The open archive for STFC research publications

Full Record Details

Persistent URL http://purl.org/net/epubs/work/53731
Record Status Checked
Record Id 53731
Title Numerical algorithm libraries for multicomputers
Abstract These notes were written to document experiments in the design of a parallel algorithm library for numerical applications in FORTRAN-77. The target computer is any one of the current multiple instruction multiple data (MIMD) distributed-memory multicomputers although the philosophy is also applicable to shared-memory computers. Versions of the software are also available for such UNIX-based multiprocessors and also networked UNIX hosts. I have used Intel iPSC/2, Intel iPSC/860, Meiko Computing Surface with i860 Convex and Alliant multi-computers and workstations. Extensive reference will be made to a previous review [1] on programming of similar parallel machines in FORTRAN-77. The communications paradigm is that of Fortnet [2,3], which is representative of the general scheme of parallel communications on multicomputers. Other public-domain and commercially available harnesses differ only in detail. Extensions of this scheme are illustrated which embody algorithms for global communication and exchange of data between subsets of processes. This is compared with global communications procedures available in other organised libraries of software such as the SUPERNUM project grid library and commercial software which has similar ideas to our own. Finally details of subroutine calling sequences and usage of a number of existing parallel numerical libraries is reviewed. Examples are also given, taken from working programs, which illustrate programming styles. These might serve as an introduction to parallel algorithm design.
Organisation CSE , SERC
Funding Information
Related Research Object(s):
Language English (EN)
Type Details URI(s) Local file(s) Year
Report DL Technical Reports DL-SCI-TM81T. 1991. DL-SCI-TM81T.pdf 1991