high performance computing on graphics processing units: hgpu.org

Extracting performance improvements from modest and cost-effective computing resources is one of the key challenges in the IT sector. CPU clock speeds have reached a plateau in recent years, with no significant clock speed improvements forthcoming. However, we see an increasing number of computational cores available on the desktop, via the CPU and, more recently, Graphical Processing Units (GPUs). This trend has placed parallel computing to the fore of application performance improvement techniques. In this paper, we select a number of open source extensions to the R Statistical Environment (a desktop test case), and the Postgres-Postgis database (a server-side test case) and seek to address how runtime improvements to the extensions can be made in a cost effective manner. For us, cost implies both development time, and, more critically, development of the programmers skillset to properly understand the design issues inherent in parallel programming. We address the important issue of how easily originally non-parallel code can be parallelized and the software engineering challenge in so doing. Our findings here illustrate that, by carefully parallelizing key specific code sections, impressive performance improvements can be made with modest enhancement to the programmers skillset.