Just a couple of screen shots illustrating what happens with I/O when 30+ Windows XP virtual desktops boot at the same time on local SSD storage.
The hardware used was an 8-core (16 thread) Nehalem 5520 IBM x3650 M2 rack server, 64GB RAM running ESXi 4.1, with two local 160GB Intel 320 MLC SSD in RAID-0 providing the VMFS datastore for the desktops
The desktops were the linked clone variety in a VMware View 4.6 floating pool, and were running XP SP3 with 1 vCPU and 1.5GB RAM. Sophos anti-virus was also running in the VM’s.
Firing up esxtop on the ESXi host and looking at the disk latency and I/O
Booting the VM’s is using the read-only replica at this point, so read IOps are hitting 8200, with 493 write IOps against the linked clones. You can see latency is under 1ms. It did peak at 3ms on occasions.
Just running two Intel SSD in RAID-0 gives anywhere between 8,000-11,000 random 4k read/write IOps with very low latency.
With normal 15,000rpm SAS or FC disks at 180 IOps, you can see you may need 45-60 spindles to get this type of performance, along with the associated cost and rack space required.
This makes local SSD the best bang for buck for VDI deployments, as it mitigates boot-storms and AV update storms as well as enhancing deployment speed.
It also provides the nice side effect of a much snappier, smoother and consistency fast desktop experience for the end user – and that’s what it’s all about.