My org has a number of Windows 2008 R2 Enterprise servers hosting WPT agents and testing Chrome/Safari/FF/IE. The servers are all dual CPU Intel L5320 quad-cores running at 1.86GHz, with 8GB RAM.
These servers are going to host VM’s, and I’m curious if anyone has VM’d in a similar situation or has recommendations about how much RAM to provide each VM. Win2K8 Ent can host up to 4 VM’s without having to purchase a new license.
Thanks in advance for any feedback
What OS are you going to use for the agents?
I tend to use VirtualBox (and have client using ESXi) but I normally run a mixture of XP and Win 7 and give each VM 1.5 - 2GB
It’s worth comparing performance across the agents once their running to check they’re consistent.
The agents are currently Win2K8 R2 Ent.
Our org has direct VM support for Win Vista, XP, 7 & 8, but I’ll keep Virtualbox in mind.
Thanks for the feedback, and specifics re: VM RAM size.
I usually allocate 1 core and 1GB of RAM to the VM’s and they work quite well. I have 10 VM’s running on a single E3-1230 (quad core, hyperthreading) with 16GB of Ram and the overall system utilization (viewed from the hypervisor) when running tests in all of the VM’s hovers around 60-80%.
I usually use VMWare but Microsoft’s Hypervisor scales well too. 4VM’s would scream on that hardware. You’re probably going to be memory limited and could stack even deeper if you wanted.
One caveat is that GPU acceleration doesn’t work in VM’s (Dulles has that issue now). I’m experimenting with adding Quadro cards that do support in-VM GPU acceleration and adding support for testing hardware-accelerated browser rendering. Not a big deal, just something to be aware of as all of the browsers move towards using GPU’s in their rendering pipeline.
Thanks for the detailed feedback, Patrick. It turns out that, with our current hardware, we’re disk i/o limited (single IDE non-RAID configuration). We’re working on new hardware, and will likely use the current stack to host 2 VM’s each.
FWIW, SSD’s work REALLY well for hosting VM’s. The insane IOPS and random performance are a perfect fit and remove disk I/O entirely as a bottleneck.