Scientific computing on public computing platforms — practices and experiences

Public computing is a type of grid computing architectures composed of autonomous volunteer workstations coordinated by a central server complex over the Internet. Exploiting these distributed resources comes at the price of unpredictable availability, fluctuating performance, and heterogeneous participant nodes. Central to addressing these problems is an efficient and accurate scheduling mechanism for the public computing architecture. In this talk, I will overview grid computing in general and a public computing platform called BOINC in particular. Several major issues and challenges in grid computing research will be identified. I will also report our new research results in this exciting area through introducing a scheduling system based on the ant colony algorithm. In addition, the bottleneck problem in the client-server architecture used in BOINC will be identified and a new peer-to-peer architecture for solving the problem will be presented. Several applications have been implemented on BOINC and our new architecture using our scheduling system. By testing these applications with totally different characteristics, we show that our scheduling system consistently provides a high performance, adaptive solution to all of them, and the system using our new peer-to-peer architecture indeed performs better than BOINC. Our work also manifests that public computing can solve the problems of large computing power requirement and huge memory demand in many scientific applications and potentially replace high-cost supercomputing systems for certain application domains.