NCTS(South)/ NCKU Math Colloquium | |
| DATE | 2013-01-03¡@16:10-17:00 |
| PLACE | R204, 2F, NCTS, NCKU |
| SPEAKER | ²Àsª@¥ý¥Í¡]Software engineer in CUDA Libraries & Algorithms Group at Nvidia¡^ |
| TITLE | Load Balance in Sparse Linear Algebra |
| ABSTRACT |
As Nvidia keeps its tile on Top500 list, developers, researchers and scientists are willing to do GPU programming in elegant way, not just rely on compiler.The most important part in computational mathematics is sparse linear solver, mainly used in computational fluid dynamics (CFD), structural mechanics, circuit simulation and reservoir simulation, ...etc. However memory-bound nature of sparse linear algebra makes it difficult to perform well, not to mention sparsity pattern may kill parallelism. In this talk, we will review some recent works in sparse linear solver, from either public domain or libraries at Nvidia. Starting from matrix-vector multiplication, through sparse matrix-matrix multiplication, triangular solver, we will cover an issue of load imbalance, which is a hot topic in parallel computing world. We will discuss how to address and mitigate load imbalance, especially recent stable tridiagonal solver published at supercomputing conference 2012.. The goal of this talk is to discuss some possibilities on direct sparse solver on GPU. Throughout the talk, you will learn what is load imbalance and how it affects performance on a many-core processor. |