DATE: Thursday, February 7, 2013
TIME: Noon - 1:00 pm
PLACE: CIC 4th Floor ISTC Panther Hollow Conference Room

SPEAKER: Guy Blelloch, CMU

TITLE: The Problem Based Benchmark Suite

ABSTRACT:
With the advent of dozens of different parallel programming models, and many different parallel architectures, it can be very hard to do an honest comparison among them. This is especially true for running more unstructured problems such as problems on graphs, strings, or trees. In the talk I will present a benchmark suite specifically designed to compare different programming models and architectures, and present some preliminary results on implementing these benchmarks. The benchmarks are defined in terms of the specific problem they solve and allow for one to code them using any programming language and run them on any architectural platform. The benchmarks include problems such as sorting strings, minimum- spanning-trees, breadth-first-search, building a suffix tree, Delainey triangulation, ray casting, and nearest neighbors. The quality of a solution is considered in terms of both runtime, but also simplicity of the code.

In the talk I will present some preliminary results on coding the benchmarks on a particular platform, a commodity multicore. I will discuss what approaches were needed to get good performance in many of the benchmarks. On an 40 core Nehalem machine we were able to get at least a 10 fold speedup over the best sequential implementation on all benchmarks, and very close to 40 fold on some of them. I will also give some comparison to implementations on various distributed platforms.

BIO:
Guy Blelloch received a BA from Swarthmore College in 1983 and a PhD degree from MIT in 1988. His research interests are in in the interaction of parallel programming languages, systems, and algorithms. He is an ACM Fellow for his contributions in parallel computation and is General Chair of the ACM Symposium on Parallelism in Algorithms and Architecture. He has written well over 100 articles and several book chapters on parallel algorithms, applications and, issues in programming languages.

HOST: Phil Gibbons

SDI / ISTC SEMINAR QUESTIONS?
Karen Lindenfelser, 86716, or visit www.pdl.cmu.edu/SDI/