Modeling of an availability driven computer network architecture
Dan Cristian Marinescu, VINICIUS FERREIRA BRITTO REGO, Wojciech Szpankowski · Purdue e-Pubs (Purdue University System) · 1987
An availiability driven high speed network architecture has been proposed in the literature.The architecrure incorporates a long'"range view of LANs that service multitudes (possibly lhollsands) of relatively slow hosts (workstations, slower minis) and a smaller number of exuemely fast hosts (supercomputers, parallel processors) and specialized service hoses (Image Processors, Lisp Machines, Data Bases Machines).In oder to guarantee simple, efficent operation and small response times (compared to scaling up existing schemes), me availibility ddven concept makes use of connectionless communication OD a broadcast bus along with the notion of scheduling protocols.The scheduling protocols are essentially multiple access algorithms that are random, limited contention (hybrids), or contention-free multi-access.The many advantages of this design include an inherent dynamic load balanced system (without the need for distributed balancing algorithms), fairness, robustness, stability, graceful perfonnance degradation with upward scaling, etc.In this report is presented the results of a set of intensive simulation experiments on a four sever, ten user ADMA system.The protocols investigated are the token-passing, coUision resolution, and two versions (fixed backoff, binary exponential backoff) of the CSivWCD datalink protocols.In this pilot study, the simulation experiments tend [Q confinn certain ADMA characteristics, while raising a host of interesting p~ffonnance questions regarding the behaviour of Utis novel but nonrivial architecture.