Graph models of social information systems typically contain trillions of edges. Such big graphs cannot be processed on a single machine. The graph object must be partitioned and distributed among machines and processed in parallel on a computer cluster. Programming such systems is very challenging. In this work, we present DH-Falcon, a graph DSL (domain-specific language) which can be used to implement parallel algorithms for large-scale graphs, targeting Distributed Heterogeneous (CPU and GPU) clusters. DH-Falcon compiler is built on top of the Falcon compiler, which targets single node devices with CPU and multiple GPUs. An important facility provided by DH-Falcon is that it supports mutation of graph objects, which allows programmer to write dynamic graph algorithms. Experimental evaluation shows that DH-Falcon matches or outperforms state-of-the-art frameworks and gains a speedup of up to 13×.
展开▼
