Research Challenges in Computing – Ganesh Gopalakrishnan

Ganesh Gopalakrishnan
Professor
https://www.cs.utah.ed/~ganesh

September 20

11:00am

3147 MEB

https://utah.zoom.us/j/99963204158?pwd=tEpfDa4bQrzT7echP32lnt3LbY8fyP.1

Meeting ID: 999 6320 4158
Passcode: 595106

ParFloat: Pushing on Performance Post-Dennard Without Abandoning Correctness

In this Post-Dennard Era, obtaining high performance in Scientific Computing and Machine Learning requires solving correctness challenges in response to each performance-enabling design decision. I will describe the following correctness questions our past research has faced as well as contributed to: (1) In message passing, how can deadlocks arise due to insufficient buffering as well as too much buffering? (2) In shared memory parallelism, what are different approaches to detect races? (3) In floating-point, how do we (a) provide tight guaranteed error bounds at scale? (b) handle exceptions that are not trapped by hardware? (c) understand the extent of non-reproducibility? (4) When the hardware is unreliable, how can we sometimes encourage programs to fail fast? My talk will cover the basics of our solutions as well as unexplored avenues. It will close with an open discussion on how one finds cool research problems (at least I can summarize how my past N students finally got through, and see if there is any lesson… or a reason for caution :-)).