Performance estimation techniques for micro-heterogeneous computing systems

Heterogeneous Computing environments are networks of loosely coupled computational nodes with different architectures that map portions of a problem to nodes whose architecture is best suited to that computation. A proposed architecture, called Micro- Heterogeneous Computing (MHC), aims to develop an environment that would facilitate the creation and utilization of computational networks as small as a single workstation. The scheduling algorithms used to schedule computations to MHC nodes rely heavily on computation time performance estimates. Thus an appropriately precise performance estimate is essential in order to enable a scheduling algorithm to make the best scheduling decisions and thus to maximize the performance of the MHC system. The unique needs of the Micro-Heterogeneous Computing environment place specific requirements on performance evaluation techniques that may be used by the MHC scheduling algorithms. This work evaluates the requirements placed on performance estimation techniques by the MHC environment. Several potential performance estimation techniques are evaluated with respect to those requirements. Two empirical techniques - polynomial and Multi-Chord approximation - are selected as the potential candidates and their characteristics are evaluated and compared. Based on this evaluation, means of automatically obtaining performance specifications for MHC computations are created. Several computational algorithms are implemented on a DSP-based peripheral and their performance specification is obtained using the newly developed automatic technique