high performance computing on graphics processing units: hgpu.org

Computer engineers are continually faced with the task of translating improvements in fabrication process technology (i.e., Moore’s Law) into architectures that allow computer scientists to accelerate application performance. As feature-size continues to shrink, architects of commodity processors are designing increasingly more cores on a chip. While additional cores can operate independently with some tasks (e.g. the OS and user tasks), many applications see little to no improvement from adding more processor cores alone. For many applications, heterogeneous systems offer a path toward higher performance. Significant performance and power gains have been realized by combining specialized processors (e.g., Field-Programmable Gate Arrays, Graphics Processing Units) with general purpose multi-core processors. Heterogeneous applications need to be programmed differently than traditional software. One approach, stream processing, fits these systems particularly well because of the segmented memories and explicit expression of parallelism. Unfortunately, debugging and performance tools that support streaming, heterogeneous applications do not exist. This dissertation presents TimeTrial, a performance measurement system that enables performance optimization of streaming applications by profiling the application deployed on a heterogeneous system. TimeTrial performs low-impact measurements by dedicating computing resources to monitoring and by aggressively compressing performance traces into statistical summaries guided by user specification of the performance queries of interest.