Indiana University is remarkably well equipped in HPC systems. You can view a list of what is available at http://www.indiana.edu/~rats/.
Our main research system is the IBM SP. It is a cluster system connected by a special high bandwidth switch network, which provides point-to-point connectivity at 150 MB/s between a variety of nodes, beginning with 136 4-way POWER3+ SMPs, 4 16-way POWER3+ SMPs, and one 16-way POWER4 SMP. This system also has a small number of other nodes, which ``don't quite fit'' and are used for auxiliary tasks. The General Parallel File System installed on our SP provides nearly 3 TBs of storage. All CPUs mentioned here are 64-bit CPUs, which can perform up to 4 double precision floating point operations in a single clock cycle, albeit on special operations (add-multiply) only.
The whole system is gauged at about 1 TFLOPS peak. Because its nodes are quite powerful SMPs (and the 16-way nodes are very powerful SMPs), the system can be used to run a large number of shared memory jobs and this makes it extremely useful in our environment.
The next system is made of the four AVIDD clusters. These are the clusters that we are going to use in this course. There is a lot of spare computational capacity on this system at present and, because it is so similar to the NCSA clusters, it can be also used as a training and testing ground for the latter. Its primary purpose is ``Analysis and Visualization of Instrument Driven Data'', hence its name, AVIDD, and the rationale for this course, ``High Performance Data Management and Processing''.
We are going to dedicate the whole chapter 3 to this cluster, so we are not going to spend more time on it here.
Although we have HPC clusters at IU, we don't have ``real supercomputers'', like Cray X1, NEC SX6, Cray T3E or Hitachi (University of Tokyo has one of these that can run LINPACK at 1.7 TFLOPS) for the following reason. They are very good for certain types of typical supercomputer jobs, but no good for a very broad range of jobs. At our university we might perhaps have two or four users who would benefit from such systems, but they would be pretty useless to the vast majority of our users, who are not typical supercomputer users. On the other hand clusters, although not efficient as supercomputers, can be used to run just about anything, beginning with little sequential scalar jobs, or a lot of little sequential jobs, and ending on parallel jobs of various sizes: some requesting 16 CPUs only, some running on 64 2-way nodes.