Table of Contents

Selectively Consistent Approximate Parallel Execution

Friday March 24, 2017
Location: CIC Panther Hollow Room
Time: 4:30PM

Vignesh Balaji (CMU)


Programmers writing shared memory parallel applications use synchronization operations for correct manipulation of shared data. They also rely on hardware based cache coherence to automatically transport data between cores. While synchronization and cache coherence are required for correct execution of parallel applications, they impose a significant performance penalty. The serialization and data movement overheads of these operations impact scalability of parallel applications. Research in approximate computing has demonstrated the error tolerance of many important application domains by improving performance with minimal impact on output quality. In this talk, I will present our work on SCAPE - a system that exploits this resilience towards errors to eliminate synchronization and cache coherence for performance. Our system allows the programmer to control the quality of such approximate executions by ensuring precise updates only for program values deemed quality critical by the programmer. Additionally,  SCAPE uses a neural network to selectively approximate executions only when the expected performance improvement justifies approximation. Our evaluations show that SCAPE can improve performance up to 20X while keeping the quality degradation less than 1% for a collection of graph applications.


Vignesh Balaji is a second year PhD student co-advised by Professors Brandon Lucia and Radu Marculescu. His research is focused on approximate computing with a particular emphasis on reducing data movement in multi-core processors.

Back to the seminar page