Adhitha Dias

The Best Paper at this year’s International Conference on Supercomputing (ICS) was awarded to a team from Purdue University’s Elmore Family School of Electrical and Computer Engineering. Graduate student Adhitha Dias was first author on the paper, titled “SparseLNR: accelerating sparse tensor computations using loop nest restructuring.” Co-authors include Purdue ECE graduate students Kirshanthan Sundararajah and Charitha Saumya, and Milind Kulkarni, professor of electrical and computer engineering. The paper was published in the Proceedings of the 36th ACM International Conference on Supercomputing (ICS).

The paper explores a means of executing computations over sparse arrays more quickly using less memory. A graph is a mathematical approach to discussing relationships between objects. Graphs are an incredible way to model real-world phenomena such as social networks, protein-protein interaction, etc. For instance, in a social network graph, nodes are people and edges are friendships. A caveat of these graphs is that they are sparse — a person could be friends with 100 people but not 1 million. If we use an array to represent friendships, only 100 of them would be non-zero and all others would be zero, and this kind of array is known as a sparse array. These sparse arrays are stored in compressed formats to save space. Hence, computations performed over these arrays are irregular in nature. A swath of modern applications, such as recommendation engines, web-page ranking, etc., rely on repeatedly operating over sparse arrays, and the performance of these applications are critical. Hence, it is important to automate code generation to perform these computations.

Naively produced code for such applications takes a long time to execute when multiple computations are performed together. If these computations are separated, then it results in multiple smaller computations which require a lot of memory usage, and in return take a longer time to execute. In this paper, the team from Purdue develop a novel optimization to combine these computations such that it executes faster using less memory. Since these applications are ubiquitous, they integrated our optimization into a compiler that produces code for sparse array computations.

The ACM International Conference on Supercomputing (ICS) is the premier international forum for the presentation of research results in high-performance computing systems. The 36th conference (ICS-2022) was held as a virtual global event in June 2022. Papers were solicited on all aspects of the architecture, software, and applications of high-performance computing systems of all scales (from chips to supercomputing systems).