Optimization Research Area

Tuning for Multiple Objectives: Power versus Performance
Lead: Paul Hovland

When a single objective, such as execution time, is available, the autotuning search
problem can be posed as a numerical optimization problem. However, it is
increasingly common to have multiple objectives, such as execution time, energy
consumption, resilience to errors, power demands, and memory footprint. When the
relative weights or constraints on these objectives are not known at search time,
one must pose the autotuning search problem as a multi-objective optimization
problem. We describe the optimization framework for search, discuss some of the
potential tradeoffs among multiple objectives, and provide empirical evidence that
such tradeoffs do exist in practice.

Optimization poster for the Sept 2012 SciDAC PI meeting