Name: Sherry Li
Pronouns:
Biography:
Sherry Li is a Senior Scientist and group lead of the Scalable Solvers Group at Lawrence Berkeley National Laboratory. She has worked on diverse problems in high performance scientific computations, including parallel computing, sparse matrix computations, high precision arithmetic, and combinatorial scientific computing. She is the lead developer of SuperLU, a widely-used sparse direct solver, and has contributed to the development of several other mathematical libraries, including ARPREC, LAPACK, PDSLin, STRUMPACK, and XBLAS. She has collaborated with many domain scientists to deploy the advanced mathematical software in their application codes. She earned Ph.D. in Computer Science from UC Berkeley. She has served on the editorial boards of the SIAM J. Scientific Comput. and ACM Trans. Math. Software, as well as many program committees of the scientific conferences. She is a SIAM Fellow.
Institution/Lab: Lawrence Berkeley National Laboratory
Website: https://crd.lbl.gov/xiaoye-li
SRP Collaboration Topic/Title: Development and deployment of autotuning tools using statistical and machine learning methods
Field or research area: ML for HPC
Please select all the topical areas that apply to your project:
Data Science (i.e., data analytics, data management & storage systems, visualization); High-Performance Computing; Machine Learning and AI
Brief Abstract:
The project is to develop an autotuning software framework via statistical and machine learning techniques, such as multitask and transfer learning using Gaussian process. The goal of this work is to help the HPC codes to choose the near-optimal parameters setting on a large-scale parallel machine, which take into account the characteristics of the input problems. The typical minimization metrics are runtime and memory usage. Since each execution (“function evaluation”) of the HPC code is expensive and takes a lot of resources, it is not feasible to use a brute-force approach (e.g.,grid-search) to search for optimal parameters. Therefore, it is critical to “learn†some knowledge from the limited number of executions with certain input instances and build a prediction model for the unseen tasks. The research will be conducted in the context of GPTune (https://gptune.lbl.gov/).
Desired relevant skills, background, or interests:
Programming language: Python / Matlab / C++; Linear algebra; Probability and statistics; HPC
Other comments:
Do any special requirements apply? other
Other, specify: none required
Keywords:
Peformance autotuning; Bayesian optimization; Gaussian Processes; HPC; Parallel sparse matrix algorithms
Lightning Talk Title: Development and deployment of autotuning tools using statistical and machine learning methods