Howard Hinnant wrote:
On Feb 8, 2008, at 2:42 PM, Anthony Williams wrote:
Howard Hinnant <hinnant <at> twcny.rr.com> writes:
3. The yields are absolutely necessary to get the high performance mentioned in the previous note.
I find this comment intriguing. Is this always the case, or does it depend on the specifics of how a mutex is implemented on a given platform?
That is a question I've been asking and myself for a long time, unfortunately without a satisfactory answer. Here's my experience, and a few guesses...
Hi Howard, If you can share the code, I'll see how it performs on my Linux systems. I bet it depends greatly on the workload. I have been considering writing an instrumented mutex that records how often it blocks. In my real code, the answer seems to be almost never. Peter's comment about evil spin-waits in lock() is interesting as I was considering implementing something like that. I wonder about this instead: void mutex::lock() { if (atomic-attempt-to-lock) { // uncontended return; }; // first wait, try yielding since it's fast: sched_yield(); if (atomic-attempt-to-lock) { // got lock on 2nd attempt return; } // We failed again. Maybe there are lots of threads waiting; in that // case it will be more efficient to futex_wait() since the kernel will // then only wake us when our lock has been woken. futex_wait(....); } I'll try to find time to experiment with this soon. Phil.