A framework for remote dynamic program optimization

Abstract

Dynamic program optimization allows programs to be generated that are highly tuned for a given environment and input data set. Optimization techniques can be applied and re-applied as program and machine characteristics are discovered and change. In most dynamic optimization and compilation frameworks, the time spent in code generation and optimization must be minimized since it is directly reflected in the total program execution time. We propose a generic framework for remote dynamic program optimization that mitigates this need. A local optimizer thread monitors the program as it executes and selects program sections that should be optimized. An optimizer, running on a remote machine or a free processor of a multiprocessor, is then called to actually perform the optimization and generate a new code variant for the section. A dynamic selector is used to select the most appropriate code variant for each code interval based upon the current runtime environment. We describe this framework in detail and present an example of its use on a simple application. We show that our framework, when used with changing input, can outperform the best statically optimized version of the application.