high performance computing on graphics processing units: hgpu.org

Advances in technology have given rise to applications that are deployed on wireless sensor networks (WSNs), the cloud, and the Internet of things. There are many emerging applications, some of which include sensor-based monitoring, web traffic processing, and network monitoring. These applications collect large amount of data as an unbounded sequence of events and process them to generate a new sequences of events. Such applications need an adequate programming model that can process large amount of data with minimal latency; for this purpose, stream programming, among other paradigms, is ideal. However, stream programming needs to be adapted to meet the challenges inherent in running it in distributed environments. These challenges include the need for modern domain specific language (DSL), the placement of computations in the network to minimise energy costs, and timeliness in real-time applications. To overcome these challenges we developed a stream programming model that achieves easy-to-use programming interface, energy-efficient actor placement, and timeliness. This thesis presents Curracurrong, a stream data processing system for distributed environments. In Curracurrong, a query is represented as a stream graph of stream operators and communication channels. Curracurrong provides an extensible stream operator library and adapts to a wide range of applications. It uses an energy-efficient placement algorithm that optimises communication and computation. We extend the placement problem to support dynamically changing networks, and develop a dynamic program with polynomially bounded runtime to solve the placement problem. In many stream-based applications, real-time data processing is essential. We propose an approach that measures time delays in stream query processing; this model measures the total computational time from input to output of a query, i.e., end-to-end delay.