CPU-intensive busy waiting (SpinLock, WaitLock)

[jorrit.schaap]

Hi,

 

running my program in valgrind (callgrind) I noticed a lot of CPU cycles are spent doing noting in spinlock and waitlock.

Looking at the code below this is no surprise, since the while loops are spinning full speed.

 

*******BEGIN ORIGINAL SOURCE****************

INLINE void SpinLock(volatile int *ptr,int old_value,int new_value) {
    #ifdef _CELLOS_LV2
        while(!AtomicCAS(ptr,old_value,new_value)) { sys_timer_usleep(1); }
    #else
        while(!AtomicCAS(ptr,old_value,new_value)) { }
    #endif
}

INLINE void WaitLock(volatile int *ptr,int value) {
    #ifdef _CELLOS_LV2
        while(!AtomicCAS(ptr,value,value)) { sys_timer_usleep(1); }
    #else
        while(!AtomicCAS(ptr,value,value)) { }
    #endif
}

*******END ORIGINAL SOURCE****************

 

 

Here is my proposal for an improved version which is also platform independent.

 

*******BEGIN IMPROVED SOURCE****************

INLINE void SpinLock(volatile int *ptr,int old_value,int new_value) {
        while(!AtomicCAS(ptr,old_value,new_value)) { Timer::usleep(1); }
 

INLINE void WaitLock(volatile int *ptr,int value) {
        while(!AtomicCAS(ptr,value,value)) { Timer::usleep(1); }
 }

*******END IMPROVED SOURCE****************

 

[ulf.schroeter]

Spinlocks are much more efficient for short-term thread blocking as they avoid context switching (which would be forced by usleep ()). Even though your CPU load is higher, overall performance with this approach is better.

 

 

http://en.m.wikipedia.org/wiki/Spinlock

[unclebob]

Thanks for explanation, Ulf!