A Message-Passing Model for Highly Concurrent Computation
Abstract
Based on essentially the same computational model as the Torus, the cube uses a binary n-cube network instead of toroidal mesh, with the requested hardware upgrading of the nodes: a more powerful processor with a floating-point co-processor, a larger storage, full-duplex asynchronous channels. A version of C, extended with communication primitives is available as programming language, together with the required support software for loading the programs into the nodes. An important difference with the Torus, however, is the availability in the Cosmic Cube of a deadlock-free protocol for routing messages between any two nodes. The Torus was built on the premiss that locality of communication should be maintained by the programs; such a system is called a processing surface. The Cosmic Cube drops the locality requirement by providing a routing protocol between arbitrary nodes. The Cosmic Cube prototype proved so successful that the design was licensed and rapidly commercialized (1985). Several manufacturers are already offering "cube machines" which are routinely used by researchers at Caltech and elsewhere to solve computationally demanding scientific problems.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1988
- Accession Number
- ADA447726
Entities
People
- Alain J. Martin
Organizations
- California Institute of Technology