Evaluating and designing software mutual exclusion algorithms on shared-memory multiprocessors

Spring 1996 hared-memory multiprocessor systems must provide facilities called critical sections for programs to share physical and logical resources. Only one processor at a time can process a CS. This requires a method to ensure mutually exclusive access to the logically atomic operations of a shared CS. Two approaches to mutual exclusion are hardware algorithms, which use primitives, and software algorithms, which require only software read and write instructions (see the “Hardware primitives versus software algorithms” sidebar). We previously examined architecture and system effects on spin-locks, a type of hardware a1gorithm.l Extensive experiments on two different shared-memory multiprocessor systems-Bolt, Beranek, and Newman’s TC2OOCl and Kendall Square Research’s KSR-l-showed that the execution Elehavior of these algorithms, and therefore their performance, differs significantly. These machines use different types of interconnection networks and cache/memory systems. This construction produces different types of nonuniform memory access (NUMA) execution patterns for the mutual exclusion algorithms. We’ve concluded that architecture and system effects should be seriously considered when developing and implementing mutual exclusion algorithms. Our experience has motivated us to investigate comprehensive performance effects on portable software mutual exclusion protocols. However, standard analyses of mutual exclusion algorithms do not fully account for these effects. For example, these analyses express the memory-access complexity of an algorithm as a function of IV, the number of