Page / Swap Files and you - Part 1

Hey another "part 1" article!

Okay, so I have a bad track record ... but I'm going to keep trying!

On a couple of occasions I've been asked "what should page files be set to" (I say page file because it invariably is a Windows deployment we are discussing). My answer is "depends." There are quite a few greatly detailed articles about how page files work etc so I'm not going to go over them here. Google (or Bing (ick)) them and have fun with the feeling of your head exploding. For my part lets keep it digestible and focused on virtualization.

VM level

So this most often comes up cause someone foolishly allocated 30gb to the main partition and now you're in a storage hole. Anyone who doesn't know the "60/x/x" rule needs to be educated ... but I digress. So going back the question is how much page file is needed. In the past the rule was always 1.5 x physical memory. This rule came about for various reasons but one of the main ones was that 4gb was the nominal limit for server memory. Giving it a big page file help performance (research 'commit charge' for more on that). The other reason, though less heavily used, was the fact that when a memory dump happened (BSOD) you needed a page file big enough to take it.

In this day and age of x64 and super high limits (if you can call them limits) on memory begs the question: do we stick with the rule? Short answer: yes. Long answer: well yes but if we are talking a really big amount then maybe no. Example: a file server is usually a 2/1/4 (cpus/cores/RAM) affair, this results in a page file size of 6gb. Not so bad and within reason. Now lets look at a virtualization server. Example: a XenApp server is typically a 2/2/16 which would result in a page file (using the old rule) of 24gb. Whoa!

So on the latest x64 OSes we have a quandary! Okay, not really cause there is an easy way to figure out what is really needed. Remember the first reason we noted, well that part is not needed quite as much with so much physical memory available. So the dump is the key and here's a question: how much memory is really needed for a dump. Turns out a lot less then maybe it used to need. Try 1/3 of what you might think. So here's what I'd suggest:

Physical Mem
2 to 8 GB == 1.5 x RAM
10 to 16 GB == 0.75 x RAM
18 to ? == 0.5 x RAM

It generally works and makes sure that there is enough page file for a typical kernel dump in the larger ranges. The storage foot print is reasonable so things don't get out of hand. Of course you can still use the 1.5x rule and be just fine if you have storage to burn.

In the next part we go to the host level!

Peace.