When implementing an apparently parallel algorithm, there are certain issues not emphasized in theoretical discussions which may play an unexpectedly important role in actual performance. In this implementation two such issues come to mind. The first one is the effect column shifting has on performance, even though it appears to be a fairly trivial task. The second one is the drastic effect data layout can have on performance on a CM.
In general, programming on the CM has required a struggle to get acquainted with the CM FORTRAN compiler's many idiosyncrasies. A great deal of effort was made to achieve the current less than optimal performance. However, once the optimization phase is complete, SEP will have an efficient program to obtain the SVD of medium to large matrices. At that point, running the program on a larger CM will allow processing of most matrices of interest in geophysical data analysis.