Alex Hopmann’s Blog

So Windows (like most OSes) maintains a number of caches where it keeps in RAM frequently accessed disk contents (as well as a few other things). Normally these are not adjustable- it just uses extra free memory for this to improve performance, and the minute you actually need the memory, it is freed up and available to a real program.

Hyper-V puts a little twist in this. When you create or start up a new VM, it needs the full memory as hard-allocated memory to that VM and apparently it doesn’t force a cache clear if there isn’t enough memory available. We are building a bunch of stuff on 48GB servers and initially thought we would easily be able to put 3x 15GB VMs on them- the remaining 3GB would be plenty for the OS that basically wasn’t doing anything. But at times we have found the Cached memory is as high as 20GB (after a bunch of big file copies, usually of VHD images), and the machine doesn’t have enough free RAM to boot the target VMs.

Any thoughts on how we can clear this? Dynamic memory coming in future releases should fix this issue, but for now it would be great if we could just run a cache clear/flush right before VM start time.

I’ve been using VMWare workstation to run some virtual machines lately. Its very helpful to be able to have a WinXP IE6 box, WinXP IE7, Vista, Linux, etc, all at once, not to mention other nice uses of VMWare.

So yesterday I’m doing some big file copies and noticing that my network transfer performance on my gigabit Ethernet is running a solid 6%. Not a blip over or under, as if something were hard-limiting it. I also notice there are these two extra network adapters that VMWare has installed which each think they are 100mbps connections.

So I try disabling them (since with a little poking around I find out they are only necessary if you are doing on-machine NAT or internal networks) and like magic my network transfers hop up to 9.5%. These get installed by default but for most people you will never need them, so get rid of them.

Now to figure out what the next bottleneck is. The machines involved have plenty of CPU, I was testing on multiple disks so disk IO should not have been limiting, and I’m pretty sure everything is hooked up with nice high quality cables on my gigabit switch.