high performance computing on graphics processing units: hgpu.org

Modern systems generates huge amounts of information right from areas like finance, telematics, healthcare, IOT devices to name a few, the modern day computing frameworks like Mapreduce needs an ever increasing amount of computing power to sort, arrange and generate insights from the data. This project is an attempt to harness the power of heterogeneous computing, more specifically take benefit of parallelism offered by modern day GPU’s and accelerate data processing of Mappers and Reducers in the Mapreduce framework. In this regard, AMD’s opensource APARAPI library was used as the foundation to develop a simple to use heterogeneous framework. The framework uses both the CPU and GPU to perform processing by translating sections of Mapreduce Java bytecodes to OpenCL and executing the same in parallel in AMD GPU’s. Typical Mapreduce tasks were then run in this framework with various types of workloads and while no perceptible improvement in speedups were noticed the framework shows excellent promise in computing complicated parallel data in upcoming high performance computing workloads.