Scientific simulations in various domains such as chemistry and material science are being routinely executed on extreme scale supercomputing systems. This increased computational capability helps in improved modeling of the real-world phenomena by enabling higher fidelity and more complex simulations. These simulations produce enormous output data which is stored and analyzed to understand the simulated phenomena. Analysis enables scientists to derive insights from their simulations. However, slower growth in storage bandwidth in comparison to the computational capability is a critical bottleneck in the simulation-analysis pipeline. Traditional post-processing of stored simulation output leads to increased analysis time due to poor I/O bandwidth and large data requirements. This also increases the time to gain insight from the simulations. Hence, it is essential to perform simulation-time analysis.<br/><br/>Simulation-time analysis can be performed during the simulation on the same compute resources/cluster in two modes - time-shared or space-shared. In time-shared mode, simulation and analysis alternate whereas in space-shared mode, simulation and analysis simultaneously run on different processors. Simulation-time analysis can also be performed at a different execution site concurrently with the simulation. In situ analysis minimally increases the simulation time due to additional analysis computations or due to the required data transfer for analysis. However, this significantly reduces the end-to-end simulation-analysis time. Such analysis ought to be scheduled at an optimal frequency so that the total analysis time is within a user-specified time threshold. However, it is challenging to decide the optimal analysis frequency and is often chosen empirically. We present the scheduling of simulation-time analysis as a numerical optimization problem to maximize the number of online analysis subject to resource constraints such as I/O and network bandwidth, rate of computation, available memory, etc. Our optimization-based scheduling recommends the different feasible analyses and their frequencies depending on the threshold for analysis time. It also takes into account the importance of the various analyses that a user may specify. We demonstrate the effectiveness of our approach for the classical molecular dynamics simulation code - LAMMPS on the IBM Blue Gene/Q and Cray XC40 systems. Our optimal scheduling results in above 80% usage of the allowed threshold.

Scalable<i> in situ</i> analysis for large-scale molecular dynamics simulations on supercomputers

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.

Scalable in situ analysis for large-scale molecular dynamics simulations on supercomputers