BEGIN:VCALENDAR PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN VERSION:1.0 BEGIN:VEVENT DTSTART:20151117T213000Z DTEND:20151117T220000Z LOCATION:19AB DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: Graph analytics are arguably one of the most demanding workloads for=0Ahigh-performance interconnection networks. The fine-grained,=0Ahigh-rate communication patterns expose the limits of the protocol=0Astacks. The load and communication imbalance generate hard-to-predict=0Ahot-spots, and require computational steering due to unpredictable=0Adata distributions. In this paper we present a lightweight=0Acommunication library, implemented on the metal of BlueGene/Q and =0APower775/PERCS that we have used to support large-scale graph=0Aalgorithms up to 96K processing nodes and 6 million threads. With this=0Alibrary we have explored several techniques, including computation in=0Athe network for special collective communication patterns such as=0Aparallel prefix and wavefront algorithms. The experimental results=0Ashow significant performance improvements when compared to equally=0Aoptimized MPI implementations. SUMMARY:Exploring Network Optimizations for Large-Scale Graph Analytics PRIORITY:3 END:VEVENT END:VCALENDAR