Why Build a Cluster Computer?

The graph below is a plot of the performance of the Ames Lab Classical Molecular Dynamics code for several common parallel systems (Spring of 2003). Medusa is an 11-node 2.4 GHz dual-Xeon Gigabit Ethernet cluster from Atipa. The code runs very well on the Xeon processor when the Intel compiler is used. The CMP cluster is a 24-node AMD Athlon MP2100 cluster connected by Gigabit Ethernet. The performance per processor is not as good using the Athlons, but it is still respectable. The ALCMD code has scaling problems on the IBM SP even when run on the 16 processes of 1 SMP node. The off-node communication does not seem to hurt when going to 64 processors (4 SMP nodes), so the code would likely scale well to many more processors.

The cluster computers provide much better performance for the price compared to traditional SMP and MPP systems. The main limitation of the clusters in this case is that reasonably priced Gigabit Ethernet switches are available only up to 24-ports. The larger MPP systems are scalable up to thousands of processors. Many also have much better communication performance.

Links to more advanced topics

Ames Laboratory | Condensed Matter Physics | Disclaimer | ISU Physics