Bilkent University
Department of Computer Engineering


Performance Portability in Combinatorial Algorithms


Dr. Mehmet Deveci
Sandia National Laboratories

There have been drastic changes in the design of the supercomputer architectures in the recent years. Modern architectures are taking various paths with multi-, many-core processors or NVIDIAs Graphic Processing Units. It has become important to design algorithms that can perform well on various platforms. Following these lines, this talk focuses on the problem of enabling performance portability in combinatorial algorithms with sparse matrix sparse matrix multiplication (SPGEMM) kernel as a case study.

SPGEMM is approached from perspectives of algorithm design and implementation, its practical usage, and a theoretical model for memory accesses. First, a hierarchical, and memory-efficient SPGEMM algorithm is proposed. Second, thread-scalable data structures that enable a portable SPGEMM method are investigated. Third, a theoretical hypergraph model is developed to study the memory accesses. I will present that the SPGEMM method achieves performance portability with the proposed guidelines on massively threaded architectures. I will also demonstrate an important aspect of SPGEMM s usage in practice with structural reuse. Finally, I will show that the hypergraph model can effectively guide us in understanding performance behavior of SPGEMM variants.

Bio: Mehmet Deveci received a Ph.D. degree in Computer Science & Engineering from the Ohio State University and a B.S. degree in Computer Engineering from Middle East Technical University. He is currently a Postdoctoral Research Scientist at Sandia National Laboratories (SNL) Albuquerque, New Mexico. Since joining SNL in 2015, he has made significant contributions to various exascale scientific computing applications as a Zoltan team member, and as KokkosKernels technical lead. His professional expertise and background range from parallel combinatorial algorithms, high performance computing, scientific computing, and bioinformatics.


DATE: 13 March 2017, Monday @ 13:40