|Title||:||Towards performance portability for parallel computations|
|Speaker||:||P. Sadayappan (Ohio State University)|
|Details||:||Mon, 15 Dec, 2014 3:00 PM @ BSB 361|
|Abstract:||:||The development of high-performance parallel applications is very
challenging, especially for applications operating on sparse data
structures, since they exhibit irregular data access
patterns. Further, multiple levels of parallelism must be utilized,
including asynchronous distributed-memory parallelism at the outer
level and synchronous parallelism at the inner level. This talk will
discuss three complementary directions of research directed at
assisting the development of efficient parallel applications. First,
we will describe an inspector/executor based approach for automatic
generation of distributed-memory parallel code from sequential input
code, for a class of loop computations with irregular data access. The
second part of the talk advocates the development of standard APIs for
key data structures and operations on them. The specific example of
sparse matrix multiplication of GPUs will be covered. Finally, the
issue of characterizing the inherent data movement costs of different
algorithms will be discussed - an issue of increasing importance with
the cost of data movement, both in terms of time and energy, greatly
eclipsing the cost of performing arithmetic/logic operations.
Speaker Bio: Prof. Sadayappan is a Professor in the Department of Computer Science and Engineering at The Ohio State University. His primary research interests center around performance optimization and compiler/runtime systems for high-performance computing, with special emphasis on high-performance frameworks that enable high productivity for application developers in scientific computing. Two recent projects include a polyhedral framework for automatic parallelization and data locality optimization, and the Tensor Contraction Engine - a domain-specific compiler/runtime system to automatically transform high-level specifications into efficient parallel programs, for a class of high accuracy ab initio models in quantum chemistry.
Prof. Sadayappan obtained a B.Tech from the Indian Institute of Technology, Madras, and M.S. and Ph.D. from Stony Brook University, all in Electrical Engineering.