Exploração de múltiplos algoritmos de escalonamento de núcleos OpenCL para sistemas de processamento em nuvem multi-inquilino baseados em CPUS e GPUS

Abstract

Cloud computing provides services of shared computational resources for the execution of applications by tenants through computation offloading. To deal with the computational load, the cloud’s nodes can integrate multi-core processing units with graphics processing units. The system’s heterogeneity covers applications with massive instruction, thread and data parallelism, which can be explored through the OpenCL programming model, leading to reductions in the application’s execution time for the tenants and a reduction of energy consumed to the cloud provider. This work proposes a framework for the execution of OpenCL applications on a multi-tenant cloud environment composed of a CPU and a GPU, scheduling the applications to the best device transparently, with the objective of best utilizing the computing resources, reducing execution time and energy consumed. Five scheduling algorithms are explored: First Come First Served, Weighted Round Robin, Max Min, Min Min e First Fit. The execution of the PolyBench benchmark suite is used to evaluate the performance of the framework in hardware configurations composed of an Intel Xeon Haswell processor combined with an NVIDIA Tesla K20m, NVIDIA Tesla K80 or NVIDIA GeForce 1080 Ti. For all of these configurations, First Fit is the highlighted algorithm, achieving an average acceleration of 3,04 up to 3,80 times and an average energy reduction of 33% when compared to GPU exclusive execution.