Agent-Oriented Approach to Simulate Exaflop Supercomputer with Application to Distributed Stochastic Simulation

A possibility of using an agent-oriented simulation system for solving a variety of problems that arise in design and implementation of exaflop supercomputers consisting of ten and hundred millions of computational nodes is discussed in the paper. We suggest two-lewel decentralized scheme of computations control and the corresponding simulation model in which all the computational nodes are distributed over computational domains controlled by their control agents. Master control agent distributes a flow of big problems over computational domains and manages common resources. Monte-Carlo method considered to be promising to use on exaflop supercomputers is given as an example of highly scalable algorithm. In this method, is essential that the lager sample size of independent realizations, the higher accuracy of estimating. We also suggest a parallel pseudorandom numbers generator suitable for large-scale computations with Monte Carlo method. When distributing stochastic computations over different nodes it is possible to simulate different sample volumes on different nodes using statistically optimal technique of results averaging. Naturally, an amount of computer resources available on each node must be quite enough to simulate the realizations effectively. The described algorithm of distributed stochastic simulation is asynchronous one and can be scaled to a practically infinite number of nodes using the described parallel pseudorandom numbers generator. An example of the highly scalable application utilizing distributed stochastic simulation on up-to-date teraflop supercomputers is the program library PARMONC. Also, the multi-agent simulation is used for the prediction and processing of possible failures of computational nodes. Architecture of dynamic system of failures prediction is given. The system consists of the agent for different purposes; each agent is playing its role to achieve the common goal.