In search of clusters (2nd ed.)
In search of clusters (2nd ed.)
I'm Not Going to Pay a Lot for This Supercomputer!
Linux Journal
High Performance Cluster Computing: Architectures and Systems
High Performance Cluster Computing: Architectures and Systems
| Hi-index | 0.00 |
During recent years clustered systems using off-the-shelf processors and standard Ethernet networks have been increasingly popular. The motivation has been primarily the cheap price of systems, but also the rapid development of standard processors. So-called Beowulf systems have spread around the world. This development is further accelerated by the LINUX-boom which provides an ideal and free operating system for these clusters. In the end of 1998 a pilot project was started at CSC in order to study PC clusters for parallel computing. The aim was to gather experience about performance, bottlenecks and quality of programming tools. We will present in this paper our experiences in building a 32-processor PC cluster and running parallel applications on it. The initial problems in getting the systems to work properly and tuned represented a fair workload. Data communication bottlenecks turned out to be a major performance issue. A small number of parallel scientific applications, mostly running on Cray T3E or other parallel systems, was transferred to the PC cluster environment in order to get porting experience. A large set of benchmarks was run on the cluster, in addition to the benchmarking of the ported applications.