Automatic Tuning of Libraries and Applications

Retargeting of applications for new architectures has been a major headache for high performance computation, requiring many person-months (or even years) of effort to retune for each new architecture, and even each new model of an established architecture.  In this new era of multi-core microprocessors and availability of petascale platforms, the challenge of achieving portable high performance will grow to a point where it is no longer feasible for application programmers to do this work without support from tools.

Automation of this retuning process has now become a fertile area of computer science research, especially for domain-specific libraries. Most of this work is based on the strategy of using large amounts of computation time to explore a space of different variants of a loop nest, running each variant on the target architecture, and picking the best one.  If this approach can be extended more generally to components and whole programs, it would help avoid the enormous human costs involved in retargeting applications to different machines.  A major research issue is how can tuning time be kept to manageable levels, given that the number of variants in a complete application can be enormous.

This workshop is an adaptation of a previous workshop series at the 2005 and 2006 Los Alamos Computer Science Institute (LACSI) Symposia.  The specific purposes of the workshop are: (1) to solicit feedback from the application and library community about tool support that they would like for tuning their codes; (2) to discuss future research directions in autotuning; and, (3) to strategize with members of the autotuning community on how to move the existing tools into broader use.  For this latter purpose, we will identify commonalities in autotuning systems, discuss strategies for sharing infrastructure, and develop a common suite of benchmarks.

Close window