Scalable reader-writer locks for parallel systems

Abstract

Current algorithms for reader-writer synchronization do not scale for readers: readers cannot acquire locks in parallel. The authors describe two new algorithms that allow parallelism among readers during lock acquisition; this is achieved by distributing the lock state among different processors, and by trading reader throughput for writer throughput. Their experiments show that when reads are a large percentage of lock requests, the throughput of each of their algorithms scales significantly better than current algorithms.<>