Client-server systems are becoming increasingly common in the world today as users move to networks of distributed, interacting computers. This form of work demands new performance models as the interactions in client-server systems are more complex than the types supported by classic queueing network solvers such as Mean Value Analysis. A Layered Queueing Network is one of these models; it uses hierarchical decomposition and surrogate delays to solve the model.
This thesis describes a new analytic modeling tool called LQNS (Layered Queueing Network Solver) which extends previous techniques used to model distributed client-server systems. The contributions of the thesis are as follows. First, the model now supports forwarding. Forwarding is a technique where a reply to a client is deferred to a lower level server in a multi-level system, improving performance by reducing message traffic. Forwarding can also be used to convert open models to closed models. Second, systems that use early replies can be modeled. Early replies are used to reduce the response time by replying to a client before all of its work at a server is completed. Previous techniques have been extended to multiservers and to allow multiple clients. Third, activities have been introduced. Activities represent the smallest unit of modeling detail and can have arbitrary precedence relationships. Finally, the solver has been extended to handle models with both homogeneous and heterogeneous threads within a task. Homogeneous threads are used to model multiservers. Heterogeneous threads are used to model fork-join interactions such as asynchronous remote procedure calls and in in RAID storage devices. The solver also incorporates accuracy improvements for models with early replies and for models with multiple layers.
The solver has been used to analyze numerous systems found in existence today including a tele-operator system and a transaction processing system. Finally, a extensive performance model of the Linux 2.0 Network File System (NFS) is presented.