The International Conference for High Performance Computing, Networking, Storage and Analysis
Towards Co-Evolution of Auto-Tuning and Parallel Languages.
Student: Ray S. Chen (University of Maryland)
Supervisor: Jeffrey K. Hollingsworth (University of Maryland)
Abstract: There exists a gap between the massive parallelism available from today's HPC systems and our ability to efficiently develop applications that utilize such parallelism. Several emerging programming languages aim to narrow this gap through high-level abstractions of data and/or control parallelism. However, this requires the compiler to make performance affecting decisions based on information unavailable at compile time, such as workload complexity or thread distribution. By evolving auto-tuning along with these languages, such decisions could be deferred until run time when it's possible to test for optimal values.
In this poster, we present our vision of co-evolution by pairing the Chapel programming language with the Active Harmony auto-tuning framework. We include results of our co-evolution experiments which demonstrate the ability to automatically improve the performance of a proxy application over the default by nearly 25%.