Description: Modern processors perform poorly on sparse workloads, which are communication intensive. In this work, we aim to enhance data communication efficiency within the system while exploring the relations between parallelism, throughput, and latency to improve performance. Techniques we use include optimizing data structure layouts, saving work via algorithmic innovations, improving bandwidth utilization through software changes, and reducing synchronization overhead in multi-node systems via asynchronous active messages. We consider multiple problems, such as clique counting in graphs (computationally intensive), sparse tensor decompositions (effects of higher dimensionality), distributed breadth-first search (synchronization and communication along a network) and parallelizing finding the succinct clique tree.

Event Host: Amogh Lonkar, Ph.D. Student, Computer Engineering

Advisor: Scott Beamer

This is a hybrid event. Join us in-person or on Zoom.