Scale Up vs Scale Out, hardware sizing

[colin270]

Look for a little advise if I can please.....

I'd really appreciate your comments on scale up vs scale out, when planning VI3 clusters for best performance.

If for example I want to virtualise 40 machines, The customer has had Capacity Planner and PowerRecon assessments done by a third party.

They have run the following two consolidation scenarios and Capacity Planner is happy to place all 40 VM's on either hardware scenario.

Scenario 1 - Two 4 way boxes running six core processors (total 48 cores), with 60GB RAM each (total 120GB), each host running 20 guest VM's.

Scenario 2 - Six 2 way boxes running quad core processors (total 48 cores), with 20GB RAM each (total 120GB), each host running 6 to 7 guest VM's.

All servers would be connected to a 4GBps Fibre Channel SAN and all VMFS volumes would be stored here, the SAN can't serve I/O faster than 4GBps, regardless of how many servers it is serving, the bottleneck is in the fibre fabric, so I/O shouldn't be a deciding factor for the server hardware itself.

So the processor cores and memory are equal, but there must be other factors affecting performance, cache, bus speed, etc. Can you really expect the same guest O/S peformance from a big box running 20 VM's, compared with a smaller box running 6-7 VM's. Which scenario would yield better performance and why?

I have the raw data from the PowerRecon assessment, but as it was carried out by a third party, I don't have access to run new consolidation scenarios. What performance counters do I need to be looking at to help me make this decision?

Please assume cost is no object. I've been through all the normal stuff about HA node failure, CPU per rack space, but the customer is still asking the specific purely technical question of "Which will give me better performance?"

Thanks v much.

[ChrisDearden]

From a Raw performance point of view I dont think there would be much in it ? With 6 cores on a die is the cache shared ? if so then Plan B would seem to give you more cache/core ?

However with 6 Hosts , there is a slightly more Service console overhead form your total pool of memory / CPU , although that would be minimal.

[mreferre]

Colin,

this is a very old paper I wrote 4 years ago on the subject. The technology has changed but many (not all though) of the the thoughts still apply.

The scale-up Vs scale-out decision point is not a performance discussion. It's more like a religious topic to me.

Massimo.

[TomHowarth]

There is a lot more to performance than Just CPU cycles, for example your 6 dual quads cores will have many more paths to the shared storage than your 2 quad/Hex way boxes, yes you will probabaly get better raw cpu performance on the hex core and a better ability to run smp guests due to there being a greate number of cores in the host. but you may run into IO contention as you are running more guests accross less path.

I have not mentioned Hardware failure, 5 machines will be much more able to take up the load of a single host failure that the remaining single quad/hex box in a pair of hosts.

If you found this or any other answer useful please consider the use of the Helpful or correct buttons to award points

Tom Howarth

VMware Communities User Moderator

[depping]

Tom is correct. Go for the 6 hosts. This will give you more flexibility and less contention. I hardly ever go for Scale Up, only VDI might be a good use case for scale up...

Duncan

Blogging: http://www.yellow-bricks.com

If you find this information useful, please award points for "correct" or "helpful".