Analytics and Machine Learning Systems on Graphs

Abstract: Graph is an important type of big data that naturally capture the relationships between entities in real-world applications. Systems that perform analytics and machine learning on graphs constitute an important kind of domain-specific systems. Irregular memory access and inter-node communication are two key problems for graph systems. In this talk, I will present two recent systems for graph pattern mining (GPM) and graph neural network (GNN) training. The first system, Khuzdul, is a distributed GPM execution engine. It can be connected to an existing single-machine GPM system with a well-defined interface, enabling efficient communication and data reuse leveraging GPM algorithm properties. The second system, GNNPipe, is a distributed deep graph neural network (GNN) training system. It can drastically reduce communication by pipelined model parallelism and ensure resource utilization with hybrid parallelism. I will also briefly discuss our recent endeavor of the cost-efficient CPU-based GNN training system.

Bio: Xuehai Qian is an associate professor of the Department of Computer Science at Purdue University. His research group mainly focuses on developing efficient domain-specific systems and architectures for graph and machine learning. In the past, he developed scalable and efficient coherence protocols and architectural techniques to support the debugging and enhance programmability of shared-memory parallel programs on multicores. He is the recipient of W.J Poppelbaum Memorial Award at UIUC, NSF CRII and CAREER Award, and the inaugural ACSIC (American Chinese Scholar In Computing) Rising Star Award. His research contributions have been published in various top architecture and system conferences such as ASPLOS, ISCA, MICRO, HPCA, PLDI, and OSDI, etc. He has been inducted into the "Hall of Fames" of all of the top-4 architecture conferences.

https://mit.zoom.us/j/96748144797