ABSTRACT
Data collection and analysis is rapidly changing the way scientific, national security,and business communities operate. Current database systems are based on the relational data model which is effective at selecting, joining, and modifying small sets of records stored in indexed tables. Using such database to discover complex relationships is very difficult as the systems are not designed to support subgraph isomorphism, typed path traversal, and community detection. To perform such memory intensive analytics, analysts frequently export a small snapshot of their data into a multidimensional database restricting their global view and paying a steep price in operational requirements. Even so, many analytical capabilities, such as determining behavior from structure, lie out of reach due to the lack of computational power.

In response to these analytic challenges, we are developing GEMS --- Graph Engine for Multithreaded Systems. The framework has three components: 1) a SPARQL front end to transform SPARQL to data parallel C code; 2) a semantic graph engine with scalable multithreaded algorithms for query processing; and 3) a custom multithreaded runtime layer called GMT (global memory and threading library) for scalable performance on conventional cluster systems. Our objectives are twofold: 1) to scale system size as data sizes increase, and 2) to maintain query throughput as system size grows.

We are accomplishing these objectives by targeting conventional clusters with large memory nodes, developing an in-memory graph engine, managing a fine-grain multithreaded runtime layer to hide memory latencies, and aggressively aggregating memory requests to maximize system bandwidth.

In this talk, I will describe GEMS focusing on the graph engine and on the GMT runtime layer. I will present performance results for GEMS and describe how the GMT runtime system enables to scale irregular computations on commodity clusters using techniques such as latency tolerance and message aggregation.

BIOGRAPHY
As a research scientist within the Center for Adaptive Supercomputing Software, or CASS, Alessandro Morari is part of the PNNL team designing the PNNL Semantic Graph Engine Multi-threaded (GEM)S. GEMS is based on the PNNL Global Memory Multi-threaded (GMT), a runtime system he and other scientists designed to efficiently execute irregular applications on large-scale systems. In addition, Alessandro Morari has contributed to several conference papers and journal articles. He has served as a program committee member for international workshops and as a reviewer for several international conferences. Prior to joining PNNL, Morari was a resident student at Barcelona Supercomputing Center in Spain and a research intern at the IBM T.J. Watson Research Center in US.

Conference Information