On Feb 8, 2008, at 5:31 PM, Howard Hinnant wrote:
On Feb 8, 2008, at 5:22 PM, Peter Dimov wrote:
Howard Hinnant:
Here are results (dating around 3 years ago):
In order: 63 seconds. try/back off/without yield: 47 seconds. try/back off/with yield: 34 seconds.
With Alexander Terekhov's "swap-based mutex", I'm not seeing a significant difference between the versions with and without yield. The penalty likely comes from spin-waits in lock() (which is why they're evil).
There's no significant difference with critical sections, too. Maybe I'm doing something wrong. :-)
Or this could be a manifestation of the fact that I never really had "hands on" tests with Windows. I have observed that if the contention isn't high, then the algorithm doesn't matter. Does your test have the threads sleeping or otherwise doing independent work? In mine the threads did nothing but fight over the locks.
Oh, also note that my single-core results were only 5% apart (comparing lock in order to try/backoff/yield). Are you running on single core? -Howard