I/O Sizing relocated VM swap files

[mrstorey]

Hi,

We are looking to use the new VI3.5 feature of relocating the VM swap files to an alternate datastore.

We beleive this will be a significant cost saving because we will not only free up space on our datastores, but we don't have to replicate this data to our DR site, saving on more storage and replication bandwidth.

I wondered if any of you have done the same, and wondered if you had a sizing strategy for the number of swaps you would allow on a single LUN / vmfs datastore?

For example, we have well over 200 VM's split across around 20 LUNs in our primary HA and DRS cluster, but I'm reluctant to place 200 VM swap files on a single LUN for fear of degraded VM performance due to disk I/O and scsi locking on the 'swap file' vmfs datastore.

However, isn't it true that VM swap files only get written to when memory on the ESX host is sparse? ...and obviously when the swap is created when powering on the VM! So maybe disk I/O and scsi locking won't be an issue in an environment with comfortable amounts of free physical RAM?

Does anyone have any thoughts on this? I was thinking maybe storing no more than 50-100 VM swap files on a single vmfs datastore....does this sound ok to you?

Alex

[weinstein5 · Accepted solution]

You are correct - the per vm vmkernel swap files are memory of last resort - so if you have your environment sized approriately they should be used rarely - I agree with your estimate 50-100 per datastore -

[mrstorey]

Cool - thanks for your advice. Most of the VMs have 2gb of RAM, so we're gonna create a number of 100gb VMFS datastores to store the swapfiles, thus approx providing space for up to 50 VMs each.

[weinstein5]

a reminder when you create a vmfs data store about 600 MB is held to store metadata -

[mrstorey]

Yep - no worries. It's all just to get a rough idea. I'll let you know how it goes!