high performance computing on graphics processing units: hgpu.org

Current trends in multicore platform design indicate that heterogeneous systems are here to stay. Such systems include processors with specialized accelerators supporting different instruction sets and different types of memory spaces among several other features. Unfortunately, these features increase the effort for programming and porting applications to different target platforms. To solve this problem, effective programming strategies that can exploit the rich feature set of such heterogeneous multicore architectures are required without leading to an increased learning curve for applying these strategies. In this paper, we explore the Multicore Task Management API (MTAPI), a standard for task-based parallel programming in embedded systems. MTAPI provides interfaces for decomposing a problem into a set of tasks while abstracting from the heterogeneous nature of the target platforms. This way, it allows to efficiently distribute work in systems consisting of different cores running at different speed. For evaluation purposes, we use an NVIDIA Jetson TK1 board (ARM + GPU) as our test bed. As applications, we employ codes from benchmark suites such as Rodinia and BOTS.