Blade servers and HA

[TheVMinator]

What are considerations to take into account when an HA clusters will be built with blade servers?  For example, if the chassis' networking goes out, all HA node heartbeats from that chassis go down at same time, etc.

[ats0401]

The biggest issue is making sure all five primaries are NOT on the same blade chassis.

If all five primaries fail, HA will not work and will not restart any VM's.

The first five hosts to join a cluster are automatically made primary nodes.

Check out the link below from Duncan's blog, I think it is exactly what you need:

http://www.yellow-bricks.com/2009/02/09/blades-and-ha-cluster-design/

[idle-jam]

this is also useful: http://www.yellow-bricks.com/vmware-high-availability-deepdiv/

[bilalhashmi]

Apart from everything else, I would try and limit no more than 4 hosts in a cluster per enclosure, simply because of the same reasons discussed above. If you have 4 hosts in the same enclosure and if the enclosure goes out, you will still have one primary node to manage your cluster..

I wrote this some time ago and maybe this might be helpful for you to understand some basic things about HA. Good luck! http://www.cloud-buddy.com/?p=28

Follow me @ Cloud-Buddy.com

[a_p_]

In addition to keep the 5 primaries in different enclosures (as mentioned by user ats0401) to make sure at least one of them "survives", you may also consider to setup Link State Tracking on the blade switches. Scott Lowe wrote a interesting blog some time ago about Blade Deployments (see http://blog.scottlowe.org/2007/06/22/link-state-tracking-in-blade-deployments/)

André

[ewilts]

In addition to the other posts already mentioned, don't forget to set affinity for your key apps such that they're spread out across hosts and enclosures too.  For example, say you have a 4-guest web farm.  Put 1 & 2 on enclosure 1 and 3 & 4 on enclosure 2 and also make sure that 1 & 2 aren't on the same blad and that 3 & 4 aren't on the same blade.

It can be a bit of a challenge getting the affinities correctly defined especially if your application folks don't tell you what pieces belong where and what dependencies exist.

[bulletprooffool · Accepted solution]

When using blades, I try design my environment so all hosts on a chassis can be removed without me violating HA constraints.

So typically, I'd have a 6 node ESX cluster with  n+2 avilability, and only 2 of those blades per chassis (though if I am spec'cing hardware, I very seldom (if ever) spec blades - I prefer the old rack mount approach for ESX)

Anyway, the problem with the above design is that this only works in large environments with large ESX clusters and large requirements. Most clients I have dealt with are finaince / banking, with 100+ ESX hosts - so this is not a problem, but sometimes you'll have a smaller company with lesser requirements trying to run everything from the 1 chassis . . .

[TheVMinator]

ok thanks all for the input

[unsichtbare]

Ok, so everyone says "don't put all your HA primaries in one chassis" which is awesome advice IF you have more than one chassis! My company ran for over a year(until we bought a second recently)  on one chassis with no problems. Lets face it, blades have a lot of redundancy built in:

• 3 to 6 power supplies
• Bunch of fans
• Multiple network switches
• Multiple onboard management

We improved our chances when running on a single enclosure by enabling the crossover ports on the two switches and teaming all the NICs on a single vSwitch, then isolating networks with VLANs. That way, we could loose a NIC or a whole switch and not disconnect.

[ewilts]

If you're buying HP blades now, note that there's a promotion on - buy 4 blades and get a free chassis (with a very limited SKU selection).  Our reseller neglected to tell us when we ordered our last 4 blades so we couldn't take advantage of it but we're doing it with our next blade order.  The promotion runs through the end of July in the US.  With the various discounts and rearranging of the blade memory and power module for the chassis, we get the chassis for about $6K cheaper than we'd otherwise pay.  Depending on what config you buy, it may actually be free for you.