Parallel Computing And Performance Analysis Of Concurrent Systems

With the wide use of concurrent systems in various fields, performance analysis to ensure the quality of the systems is becoming a hot issue for the developers and users. In the early stage of the software engineering, if we can model the system and conduct the performance analysis, then the design error can be founded earlier which could decrease the cost to modify and test the software.A few formal methods of modeling and performance analysis have been proposed, the follows are the typical ones: Queuing Theory, Stochastic Petri net (SPN) and Stochastic Process Algebra (SPA). These methods are based on the Markov processes, thus state-space explosion may occur. Besides the method based on Queuing Theory can not describe the phenomenon of synchronization and resource sharing. Stochastic Petri Net (SPN) is popular because of its strong ability to model and analyze concurrent systems. Although many techniques have been proposed to reduce the state space, the state-space explosion is still there.To address the issue, this paper proposes a method for the modeling and performance analysis of large-scale concurrent systems by parallel computing ordinary differential equations. The follow list is our work:(1) Modeling. In this paper, Petri Net is used to describe concurrent system, after it has been relaxed, ordinary differential equations is constructed to model the system. In this way, state-space explosion can be avoided and system performance is analyzed quantitatively by solving the equations using MATLAB.(2) Parallel computing. If the system is very large, then the number of states will be large, and it becomes difficult to solve using MATLAB. Considered the special structures of the equations, transformation rules from Petri nets to hypergraph are given. Then hpergraph is partitioned, based on the partition result, equations is grouped and solved in parallel with SUNDIALS. Based on the Load-balance optimizing and communication minimization, the bottleneck problems of parallel computation can be solved.(3) Performance analysis. The method to analyze and optimize performance is given according to the solutions of ordinary differential equations. Our method compared with the method based on SPN.Finally, we use typical examples to analyze the complexity of parallel algorithm and the performance of system. It demonstrates the advantage of our method.