Colloquium
Greg Bronevetsky
LLNL
Friday, March 18, 2011
3147 MEB
Refreshments 3:20 p.m.
Lecture 3:40 p.m.
Title: Compiler Analyses for Explicitly Parallel Applications
Greg Bronevetsky
LLNL
Friday, March 18, 2011
3147 MEB
Refreshments 3:20 p.m.
Lecture 3:40 p.m.
Title: Compiler Analyses for Explicitly Parallel Applications
Abstract
Parallel computing has become an integral part of computer science research and practice. Today parallel systems appear in all types of devices, including desktops, routers, graphics cards and servers, with the largest systems reaching hundreds of thousands of computing cores. This growth in parallel computing is driving a breadth of research in programming models that enable application developers to exploit the power of these systems. The most popular of these are explicit models, such as messages passing and shared memory, where the developer explicitly describes the communication to be performed. Despite significant work to improve runtime support for these programming models, there has been relatively little research on compiler analyses that can understand explicitly parallel applications.
I will discuss my work on developing a new framework that extends dataflow analyses to explicitly parallel applications in a way that is sensitive to the applications' communication topology. By defining dataflow over a parallel extension of traditional control flow graphs, this analysis enables a variety of novel optimizations, transformations, debugging tools and source code validation tests. Furthermore, the use of key symbolic abstractions enables the framework to support applications that may execute on arbitrary number of processes without suffering from exponential algorithmic blowup. The result is a new bridge between decades of research on dataflow and symbolic compiler analyses and explicit parallel programming models. This framework is thus a key enabler for a variety of future techniques that will use compiler intelligence to support and enhance application development on parallel systems.
Finally, I will discuss a new project at LLNL on using the above compiler framework to develop analyses and transformations for MPI applications and the work we hope to do in the near future on extending the current MPI specification to make it more amenable to compiler analyses and enable easy composition of parallel components.
Parallel computing has become an integral part of computer science research and practice. Today parallel systems appear in all types of devices, including desktops, routers, graphics cards and servers, with the largest systems reaching hundreds of thousands of computing cores. This growth in parallel computing is driving a breadth of research in programming models that enable application developers to exploit the power of these systems. The most popular of these are explicit models, such as messages passing and shared memory, where the developer explicitly describes the communication to be performed. Despite significant work to improve runtime support for these programming models, there has been relatively little research on compiler analyses that can understand explicitly parallel applications.
I will discuss my work on developing a new framework that extends dataflow analyses to explicitly parallel applications in a way that is sensitive to the applications' communication topology. By defining dataflow over a parallel extension of traditional control flow graphs, this analysis enables a variety of novel optimizations, transformations, debugging tools and source code validation tests. Furthermore, the use of key symbolic abstractions enables the framework to support applications that may execute on arbitrary number of processes without suffering from exponential algorithmic blowup. The result is a new bridge between decades of research on dataflow and symbolic compiler analyses and explicit parallel programming models. This framework is thus a key enabler for a variety of future techniques that will use compiler intelligence to support and enhance application development on parallel systems.
Finally, I will discuss a new project at LLNL on using the above compiler framework to develop analyses and transformations for MPI applications and the work we hope to do in the near future on extending the current MPI specification to make it more amenable to compiler analyses and enable easy composition of parallel components.