On 30.10.2011 12:30, Sad Clouds wrote:
Hi, I was expecting for a contended lock, spin lock to be at
least as efficient as adaptive mutex. However a simple test program
seems to indicate quite a big difference:
atom$ gcc -O2 lock_test.c -DMUTEX_LOCK -lpthread
atom$ time ./a.out
4.54 real 16.76 user 0.00 sys
atom$ gcc -O2 lock_test.c -DSPIN_LOCK -lpthread
atom$ time ./a.out
14.04 real 51.83 user 0.00 sys
Any idea why pthreads spin lock is 3 times slower compared to pthreads
mutex?
At first glance I'd say that it largely depends on the micro-benchmark you are using, especially /how/ you are hammering on the contended lock.
Sometimes it may be easier for a program to make progress with mutexes because sleeping threads can let other threads run uncontended for a long time; if a thread is performing pthread_mutex_lock/unlock in a tight loop, I bet the loop will be faster as the lock/unlock will take the fast path most of the time.
Whereas spin locks could generate live locking situations where progress happens, albeit slower. Like a 4x100 metres relay but with a lot more changeovers.
-- Jean-Yves Migeon jeanyves.migeon%free.fr@localhost