high performance computing on graphics processing units: hgpu.org

Matching incoming event notifications against received subscriptions is a fundamental part of every publish-subscribe infrastructure. In the case of content-based systems this is a fairly complex and time consuming task, whose performance impacts that of the entire system. In the past, several algorithms have been proposed for efficient content-based event matching. While they differ in most aspects, they have in common the fact of being conceived to run on conventional, sequential hardware. On the other hand, modern Graphical Processing Units (GPUs) offer off-the-shelf, highly parallel hardware, at a reasonable cost. Unfortunately, GPUs introduce a totally new model of computation, which re- quires algorithms to be fully re-designed. In this paper, we describe a new content-based matching algorithm designed to run efficiently on CUDA, a widespread architecture for general purpose programming on GPUs. A detailed comparison with SFF, the matching algorithm of Siena, known for its efficiency, demonstrates how the use of GPUs can bring impressive speedups in content-based matching. At the same time, this analysis demonstrates the peculiar aspects of CUDA programming that mostly impact performance.