Date: Feb 15, 2016 12:00pm - 1:00pm
Location: GDC 6.302
Talk Audience: Faculty, Graduate Students, Undergraduate Students, Outside Interested Parties, ECE, ICES
Talk Abstract: Writing high-performance parallel implementations of graph algorithms is challenging. Standard compiler parallelization techniques are usually inapplicable due to the irregular parallelism exhibited by these algorithms. On the CPU, therefore, the standard strategy is to expose and exploit irregular parallelism using sophisticated runtime frameworks such as Galois. Unfortunately, Graphics Processing Units (GPUs) make it difficult to use this strategy.
The GPU is an attractive target for graph algorithms because it supports massive parallel execution and features much higher memory bandwidth than the CPU. However, its programming model is also very restricted. In particular, communication among threads is limited, so standard strategies for irregular programs (such as runtime... Read more