HA and DRS with mixed VMotion compatibility

[ChrisYearsley]

What are the implications of creating a single HA and DRS cluster with a mix of host types, say two groups of servers that are only VMotion compatible within each group? Of course DRS could find itself unable to balance load if all of one 'flavour' of server is overloaded; what happens in that case - anything nasty?

Perhaps the real question is whether there are advantages to multiple clusters grouped into VMotion 'islands'? There are certainly plenty of disadvantages to this. It would be more elegant to throw a mixture of servers together in a "one site, one cluster" approach, but how well does VMware cope?

I'm assuming that everything can be on the same shared storage, of course. Hmm... that's another issue!

I'm fortunate enough to have a farm of identical servers at the moment, but upgrades will soon make this no longer the case.

Chris

[letoatrads]

This is one of those ' it depends' answers. By default - VC is pretty strict about VMotion and DRS and when 2 systems are in a single large DRS/HA cluster that has hosts that won't VMotion it simply will fail the check and not use DRS.

You certainly could end up with an unbalanced load, say more on an Intel set of machines vs and AMD set...but that is manageable as long as you have enough resources of both types around. I think the advantage of having one large cluster vs several smaller is the HA. If HA comes in to play, your VM doesn't care where it comes from ( AMD or Intel) as it is essentially a cold migration.

Another option, is using CPU masking you COULD relax VMotion constraints to allow some that normally don't fly to work. However, this is a YMMV, and definitely needs to be tested in your environment before doing anything along these lines.

[ChrisYearsley]

Does it decline to use DRS at all, even between compatible hosts? Or is it more intelligent, and automates where it can, and generates recommendations where manual cold migrations are needed?

[letoatrads]

Nope, its smart enough to use DRS but only between the hosts that are compatible with it ( VMotion capable)

However I don't believe it will generate recommendations on where a cold migration of a server would be effective ( I could be wrong, I just don't recall ever seeing one of my mixed clusters make a recommendation that went Intel to AMD or vise versa.

You can always set DRS/HA affinity rules inside a cluster as well...

[peetz · Accepted solution]

Hi,

we are running a mixed cluster (4x Intel and 5x AMD) for quite a while now, and have not had any real issues.

DRS will not try (and fail with) VMotion from an Intel to AMD host although both types are in the same cluster, but it will automatically and successfully use VMotion inside an island (i.e. among compatbile hosts) to balance the load there.

Of course there is a risk that both islands become unbalanced among each other, but is is lowered by the fact that the automatic placement feature is working flawlessly with mixed hosts. So, if you power on a new VM or an existing VM that was powered off previously it will be automatically placed on the least loaded island.

However, DRS will never suggest cold migrations (when you run it in suggestion mode), it is focussed on VMotion only.

HA operations are not affected at all (i.e. it runs as well as with homogeneous clusters), because it does not use VMotion, but powers on VMs that have been powered off after a host failure, and then uses the above mentioned automatic placement feature (if DRS is also enabled).

There are no issues with shared storage.

Regarding CPU masking for forcing VMotion compatibility, this will work only among CPUs of the same vendor. There is no mask available that will allow VMotion from an Intel to an AMD processor (unfortunately).

In summary - to answer your "real question": No, I do not see any advantages to multiple clusters grouped into VMotion 'islands'.

Regards

Andreas

Message was edited by:

peetz

to also answer "the real question".