever seen from a VM is a 20% performance tax.
50% is a bit steep, but not unexpected (both Oracle and ArcSDE are I/O dependent,
and while both Oracle and Esri support virtual  architectures, neither recommends
their use in production systems). 

It's not generally recommended to run ArcGIS Server on the RDBMS host.

- V

[runegullstrom]

Yes it's 3 processors, see attachment. Is that strange? I didn't set the VM up and I'm not an expert on VM at all.

The Oracle 11G/SDE10 database is on the VM. ArcGIS Server is not installed on the VM.
I loaded one big cadastral table with 3 million rows on both machines and tested it with
sdeanalyze -o square, firing random queries on the whole extent for 1 and 5 minutes.

The VM has a throughput of ~ 11.000, the physical server ~20 / 25k.

What's nagging me is that I've saved a log file a while back where the VM server reached a TPS of 20k as well. I can't reproduce it and I don't know what's changed in the mean time (I'm not the only one using this test server). That's why I've set all oracle init parameters the same - on the VM and the physical. The statistics are very consistent now, the physical server is > 2 as fast as the VM server. We're in the midst of setting up a new architecture, so I guess its clear, we should go for the physical server.

Rune

[VinceAngelo]

Three processors is exceedingly strange. I'm amazed the OS will recognize that configuration.

VMware recommends against 4-CPU VMs on a system with VMs with fewer than that because
the system must block until a larger set of CPUs is available, placing the 4-CPU host in contention
with *all* the 2-CPU hosts. I didn't see a significant performance benefit of a 4-CPU VM over a
a 2-CPU VM when simulating load for 16 simultaneous ArcGIS connections, and both were blown
out of the water by an 8-CPU physical host (900 v 800 v 72 seconds). A second 4-CPU VM did
a bit better (600 seconds), but still not twice the 2-CPU VM. Linux VMs were often slghtly faster
than Windows VMs, and not subject to the simultaneous user limitation of ~72-76 connections.
I never measured a benefit to using 64-bit ArcSDE over 32-bit on Windows hosts, though there
was a slight benefit (10-15%) on Linux.

It would be interesting to see what your performance was like with a Direct Connect client --
it's possible that the bottleneck also involves Oracle (my testing was just ArcSDE application
server on VMs, with 2 and 4 node Oracle RAC on physical hosts).

- V