Increasing Resource Utilization in Heterogeneous Architectures

Colloq: Speaker: 
Mehmet Belviranli
Colloq: Speaker Institution: 
UC Riverside
Colloq: Date and Time: 
Fri, 2016-09-23 10:00
Colloq: Location: 
Building 5700, Room MS-A106
Colloq: Host: 
Jeff Vetter
Colloq: Host Email: 
vetter@ornl.gov
Colloq: Abstract: 
Heterogeneous systems employ accelerators (e.g. GPU, FPGA and many-core) to offer massive parallelism. Scientific applications benefit from these systems by offloading their data-parallel regions to accelerators. The heterogeneity across these architecturally diverse computation units, however, may introduce many challenges to software design. Better multi-processor utilization and efficient data transfers are among the most important objectives that application developers target in order to maximize the total system throughput. In this talk, I will introduce both accelerator-level and system-wide techniques to improve utilization of various resources in heterogeneous architectures. In the first part, I will describe a new synchronization technique that replaces global barriers in wavefront parallelism to reduce SM idleness in GPUs. In the second part, I will present a novel scheduler for efficient data transfer and execution overlapping to increase accelerator and interconnect utilization. The proposed studies demonstrate significant speedups over traditional heterogeneous programming techniques like fork/join type of parallelism and the default pipelined data transfers. Our techniques can be leveraged further to obtain higher execution efficiency for a wide range of scientific applications on larger scale computing platforms.
Colloq: Speaker Bio: 
Mehmet Esat Belviranli has recently received his Ph.D. degree in Computer Science from University of California, Riverside under the supervision of Prof. Laxmi Bhuyan. His research interests are on systems. Particularly, he focuses on developing run-time solutions and scheduling algorithms to increase multi-processor utilization on heterogeneous architectures. He has authored multiple papers in ICS and also published in PACT and TACO. He received his M.S. and B.S. in Computer Science from Bilkent University, Turkey.