A hierarchical approach to reducing communication in parallel graph algorithms
KAUST Grant NumberKUS-C1-016-04
MetadataShow full item record
AbstractLarge-scale graph computing has become critical due to the ever-increasing size of data. However, distributed graph computations are limited in their scalability and performance due to the heavy communication inherent in such computations. This is exacerbated in scale-free networks, such as social and web graphs, which contain hub vertices that have large degrees and therefore send a large number of messages over the network. Furthermore, many graph algorithms and computations send the same data to each of the neighbors of a vertex. Our proposed approach recognizes this, and reduces communication performed by the algorithm without change to user-code, through a hierarchical machine model imposed upon the input graph. The hierarchical model takes advantage of locale information of the neighboring vertices to reduce communication, both in message volume and total number of bytes sent. It is also able to better exploit the machine hierarchy to further reduce the communication costs, by aggregating traffic between different levels of the machine hierarchy. Results of an implementation in the STAPL GL shows improved scalability and performance over the traditional level-synchronous approach, with 2.5 × - 8× improvement for a variety of graph algorithms at 12, 000+ cores.
CitationHarshvardhan, Amato NM, Rauchwerger L (2015) A hierarchical approach to reducing communication in parallel graph algorithms. Proceedings of the 20th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming - PPoPP 2015. Available: http://dx.doi.org/10.1145/2688500.2700994.
SponsorsWe would like to thank our anonymous reviewers. This research is supported in part by NSF awards CCF 0702765, CNS-0551685, CCF-0833199, CCF-1439145, CCF-1423111, CCF-0830753, IIS-0917266, by DOE awards DE-AC02-06CH11357, DE-NA0002376, B575363, by Samsung, IBM, Intel, and by Award KUS-C1-016-04, made by King Abdullah University of Science and Technology (KAUST). This research used resources of the National Energy
JournalProceedings of the 20th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming - PPoPP 2015