Speculative prefetching of optional locks in distributed systems

We present a family of methods for speeding up distributed locks by exploiting the uneven distribution of both temporal and spatial locality of access behaviour of many applications. In the worst case, some of our methods will not produce higher network latencies than equivalent conventional distributed locking methods. In best case, the total number of messages can be constantly bounded, approximating the impression that no network latencies exist at all. Measurements and simulations based on variants of TPC database benchmarks show that hit rates enabled by speculative prefetching of optional locks appear to be similar to hit rates of conventional data caches, typically in the range from 90% to 99%. Thus overall speedup factors of 10 or more for the average latencies of distributed locks are possible. Compared to purely temporal prefetching, adding exploitation of spatial locality may significantly improve performance, typically by factors of 2 or more. We discuss some implications for the construction of distributed systems. For the class of programs well-suited for distributed systems, network latencies will nearly vanish, blurring performance differences between local and distributed systems. For program classes exposing poor locality, there is likely no help independent from distributed computing paradigms. We explain how the communication paradigm can be efficiently implemented on top of distributed shared memory (DSM) using region locks. Thus we believe that the DSM paradigm will become more attractive than explicit communication (e.g. RPC, CORBA) for the construction of distributed applications.