PhD Proposal: Advancing the State of Auto-Tuning with Programming Languages

Talk
Ray Chen
Time: 
06.28.2019 14:00 to 16:00
Location: 

IRB 5137

The scale of next-generation leadership-class computing systems presents a tremendous challenge to HPC application developers. To fully utilize such systems, an application must distribute data to thousands of nodes and coordinate execution among millions of heterogeneous computational cores. Ensuring correct execution on such a large scale currently represents a significant investment of development time. Requiring that execution must also run efficiently increases this burden beyond what is economically feasible.Auto-tuning has long been identified as the solution to this problem. Yet, the forefront of auto-tuning research has focused on isolated libraries and computational kernels. Optimization at the system-level is unlikely to be achieved through individual greedy actors. I argue that programming languages can be the central coordinating figure to ease the adoption of, and facilitate the coordination between, multiple tuners. This proposal outlines work in three separate research areas that would enable auto-tuning to grow and scale beyond what is possible with existing tuning systems.Examining Committee:

Chair: Dr. Jeffrey Hollingsworth Dept rep: Dr. Michael Hicks Members: Dr. Alan Sussman