Performance-related reliability management for fault-tolerant multistage interconnection system

Abstract

Multistage interconnection system (MIS) which is a collection of switching elements (SEs) and links connecting the SEs has been widely used as a promising fabric to communicate between processors and/or memories in multiprocessing (parallel computing) and distributed systems. In the MIS, some failures of SEs may cause disconnection of data paths between several input/output pairs, so the system performance should be in degradation mode before the faulty SEs are repaired. This brings about fault-tolerant MIS for settling fault-sensitive system and enhancing the fault-tolerance of the system. This thesis deals with performance-related reliability (PRR) management focused on MIS. The first issue in the MIS is how to measure the reliability of MIS more accurately. The more advanced a system, the more important its performance not only its reliability. Therefore we propose PRR measures to characterize the degrading features of fault-tolerant MISs in the presence of faulty SEs. The measures include: mean number of connected input/output pairs, mean number of survivable inputs that are connected to at least one output and finally mean number of outputs connected from a survivable input. We also evaluate the measures for the well-known unique-path shuffle exchange network (SEN) and its two fault-tolerant variants, extra-stage shuffle exchange network (SEN+) and augmented-SEN (ASEN), and then compare the networks through some numerical tests. The second issue is how to enhance the fault-tolerance of a certain system and to extend the methods. By creating redundant paths we make the system tolerate the faults and enhance its reliability. In chapter 3, we introduce four useful types of system which represent actual topologies of real world communication networks or parallel computing systems. Similar procedures used before are passed through when measuring PRR and comparing its results. Of course it is difficult to describe actual system structures or network classes as c...