Virtualized Server - best practice?

[PaulDavidson1]

From a number of threads and talks, it seems that Esri has recommended that splitting servers is more efficient than using a larger server.

For example (2) 2 core, 4 GB servers typically are more efficient than (1) 4 core 8GB server.

e.g.: https://community.esri.com/message/523098?et=watches.email.thread#523098

However, just today I was reading the exact opposite in the new 10.3.1 documentation:

Virtualization and ArcGIS Server—Installation Guides (10.3 and 10.3.1) | ArcGIS for Server

CPU recommendations

It's recommended that you use a few large virtual machines as opposed to many small virtual machines. For example, a single virtual machine with eight CPUs will perform better than four virtual machines each with two CPUs.   If you're configuring failover, high availability, or separate ArcGIS Server sites for separate purposes, follow the recommendation above to determine the size of machines in each site.

Anyone willing to comment on this?

[JacobBoyle]

Thanks for posting, I hadn't caught the change either.  I've been running four 2 core, 16 gb machines for 2 years now without issues, but i'll look at performance since updating to 10.3.1.

[PaulDavidson1]

Looking forward to your results Jacob.

Might be worth running System Test against your setup, especially if you had similar results from the prior version.

https://www.arcgis.com/home/item.html?id=e8bac3559fd64352b799b6adf5721d81

I was getting hung up a bit in a system design as to whether to separate or combine components.

Obviously, it depends a lot on load and resource availability.

I think I'm going with your method of more VMs with fewer cores than one monster machine.

I've tried to arrange the design so that taking out components may not necessarily bring down the entire system.

And generate fail overs (even though we have full DR setup.  I sure do like working in the Virtual world!)

[JacobBoyle]

I have professional services coming in at the end of September, I've been running the most current 1.x version for 6 months now.  It'll be of interest to see what the results will be when we migrate to the new architecture.

The one thing I would consider is using multiple machines per cluster in the site, this will give you the failover you're looking for.

[PaulDavidson1]

Thanks Jacob, look forward to your results.  And yes, I think you're correct on the clusters.

We have a Health Check being done in September so I hope to have an initial production setup up and running for a look by Professional Services.