The SC10 Doctoral Showcase is great forum for Ph.D. students in HPC, networking, storage and analysis. To qualify, Doctoral candidates who plan on graduating within the next 12 months submit a short summary of their research. Selected speakers get to present a 15-minute summary of their best research to experts in academia, industry and research laboratories.
We present Parallel Active Harmony — a scalable end-to-end auto-tuner for scientific codes. Harmony takes a search-based collaborative approach to auto-tuning. Application programmers and end-users collaborate to describe and export a set of performance related parameters to the Harmony system. These parameters define a tuning search-space. The auto-tuner monitors the program performance and suggests application adaptation decisions. The decisions are made by a central controller using a parallel search algorithm. Recently, we have been working on making online tuning practical. Our goal is to enable application developers to write applications once and have our auto-tuner adjust the application behavior automatically when run on new systems. We evaluated our system on a parallel multi-block lattice Boltzmann code. The code has five hot-regions and all five are tuned simultaneously at runtime. Performance improved up to 48% for a 512-core run with a minimal cost of 2% (8 cores) extra cores for code-generation.
The Doctoral Showcase has an impressive lineup of talks in store for SC10, so why did I pick this one? Well, I think the world could use a whole lot more Parallel Active Harmony, don’t you?