Conference Agenda

In high-performance computing (HPC), in situ analysis and visu- alization avoid costly I/O by extracting insight during the simula- tion run. When these tasks execute on the same GPUs that drive the simulation, resource contention can degrade performance. We propose an asynchronous framework that offloads visualization to idle CPU cores and overlaps in situ work with simulation execution. The framework uses ASCENT for efficient data movement and ren- dering and applies core pinning. Evaluated with NekRS on Polaris and JUWELS Booster, it reduced end-to-end runtime by 21–40% for slice-based visualizations compared to inline GPU instrumenta- tion. Slice outputs incurred little overhead and overlapped cleanly with the simulation. Heavier filters like isovolumes, multilevel con- tours, and volume rendering were “free” (i.e., visualization time was encapsulated by simulation time) only at lower node counts, consistent with a CPU-budget model: filter cost and output cadence must fit the cores available per node.

Data visualization and analysis are increasingly becoming a bottleneck in exascale simulations because of the I/O bottleneck both with respect to speed and memory. In-situ visualization and analysis provides a viable solution for analyzing extreme scale simulations by processing data in memory, instead of dumping simulation snapshots on disk. It can also be used for debugging and ease the process of developing applications. In short, with in-situ tools there is a tighter integration of simulation and data processing as opposed to the conventional post-processing techniques.

Parallel File Systems (PFS) on supercomputers, which are shared resources, can be a massive bottleneck when running I/O-intensive applications at large scale. If multiple optimizations are possible (optimized I/O libraries, burst-buffers & pre-fetching, I/O-aware scheduling, etc.), they first require having a good understanding of how the applications read and write data on the PFS.
To this end, in this project talk, we present our methodology implemented through MOSAIC, a Python library, to detect the I/O behavior of HPC applications. MOSAIC captures patterns through three key facets: temporality (when do the applications perform I/O?), periodicity (can we find repeated I/O operations?), and metadata impact (do the applications perform many metadata requests?). Those categories cover multiple aspects of I/O loads and enable an extensive coarse-grain description of the I/O activity.
We analyzed Darshan traces from the Blue Waters and Polaris supercomputers to extract the I/O patterns and understand the I/O behavior of the workloads. In particular, we are interested in finding pattern correlations and variations in order to identify I/O patterns that usually occur together, and better understand how the execution of the same application can vary over time. These analyses also make it possible to compare the I/O behavior between two leading supercomputers and thus determine whether certain patterns are machine-dependent or common.

Summary of the work on scheduling for resources that vary over time (e.g., various levels of green versus brown power in HPC centers) conducted in cooperation with Argonne/U. Chicago and Inria.