Skip to main content

Xulong Tang

Pennsylvania State University

Thursday, February 28, 2019
11:00AM – 12:00PM – HEC 356

Abstract

The amount of resources in computing systems are scaling out over the past decade, forming various parallel computing platforms such as Graphics Processing Units (GPUs), manycore CPUs and heterogeneous datacenters. While applications with intrinsic parallelism and regular data structures (e.g., matrix multiplication) can be naturally mapped to manycore systems and get benefits from parallelization and data locality, such mapping for applications with irregular algorithms and irregular input data sets (e.g., graph applications) is non-trivial and challenging, usually ending up with ineffective and inefficient resource utilization. As a result, the delivered performance of these applications on manycore platforms is rarely keeping up with the growing amount of parallel resources.

Using a general-purpose GPUs (GPGPUs) as an example of a typical manycore platform, this talk will demonstrate how irregular computations and data accesses present challenges in 1) intelligent computation balancing among computing cores, and 2) enhancing data locality. I will first introduce our light-weight runtime system that smartly controls the tradeoff between parallelism and incurred overheads for irregular GPU applications. I will then demonstrate an approach which improves the cache performance by dynamically capturing the data reuse in irregular GPU applications. I will summarize my talk with a vision of future software-hardware collaborative design for heterogeneous, exascale computing systems.

Biography

Xulong Tang is a Ph.D. candidate at The Pennsylvania State University where he works with his advisors Dr. Mahmut T. Kandemir and Dr. Chita R. Das. He received his M.S. from University of Science and Technology of China (2013) and his B.S. from Harbin Institute of Technology (China, 2010), both in computer science. His research interests lie in the fields of high-performance computing and parallel computer architectures and systems. He has published extensively in these areas in venues including MICRO, HPCA, PLDI, SIGMETRICS, and PACT.