Programming Large Dynamic Data Structures on a DSM Cluster of Multicores∗

Applications in increasingly important domains such as data mining and graph analysis operate on very large, dynamically constructed graphs, i.e. they are composed of dynamically allocated objects linked together via pointers. Parallel algorithms on large graphs can greatly benefit from software Distributed Shared Memory’s (DSM) convenience of sharedmemory programming and computational scalability of multiple machines. However, prior DSM systems did not provide programming and memory abstractions suitable for working with large dynamic data structures. In this paper, we present dynamic DSM (dyDSM), an object-based DSM that targets applications which operate on large dynamic graphs. To achieve this goal, dyDSM differs from traditional DSM systems in four important ways. First, dyDSM provides programming abstractions that expose the structure of the distributed dynamic data structures to the runtime enabling the communication layer to efficiently perform object level data transfers including object prefetching. Second, the dyDSM runtime enables the management of large dynamic data structures by providing various memory allocators to support parallel initialization and storing of dynamic data structures like graphs. Third, dyDSM provides support for exploiting speculative parallelism that is abundant in applications with irregular data accesses. Fourth, dyDSM runtime actively exploits the multiple cores on modern machines to tolerate DSM access latencies for prefetching and updates. Our evaluation on a 40-core system with data mining and graph algorithms shows that dyDSM is easy to program in the very accessible C++ language and that its performance scales with data sizes and degree of parallelism.