Date(s) - 28 Oct 2016
1:10 PM - 2:00 PM
3043 ECpE Building Addition
Speaker: Penporn Koanantakool, Electrical Engineering Research for the College of Engineering at University of California – Berkeley
Title: Communication Avoidance for Algorithms with Sparse All-to-all Interactions
Abstract: In parallel computing environments from multicore systems to cloud computers and supercomputers, data movement is the dominant cost in both running time and energy use. This is true for memory systems and interprocessor networks with hardware trends suggesting the gap between computing and data movement will continue to grow. Minimizing communication is therefore necessary in devising scalable parallel algorithms. In this talk, I will give the intuition behind communication-avoiding, originally designed for linear algebra, and show that the idea of replicating data to reduce communication costs are applicable to other problems that involve all-to-all interactions between elements of arrays. I will show how we applied these ideas to important scientific kernels such as N-body and matrix multiplication, and how we handled sparsity. Our algorithms are provably optimal in communication and scalable to tens of thousands of processors, exhibiting orders of magnitude speedup over more commonly used algorithms. I also will describe how these algorithms can be used in machine learning applications, for example, sparse inverse covariance matrix estimation for graphical model estimation. Finally, I will discuss some of the open problems in this area and implications for future architectures and applications.
Bio: Penporn Koanantakool is a Ph.D. candidate in computer science at University of California, Berkeley, advised by Professor Katherine Yelick. She holds a B.Eng. in computer engineering from Kasetsart University in Thailand where she was granted an academic excellence medal from His Royal Highness Crown Prince Maha Vajiralongkorn. She received a Fulbright scholarship for her graduate study in the United States. Her research focuses on optimizing parallel scientific applications in domains such as remote sensing, molecular dynamics, linear algebra, and machine learning.