Measuring TLB Misses

[charlie88]

Hi all,

Is there any way to measure TLB misses to determine whether a VM should use large pages / hwMMU? Specifically I'm looking to measure this for a W2K8 64-bit VM at present.

I can't find any tool that provides this info - I've seen references to "use performance counters" but can't find any counter which appears to give this info. It seems to be a very "low level" counter so not sure whether it would actually be visible from a VM itself or whether I'd need to use a counter from ESX.

Thanks,

Charlie

[msatya]

hi please find details -web.eecs.utk.edu/~shirley/papers/iccs04.pdf

-------------------------
please award points if you see fit.Thanks

[rickardnobel]

charlie88 wrote:

Is there any way to measure TLB misses to determine whether a VM should use large pages / hwMMU?

I think that the TLB handling is done deep in the the internals of the CPU which is invisible the the operating system, in this case the Vmkernel. It would be very interesting to some logging about this, but since this handling is such low level where the system is trying to optimize RAM access there should be little room for logging up to the operating system regarding this.

Do you have the possibility to actually try setting your VM (or test version) in software and hardware MMU and do some performance benchmarking?

[charlie88]

That was my suspicion but couldn't find anything that definitively said so.

Yes for the VM in question I made the change to swMMU and reran the job. Made no improvement so it's presumably not got heavy TLB misses.

I was looking for a tool so we could proactively look for any VM that might be a candidate for this.

Thanks.