Synchronization on nulticore architectures

Abstract

The rise of multicore computing has made synchronization necessary to overcome the challenge of sharing data between multiple threads. Locks are critical synchronization primitives for maintaining data integrity and preventing race conditions in multithreaded applications. This thesis explores the lock design space. We propose a hardware lock implementation, called the lock arbiter, which reduces lock latency while minimizing hardware overheads and maintaining high levels of fairness between threads. We evaluate our mechanism against state-of-the-art software lock algorithms and find that our mechanism has comparable performance and fairness.