INCEPTION
In truth, the LIB Cluster was created in 2005 as a tool for experimenting with Beowulf cluster administration. A strict policy was maintained: no out-of-pocket money could be spent on the LIB cluster equipment. Compute nodes at the end of their life-cycle were gathered from curbsides and recycling depots, and non-useful machines were stripped for parts. Sales of parts (such as SGI parts and older SCSI hard drives) were used to purchase networking cable, switching equipment, and power strips.

INHOMOGENEOUS CLUSTERING
Like other aged and lower-performance clusters, the LIB cluster eventually proved to be useful as a test-platform for in-house code development and debugging. When used homogeneously, the LIB cluster peak performance is limited by the slowest machine. However, when used inhomogeneously at the maximum computational efficiency of each node, the peak performance (ie. Gflops) should match that of a homogeneous cluster of the same performance. It should be noted that only certain problem types are amenable to inhomogenous clustering, such as embarissingly parallel or exploratory problems with master/slave parallelization. Ring-buffer parallelizations are not amenable, as the ring would be clogged at the slowest node. Additionally, to successfully implement inhomogeneous clustering, the computational domain needs to be scattered to each node according to an evaluation of relative performance.

PROJECT GOALS
1)Conceive a methodology for evaluating relative nodal performance
2)Develop and maintain a library of inhomogeneous scatter/gather wrappers for MPI functions
3)Benchmark the performance of LIB cluster to equivalent LHB (homogeneous) cluster.

CURRENT STATUS
Currently, an open-source library (goals #1 and #2) is in the process of being developed and tested. The beta version is expected to be released by 2008.