The authors have worked together over five years
to develop a general methodology for parallel disk based computation.
This includes: construction of a permutation representation for
Thompson's group, acting on 143,127,000 points; construction of a
permutation representation for the Baby Monster group (second largest of
the sporadic simple groups), acting on 13,571,955,000 points; and a
condensation computation for Fi_23 acting on 11,739,046,176 points that
resolves an open problem in the Modular Atlas Project. At heart, these
problems are search problems. The work typically required a
multi-threaded, distributed program using the 30 nodes and corresponding
local disks in parallel. Aggregate disk space used ranged as high as 8
terabytes. The need for the highest efficiency led to the discovery of
one new search method, and the re-discovery of several other search
methods. This is summarized in a taxonomy of parallel disk-based search
algorithms. In addition, a new, open source package is presented that
automates the difficult task of developing such parallel disk-based
software. Lessons learned about the difficulties of such large
computations are also presented.
|