Benchmarks on new HyperThreading logical core avoidance in beta

Started by Jeremy Collake, December 31, 2010, 04:39:28 PM

Previous topic - Next topic

Jeremy Collake

I have not done these yet, but want to (did you like that tease? lol). They are going to show some interesting phenomenon. In *some* cases applications will run faster with an appropriately tweaked CPU affinity. You see, if the thread is moved from one CPU to another in its time slice, it may lose code/data in the CPU cache. This is known as core/CPU thrashing. Allowing it to never hit the fake logical cores may make a big difference. Then again, maybe the type of application that it really matters on is too rare to care about. Or maybe the difference is too negligible. Only benchmarks will give us some insight. From a theoretical stand-point, I believe that, even with XP+'s awareness of HyperThreading, there will be a performance gain in applications that are primarily single or dual threaded and highly CPU bound. Now, this isn't the typical app -- most are I/O bound. So, ... We'll see ;)
Software Engineer. Bitsum LLC.