I have recently received the wonderful news that I have been granted the SIAM Travel Award for participating at the SIAM Conference on Parallel Processing for Scientific Computing (SIAM PP’16) in Paris, and presenting my work on parallel sparse Tucker tensor decompositions. In addition, there will be numerous presentations involving parallelization issues in various problem domains, and I will have the chance to attend some of these interesting talks. My talk will specifically be at the Salle des Theses on April 14, Thursday at 2:40PM.
In the talk, I will present the parallelization of the Higher Order Orthogonal Iteration (HOOI) algorithm which involves tensor-times-matrix multiply and truncated SVD steps. Specifically, I will explain how to perform the expensive tensor-times-matrix multiply (TTM) operations in a communication avoiding manner where we increase the total amount work in order to asymptotically reduce the communication requirements. Using this technique, our code graciously scales up to 4096 cores on an IBM BlueGene/Q machine using real-world tensors having over 100M nonzeros, which enables us to perform Tucker analysis on such big tensors in the order of seconds.