Scheduling strategies for parallel patterns on heterogeneous architectures

To help shrink the programmability-performance efficiency gap, we discuss that adaptive runtime systems can be used to facilitate the management of heterogeneous architectures. A runtime system can provide a significant performance boost while reducing the energy consumption, because it is aware of processors’ architectures and application’s requirements. We analyse how applications map onto hardware by inspecting built-in processor counters, and therefore build models to describe the observed behaviour. In this thesis, we discuss how parallel patterns, such as parallel for loops and pipelines, can be decomposed and efficiently executed on heterogeneous plat- forms. We propose several scheduling strategies aiming at reducing execution time and energy consumption. We demonstrate how applications can be run faster by mapping the application level parallelism onto the hardware process- ing units that best fit the application requirements, and by selecting the right task size. First, we devise a load balancing technique, that targets heterogeneous CPU and multi-GPU architectures. It monitors the relative speed of each processing unit, and distributes the remaining workload based on these relative speeds. By making all processing units to finish at same time, we avoid unnecessary waits between processors. Along with this load balancing technique, we propose a performance-sensitive partitioner that adapts the amount of computation offloaded to the accelerator for better performance and utilisation. We also present an accurate performance model for streaming applications, such as face recognition or object tracking. This model targets pipelined applications, as a series of stages, and performs a scalability analysis of each stage by using coarse and medium grain parallelism. Additionally, it also considers executing the stage on the GPU or not. By applying the model, we always find the best pipeline configuration among all possible, and get substantial performance and energy savings. All experiments in this thesis have been performed by using state-of-the-art hardware accelerators and benchmarks of the field of HPC. Specifically, we use the Rodinia and SHOC benchmark suites, for the evaluation of the parallel for partitioner. Moreover, we use the the ViVid application, along with tracking and SRAD applications from Rodinia Benchmark Suite, all of them are good candidates of vision applications. Finally, we rely on Intel Threading Building Blocks, the core engine of our schedulers; the Intel OpenCL SDK and CUDA SDK to offload computations to the GPU accelerators and Intel PCM library to monitor energy consumption and cache memory metrics.During the last decade, power consumption and energy efficiency have become key aspects in processor design. Nowadays, the power consumption is the principal limitation for further scaling of chip multiprocessors design (CMPs). In general, the research community agrees that current chip multiprocessor technology trends will not scale performance without an increase of power budget. Hardware design innovations as the recent Heterogeneous Architectures and Near Threshold Computing are needed to cope with the performance-power barrier. As a result of this, there has been a shift away from chip multiprocessors to heterogeneous processor architectures. Recently, we have witnessed an explosion in the availability of this kind of architectures. Many hardware vendors have released a number of heterogeneous processors to overcome the aforementioned limitations. However, software also requires changes to allow further performance scaling on these architectures. With the advent of heterogeneous architectures, hardware manufactures have impose the burden of explicit accelerator management on software developers. In general, programmers are used to sequential programming, but writing high-performance programs for heterogeneous architectures is a complex task. Programming for this kind of platforms requires the understanding of new hardware concepts, orchestration of different parallelism levels, the explicit management of different memory spaces and synchronisations between processing units, and finally the usage of low-level programming models such as OpenCL or CUDA. Moreover, heterogeneous architectures suffer from performance portability, as one program can exhibit unequal performance on different devices