Autonomic orchestration of in-situ and in-transit data analytics for simulation studies

Modern parallel/distributed simulations can produce large amounts of data. The historical approach of performing analyses at the end of the simulation is unlikely to cope with modern, extremely large-scale analytics jobs. Indeed, the I/O subsystem can quickly become the global bottleneck. Similarly, processing on-the-fly the data produced by simulations can significantly impair the performance in terms of computational capacity and network load. We present a methodology and reference architecture for constructing an autonomic control system to determine at runtime the best placement for data processing (on simulation nodes or a set of external nodes). This allows for a good tradeoff between the load on the simulation's critical path and the data communication system. Our preliminary experimentation shows that autonomic orchestration is crucial to improve the global performance of a data analysis system, especially when the simulation node's rate of data production varies during simulation.