Host memory monitoring thresholds

[EdWilts]

I'm trying to determine at what level I should be warning and alarming on host memory utilization.  We're currently warning at 82% but I regularly see hosts in a cluster above 82% with a warning yet another host in the cluster may be sitting at 65% yet DRS isn't balancing the workload.  This leads me to think that DRS doesn't consider this to a problem, so perhaps my thresholds are too low.

From the memory management guide, "ESX maintains four host free memory states: high, soft, hard, and low, which are reflected by four thresholds: 6%, 4%, 2%, and 1% of host memory respectively."  This would suggest that we may not even want to warn until we get below the high threshold of 94% utilization but this seems awfully high to me - my personal best practice has been to keep any resource below 70% but perhaps I'm not understanding host memory management well enough yet.

At what level do other people starting warning and alarming?  How are you determining that you're "out of memory"?  Or are you not monitoring memory utilization at all and focusing on ballooning and swapping instead (which could be an indication that you're already in trouble)?

Opinions greatly appreciated (and I'm sure that there are a lot of differing opinions).

Thanks!

[mcowger]

I wouldn't alarm on usage at all.  I would alarm if any VM starts aggressively ballooning or swapping.

[EdWilts]

Using that logic, that assumes that host memory utilization must be at least 94% because it won't start ballooning before that.

Are other admins out there really running their hosts with over 90% memory utilization on a somewhat consistent basis and not concerned?

[mcowger]

Yup - I ran many many hosts like that (hundreds).

[DoDo201110141]

What do you mean by Host Memory utilisation?

In my "Virtual Machines" list, there are 3 columns: Hostmem - MB, Guest Mem - %, and Memory Size.

Are you referring to Hostmem?

(Sorry, should have done my own googling. http://communities.vmware.com/message/620581)

[Prost]

I have looked through the alarms and cant find a setting on alarming for Balloning?  Can you help?

Also do setup a Condition Length?  If so what do you set it to?  I also alert on 70% memory usage and have a Condition length of 20 min.

Thanks-

[shishir08]

I suppose DRS will kick in only when it will see high resource drain on the Host.So whenever the host will hit the hard state the DRS will come in picture and start balancing the load.

These all  Ballooning/Compression/Swapping will come in action only when we will be at hard state of the Host at 94% of the memory.