Topology-Aware Asynchronous methods and the Sparse Matrix-Vector Multiply

Speaker: Amanda Bienz (UIUC)
Date: Thursday, 3 December 2015, 15:30-17:30
Session: Numerics I
Talk type: Short talk (15 min)

Abstract: The sparse matrix-vector multiply (SpMV) is a key component of many linear solvers. In distributed computing, large communication costs associated with each SpMV can yield unscalable methods. The cost of messages varies, dependent on both the size and distance traveled along the network. Topology-aware methods, which calculate the relative position of nodes in the network, can be used to target the costly messages. Furthermore, a portion of the costly communication can be attributed to idle time, during which processes are waiting for all communication to complete. This idle time can be greatly reduced through the use of asynchronous SpMVs, allowing processors to multiply a portion of the matrix as soon as communication completes.