Revolutions in program analysis and optimizations are imperative for
meeting the challenges imposed by the trend towards exascale computing
and the processor development towards massive parallelism and
heterogeneity. This research addresses these challenges in two
aspects. The first is on how to make compilers and runtime systems
fully understand and accurately predict a program's dynamic
behaviors. Such predictive capability is fundamental: Virtually all
optimizations rely on the knowledge of how a program will behave to
make appropriate decisions. The second is on how to translate the
understanding of program behaviors into performance by effectively
exploiting the memory hierarchy and growing concurrency in modern
computing systems.

The talk starts with an overview on both aspects, and then concentrate
on a new paradigm---input-centric paradigm---for program behavior
analysis and optimizations. The new paradigm is distinctive in
bringing program inputs into the focus of program optimizations. It
addresses some dilemma in traditional program behavior analysis, and
creates many new opportunities for modern computing.

Short Bio:

Xipeng Shen is an Assistant Professor of Computer Science at the
College of William and Mary. He received his Ph.D. and Master degrees
in Computer Science from University of Rochester. He is an IBM CAS
Faculty Fellow, and a recipient of the Best Paper Award of ACM PPoPP
2010 and the National Science Foundation CAREER Award. His research in
Compiler Technology and Programming Systems aims at helping
programmers achieve high performance as well as good programming
productivity on both uniprocessor and multiprocessor architectures. He
is particularly interested in the effective usage of memory
hierarchies, the exploitation of program inputs in program behavior
analysis, and the employment of statistical learning in runtime
systems and dynamic program optimizations.