This page contains slides for Rich Vuduc's double-header talks at Illinois in February 2020.
Your public critiques (and praise) are welcome: @hpcgarage
- Talk 0: “Is deep learning wasting our time, energy, and power?” by Rich Vuduc. Wednesday, February 26, in the CS Dept. [PDF slides (~ 29 MiB)]
Tweet-ish abstract: Applications of deep learning are good, and they utilize today's machines well. But are they truly “energy-efficient?” What about other workloads? And what might machines tuned for such workloads look like, if we had fine-grained control of the distribution of resources that affect performance, like power and die area?
- Talk 1: “A communication-avoiding sparse direct solver for linear systems on CPU+GPU platforms.” (Friday, February 28, in the CSL Student Conference): [PDF slides (~ 25 MiB)] Work led by Piyush Sao (ORNL), jointly with Xiaoye (Sherry) S. Li (LBNL), and presented by Rich Vuduc (Georgia Tech).
Here are some links to some of the key papers referenced in these talks.
- S. Karamati, J. Young, R. Vuduc. “An energy-efficient single-source shortest path algorithm.” In IPDPS'18. (to appear; preprint available upon request)
- P. Sao, O. Green, C. Jain, R. Vuduc. “A self-correcting connected components algorithm.” In FTXS'16.
- J. Choi, X. Liu, M. Dukhan, R. Vuduc. “Algorithmic time, energy, and power on candidate HPC building blocks.” In IPDPS'14.
- K. Czechowski et al. “Improving the energy efficiency of big cores. In ISCA'14.
- P. Sao, R. Vuduc. “Self-stabilizing iterative solvers.” In ScalA'13.
- J. Choi, D. Bedard, R. Fowler, R. Vuduc. “A roofline model of energy.” In IPDPS'13.
- Piyush Sao, Ramki Kannan, Prasun Gera, Richard Vuduc. “A supernodal all-pairs shortest path algorithm.” In PPoPP'20. doi:10.1145/3332466.3374533.
- Piyush Sao, Xiaoye S. Li, Richard Vuduc. “A communication-avoiding 3D algorithm for sparse LU factorization on heterogeneous systems.” JPDC, 2019, doi:10.1016/j.jpdc.2019.03.004.
- “A communication-avoiding 3D sparse triangular solver.” Piyush Sao, Ramki Kannan, Xiaoye S. Li, Richard Vuduc. In Proc. ICS'19.
- “SuperLU_DIST: A scalable distributed-memory sparse direct solver for unsymmetric linear systems.” Xiaoye S. Li and James W. Demmel. ACM Transactions on Mathematical Software (TOMS), 29(2). doi:10.1145/779359.779361.
This work is generously supported in part by grants from the US Department of Energy, DARPA SDH, and the National Science Foundation (NSF Award 1710371).