PERFORMABILITY OF INTEGRATED SOFTWARE-HARDWARE COMPONENTS OF REAL-TIME PARALLEL AND DISTRIBUTED SYSTEMS

This paper presents a method for performability analysis of the integrated software-hardware components of parallel & distributed systems. The contributions of this paper are: Present a general performability evaluation technique using the Generalized Stochastic Petri Nets at the system level (analysis of integrated software-hardware systems) Address the intractable problem of evaluating a parallel software environment consisting of interacting fault-tolerant parallel tasks. This is accomplished using a decomposition technique at the task-graph level, where the task graph is decomposed into segments. The modeling of failure & repair events of interacting parallel modules in a parallel processing system is difficult because failure of a module can force other interacting modules to rollback, which increases the overall response time of the computation. This paper effectively models the recovery blocks in the interacting parallel modules as well as their supporting hardware. This method greatly facilitates the analysis of performability at the system level. However, an integrated performability model of the softwarehardware system increases the size, perhaps prohibitively, of the Markov generator matrix. This issue is addressed, and a performability decomposition technique at the task-graph level is presented using the properties of Laguerre function coefficients. The details of the methodology are illustrated by a simple example of a radar command center where the interacting parallel modules run on a shared-memory multiprocessing system.