Speaker:

Cynthia Phillips

Sandia National Laboratories, USA

Cynthia Phillips is a senior scientist in the Computing Research Center at Sandia National Laboratories. She has historically worked in combinatorial optimization, computer algorithm design and analysis, and parallel computation with elements of operations research. Her work has spanned theory, general solver development, and applications. Specific topic/applications change regularly, frequently involving multidisciplinary teams. Recently she has worked in some "big data" areas, including data structures and algorithms for streaming data analysis, complex (social) network analysis, secure multi-party computation, cooperative computing among autonomous data centers, and co-design of algorithms and architectures for extreme-scale future computers. She is also currently leading a new effort at Sandia on mixed-integer PDE-constrained optimization, which brings together massively parallel discrete optimization and massively parallel numerical optimization for PDE-constrained design decisions. She received a B.A. in applied mathematics from Harvard and a Ph.D. in computer science from Massachusetts Institute of Technology. She was named a fellow of the Society for Industrial and Applied Mathematics (SIAM) in 2016, and an Association for Computing Machinery Distinguished Scientist in 2015.

Current supercomputers have complex nodes such as multicores or GPUs. These nodes have multiple levels of memory which vary in size, latency, and bandwidth. We will discuss results and partial results from a project at Sandia National Laboratories exploring the best way to use multi-level memory for scientific-computing and general-computing primitives. We explore the value of explicit management of extra memory levels vs the system using that memory as a cache. We consider theoretical models and results relevant to cache-management decisions. We give algorithms and experimental results for problems such as sparse matrix-matrix multiplication (and related graph triangle counting), sorting, and label propagation. We consider how to lay out problems on nodes of a supercomputer to maximize "science per unit time."