This work seeks to construct data streaming abstractions for clusters composed of graphics processing units (GPUs). The proposed work covers development of stream operations, such as abstract split/join operations, demonstrated through concrete text application, such as text search, for massive data processing requirements in clusters of GPUs. We propose to demonstrate the feasibility of GPU clustering for data-dependent programming problems with fixed input data as well as live data streams.

The outline of this work includes:

• Design the streaming abstraction and language syntax suitable for GPU processing [10/31/09];
• Implementation of run-time system that efficiently controls data flow in streaming applications [11/10/09];
• Source to source conversion from streaming kernel abstraction to CUDA source code [optional 11/14/09];
• Validate the concept by realizing streaming benchmarks on single GPU or multi-GPU platform [11/30/09];
• Real-world streaming applications such as on-line intrusion detection, text mining [optional 12/10/09].