Autotuning de aplicações paralelas em sistemas heterogeneous

Abstract

Nowadays computational platforms have been evolving to the high computational power direction, however it requires a lot of energy to achieve such high performance with single but powerful processing unit.To manage this energy cost and keep with high performance, computers are built under the assumption of heterogeneous systems, in other words, computers that have different kind of processing units with different functions, such as CPU, GPU, Xeon Phi and FPGA. So, developers should take advantage of parallel activity and scheduling tasks by using the various parts of the heterogeneous systems.Now the problem is how to efficiently achieve the highest performance possible when running software applications by taking the most advantage of such heterogeneous systems and keeping the energy cost at the minimum level without jeopardizing the application performance and its results. Overall, the problem consists in the coexistence work of multicore, its parallelism and its shared cache problems; CPU and GPU parallelism and scheduling tasks; performance; and energy costs.For this problem's solution is expected to find/create an autotuner, or at least a concept proof, that can achieve the best performance in a software application by enhancing the application's code automatically in a level that takes the best benefit of the available hardware without elevated energy costs. To do so, after creating its code, the developer runs the autotuner and it will enhance, automatically, the code to get the best performance.This kind of solution requires some validation process and metrics to make sure that it is doing its work and with proper results. To do so, the idea of the process' validation is going to be about comparing the behaviour of three different codes: a version of a serialized code; a version of the same code but with an expert manually paralleling it; and a version of the serialized code but automatically parallelized. The metrics that will be used to compare these three code versions are the following: processing power; execution time; number of memory accesses; and energy consuming. With this solution, applications will achieve its highest performance possible in an automatic way and developers will have less burdened about creating parallel code, consequently, saving them time.