Communication Overhead on the Intel iPSC-860 Hypercube
Abstract
A key measure of the power of a multiple computer system is the interprocessor communication overhead. Despite major improvements in design and technology, the time to communicate a datum from one processor to another remains one or more orders of magnitude greater than the time to access it on the same processor. This time is influenced by the type of interconnection network used as well as the strategy used for routing messages. This report presents the results of experiments carried out on the recently introduced Intel iPSC-860 hypercube. This machine is based on the powerful i860 microprocessor and uses circuit-switched communications through a hypercube interconnection. The results of these experiments permit the following major conclusions to be made about this machine: (1) Contrary to popular belief, the time required to communicate between nodes does depend on the number of intervening hops on both machines. Although this variation can be neglected for very large messages, it cannot be ignored for short messages; (2) Edge contention (the sharing of a communication link by two or more paths) leads to severe overhead for all message sizes. This can increase the time to communicate by a factor of more than seven; and (3) Node contention (the sharing of a node by two or more paths) has no measurable impact.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1990
- Accession Number
- ADA227204
Entities
People
- Shahid H. Bokhari