high performance computing on graphics processing units: hgpu.org

A variety of hardware platforms for signal processing has emerged, from distributed systems such as Wireless Sensor Networks (WSNs) to parallel systems such as Multicore Programmable Digital Signal Processors (PDSPs), Multicore General Purpose Processors (GPPs), and Graphics Processing Units (GPUs) to heterogeneous combinations of parallel and distributed devices. When a signal processing application is implemented on one of those platforms, the performance critically depends on the scheduling techniques, which in general allocate computation and communication resources for competing processing tasks in the application to optimize performance metrics such as power consumption, throughput, latency, and accuracy. Signal processing systems implemented on such platforms typically involve multiple levels of processing and communication hierarchy, such as network-level, chip-level, and processor-level in a structural context, and application-level, subsystem-level, component-level, and operation- or instruction-level in a behavioral context. In this thesis, we target scheduling issues that carefully address and integrate scheduling considerations at different levels of these structural and behavioral hierarchies. The core contributions of the thesis include the following. Considering both the network-level and chip-level, we have proposed an adaptive scheduling algorithm for wireless sensor networks (WSNs) designed for event detection. Our algorithm exploits discrepancies among the detection accuracy of individual sensors, which are derived from a collaborative training process, to allow each sensor to operate in a more energy efficient manner while the network satisfies given constraints on overall detection accuracy. Considering the chip-level and processor-level, we incorporated both temperature and process variations to develop new scheduling methods for throughput maximization on multicore processors. In particular, we studied how to process a large number of threads with high speed and without violating a given maximum temperature constraint. We targeted our methods to multicore processors in which the cores may operate at different frequencies and different levels of leakage. We develop speed selection and thread assignment schedulers based on the notion of a core’s steady state temperature. Considering the application-level, component-level and operation-level, we developed a new dataflow based design flow within the targeted dataflow interchange format (TDIF) design tool. Our new multiprocessor system-on-chip (MPSoC)-oriented design flow, called TDIF-PPG, is geared towards analysis and mapping of embedded DSP applications on MPSoCs. An important feature of TDIF-PPG is its capability to integrate graph level parallelism and actor level parallelism into the application mapping process. Here, graph level parallelism is exposed by the dataflow graph application representation in TDIF, and actor level parallelism is modeled by a novel model for multiprocessor dataflow graph implementation that we call the Parallel Processing Group (PPG) model. Building on the contribution above, we formulated a new type of parallel task scheduling problem called Parallel Actor Scheduling (PAS) for chip-level MPSoC mapping of DSP systems that are represented as synchronous dataflow (SDF) graphs. In contrast to traditional SDF-based scheduling techniques, which focus on exploiting graph level (inter-actor) parallelism, the PAS problem targets the integrated exploitation of both intra- and inter-actor parallelism for platforms in which individual actors can be parallelized across multiple processing units. We address a special case of the PAS problem in which all of the actors in the DSP application or subsystem being optimized can be parallelized. For this special case, we develop and experimentally evaluate a two-phase scheduling framework with three work flows – particle swarm optimization with a mixed integer programming formulation, particle swarm optimization with a simulated annealing engine, and particle swarm optimization with a fast heuristic based on list scheduling. Then, we extend our scheduling framework to support general PAS problem which considers the actors cannot be parallelized.