Eliminating Cascading Stall on Hardware Transactional Memory

Multi-core processors are equipped in almost every computer systems from smartphones to high-end server machines, and shared memory programming becomes increasingly important for programmers to utilize the multi-core systems. Lock-based thread synchronization techniques have been commonly used in parallel programming on multi-core processors. However, lock can cause deadlocks and this leads to poor scalability. To make up for the shortcomings of lock, transactional memory (TM) is proposed and widely studied. On TMs, transactions are executed speculatively while any conflicts do not occur on shared variables. However, wasteful re-executions and waits can cause low concurrency and drastic performance degradation. In this paper, we propose a method for resolving Cascading Stall which is one of the main factors of low concurrency on TM. The result of the experiment shows that the method can reduce execution time 56.5% in maximum and 11.1% in average with 16 threads.