Doctor of Philosophy

dissertationThe increase in computational power of supercomputers is enabling complex scientific phenomena to be simulated at ever-increasing resolution and fidelity. With these simulations routinely producing large volumes of data, performing efficient I/O at this scale has become a very difficult task. Large-scale parallel writes are challenging due to the complex interdependencies between I/O middleware and hardware. Analytic-appropriate reads are traditionally hindered by bottlenecks in I/O access. Moreover, the two components of I/O, data generation from simulations (writes) and data exploration for analysis and visualization (reads), have substantially different data access requirements. Parallel writes, performed on supercomputers, often deploy aggregation strategies to permit large-sized contiguous access. Analysis and visualization tasks, usually performed on computationally modest resources, require fast access to localized subsets or multiresolution representations of the data. This dissertation tackles the problem of parallel I/O while bridging the gap between large-scale writes and analytics-appropriate reads. The focus of this work is to develop an end-to-end adaptive-resolution data movement framework that provides efficient I/O, while supporting the full spectrum of modern HPC hardware. This is achieved by developing technology for highly scalable and tunable parallel I/O, applicable to both traditional parallel data formats and multiresolution data formats, which are directly appropriate for analysis and visualization. To demonstrate the efficacy of the approach, a novel library (PIDX) is developed that is highly tunable and capable of adaptive-resolution parallel I/O to a multiresolution data format. Adaptive resolution storage and I/O, which allows subsets of a simulation to be accessed at varying spatial resolutions, can yield significant improvements to both the storage performance and I/O time. The library provides a set of parameters that controls the storage format and the nature of data aggregation across he network; further, a machine learning-based model is constructed that tunes these parameters for the maximum throughput. This work is empirically demonstrated by showing parallel I/O scaling up to 768K cores within a framework flexible enough to handle adaptive resolution I/O