IBM RS/6000 cluster

arapahoe, comanche, navaho, chief
Clock Speed: 
Dates Used: 
Saturday, February 1, 1992 to Monday, September 1, 1997
Microprocessor Peak Teraflops: 
Memory (terabytes): 
Number of Processors: 
Experimental and Production

On June 1, 1992, SCD made the IBM RS/6000 Cluster available to users on a production basis as a low-cost alternative to supercomputing on the CRAY Y-MP8/864 (shavano). The Cluster was an experimental project, used for early experiments in parallel code development and in the support of non-Cray serial compute servers.

The Cluster originally consisted of two Model 590 RS/60000 nodes (arapahoe and comanche) and a model 550 RS/6000 550 gateway server (chief). The nodes were neworked together and ran IBM's UNIX-based AIX operating system. Chief provided the user gateway functions, the login point, the batch submittal facility, and spooling of data to and from the Mass Storage System.

The nodes had large memories (512 megabytes for arapahoe, 384 for chief), 6 gigabytes of disk space, and a fast clock (66 MHz). The gateway server had 128 megabytes of memory, 6 gigabytes of disk, and a clock speed of 42.5 MHz.

Later another IBM RS/6000 550 node (navajo) was added for exclusive use by NCAR's Climate and Global Dynamics division.

The Cluster was suited for single-tasked jobs since the nodes ran independently, without Parallel Virtual Machine (PVM) or other software provision for job parallelization. The Cluster was also suited for jobs running NCAR Graphics or requiring intense scalar computation or for interactive work or Fortran 90 job development. However, jobs on the Cluster had to have modest input/output (I/O) requirements limited to tens, rather than hundreds, of megabytes per hour.

The Cluster was not binary compatible with the Cray systems, although local software was available to help with tasks such as reading Cray binary data sets with numeric conversion, timing routines, and data compression routines.

The IBM Cluster was decommissioned the first week in October 1997.

Output of the SP1 was ~26 megaflops on a single processor, which was multiplied by an estimated .15 efficiency for parallel use.

Software included math libraries, the Distributed Queueing System (DQS) for batch job submittal, Cray conversion routines, the Fortran 90 XLF compiler, and Parallel Virtual Machine (PVM).