Here's the guts of a response that I posted a while back to an internal mailing list re: tradeoffs of runing your managed code as 64-bit vs 32-bit. YMMV, and I'll remind you that every perf question has a thousand answers depending on the situation.


>>>>>>> snip >>>>>>>>>>>>>>>>>


Here's my own personal list of the big pluses and minuses of moving to 64-bit code...



- more memory (+++++)

- better 64-bit math (+++)

- X64 OS kernel takes advantage of more memory to do good things for a lot of stuff (+++)


- things need more memory (pointers are bigger, and especially in managed code references are everything and are everywhere) (--)

- the processor's cache is effectively smaller (when comparing against the same machine in 32-bit vs 64-bit mode) because of the prior point (----)

- code also tends to be bigger because of extra prefix bytes and instructions that carry around 8-byte immediate values instead of 4 byte immediate values


What this tends to mean is that code that runs extremely well on 32-bit, doesn't have any 64-bit math (or otherwise take advantage of improvements in the 64-bit processor) and runs well in < 2GB of memory without having to bother hitting the disk for anything will likely continue to run on 64-bit with somewhat MORE memory usage and a little bit slower because the processors cache is effectively smaller when compared to the bloated size of the things that need to be in it.

In the scenario described above you get the minuses of the platform without taking advantage of the pluses.

If however you have an application or set of applications that can take advantage of the pluses to offset the minuses they can come out in the black (sometimes _very_ much so). We have seen a number of large applications which used to be memory starved on 32-bit and had some type of home-grown paging able to throw that more or less out the window and see their performance go up by 2, 3 or even 4X. PaintDotNet (which is a pretty cool photo editing application, Rick Brewster's blog: rewrote a bunch of their filters to take advantage of 64-bit math and saw speed boosts moving to x64 of 3X+ for those filters. I just saw a presentation the other day where was saying that they have seen both significant reliability boosts and throughput increases moving to 64-bit (however they were running 12 app pools on a box and were definitely running into the memory limits of the 32-bit system).