No abstract

Replication is a technique that maintains and allows access to copies of data and ser vices on multiple servers. It is usually used in multi-tier architecture systems, as found in distributed software systems, to increase performance, reliability, and availability.The growing complexity of modern software systems, especially very large distributed systems, increases the difficulty of achieving performance objectives. A designer needs a high-level analytic tool to consistently compare the performance patterns of different replication techniques for a very large multi-tier replicated system. Analytic tools enable the designer to rapidly compute the relevant performance measures at an early stage of the design. This enables the designer to gain a competitive advantage by delivering his product in time, and at a lower cost.The Layered Queueing Network (LQN) is the formalism selected in this thesis, as it is designed for performance modeling of software and hardware systems with a multi-tier (layered) architecture. The Layered Queueing Network Solver (LQNS) is an analytic tool used for performance analysis. This thesis describes new analytic algorithms implemented in the LQNS to compute relevant performance measures for parallel and distributed sys tems with replication. The contributions of the thesis are as follows. First, replication performance patterns are identified and classified. Second, a general quorum pattern is in corporated in the LQN formalism to model and compute the performance measures of a replication model with a quorum consensus protocol. Third, the accuracy of the LQNS is improved by introducing new closed-form formulas for the thread delay distribution. The iii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.ABSTRACT iv distribution is used to compute the service time of a client when the client has a quorum pattern. Finally, algorithms that exploit the symmetry in a system with replicated com ponents are designed and implemented to hasten model solution time. Those replication components can contain internal parallelism.The results of this research are used to study the performance of an Air Traffic Control system and an Industrial Information Management system. The LQNS solutions for these large systems are rapid, and the accuracy of the solutions is adequate for most purposes during the early stage of system modeling.Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. DedicationI dedicate this thesis to my parents. VReproduced with permission of the copyright owner. Further reproduction prohibited without permission.