Scientific deep learning (DL) at scale typically trains on terabyte-scale datasets across thousands of accelerators, placing immense pressure on storage systems to keep pace with computation. Existing solutions respond to this demand by tuning individual I/O parameters to accelerate training performance. However, these techniques are limited by costly experiments, configuration space explosion, and inability to generalize application-specific optimizations. This leads to applications running with suboptimal configurations that reduce training efficiency, system utilization, or both. To address the challenge of finding the optimal configuration efficiently, we developed GLANCED-IO, a cross-layer I/O optimization framework that optimizes DL pipelines with high-fidelity approximation and efficient configuration space exploration. Through this work, we identified the following three key findings. First, independently optimizing either the application or system configurations leaves up to 2.4 × performance on the table for scientists to efficiently run DL pipelines on HPC systems. Second, GLANCED-IO’s one-factor-at-a-time (OFAT)-guided greedy exploration strategy achieved results comparable to more-expensive autotuning techniques while removing the pre-training required by ML-based approaches. Third, GLANCED-IO avoids executing the full application during optimization by operating on representative data subsets without GPUs, yet preserves 93% performance fidelity on average when deployed in DL pipelines. We demonstrate the efficacy of GLANCED-IO by optimizing large-scale global weather forecasting DL workloads, achieving up to 1.57 × better performance than state-of-the-art with 2.3 × fewer configuration evaluations than AIIO and 3.3 × faster optimization than DeepHyper.

Generative machine learning offers new opportunities to better understand complex Earth system dynamics. Recent diffusion-based methods address spectral biases and improve ensemble calibration in weather forecasting compared to deterministic methods, yet have so far proven difficult to scale stably at high resolutions. We introduce AERIS, a 1.3 to 80B parameter pixel-level Swin diffusion transformer to address this gap, and SWiPe, a generalizable technique that composes window parallelism with sequence and pipeline parallelism to shard window-based transformers without added communication cost or increased global batch size. On Aurora (10,080 nodes), AERIS sustains 10.21 ExaFLOPS (mixed precision) and a peak performance of 11.21 ExaFLOPS with 1 × 1 patch size on the 0.25° ERA5 dataset, achieving 95.5% weak scaling efficiency, and 81.6% strong scaling efficiency. AERIS outperforms the IFS ENS and remains stable on seasonal scales to 90 days, highlighting the potential of billion-parameter diffusion models for weather and climate prediction.

Modern HPC workflows involve intricate coupling of simulation, data analytics, and artificial intelligence (AI) applications to improve time to scientific insight. These workflows require a cohesive set of performance analysis tools to provide a comprehensive understanding of data exchange patterns in HPC systems. However, current tools are not designed to work with an AI-based I/O software stack that requires tracing at multiple levels of the application. To this end, we developed a data flow tracer called DFTracer to capture data-centric events from workflows and the I/O stack to build a detailed understanding of the data exchange within AI-driven workflows. DFTracer has the following three novel features, including a unified interface to capture trace data from different layers in the software stack, a trace format that is analysis-friendly and optimized to support efficiently loading multi-million events in a few seconds, and the capability to tag events with workflow-specific context to perform domain-centric data flow analysis for workflows. Additionally, we demonstrate that DFTracer has a $1.44 x$ smaller runtime overhead and 1.3-7.1x smaller trace size than state-of-the-art tracing tools such as Score-P, Recorder, and Darshan. Moreover, with AI-driven workflows, Score-P, Recorder, and Darshan cannot find I/O accesses from dynamically spawned processes, and their load performance of 100 M events is three orders of magnitude slower than DFTracer. In conclusion, we demonstrate that DFTracer can capture multi-level performance data, including contextual event tagging with a low overhead of 1-5% from AI-driven workflows such as MuMMI and Microsoft’s Megatron Deepspeed running on large-scale HPC systems.

We present a scalable, end-to-end workflow for protein design. By augmenting protein sequences with natural language descriptions of their biochemical properties, we train generative models that can be preferentially aligned with protein fitness landscapes. Through complex experimental-and simulation-based observations, we integrate these measures as preferred parameters for generating new protein variants and demonstrate our workflow on five diverse supercomputers. We achieve >1 ExaFLOPS sustained performance in mixed precision on each supercomputer and a maximum sustained performance of 4.11 Ex-aFLOPS and peak performance of 5.57 ExaFLOPS. We establish the scientific performance of our model on two tasks: (1) across a predetermined benchmark dataset of deep mutational scanning experiments to optimize the fitness-determining mutations in the yeast protein HIS7, and (2) in optimizing the design of the enzyme malate dehydrogenase to achieve lower activation barriers (and therefore increased catalytic rates) using simulation data. Our implementation thus sets high watermarks for multimodal protein design workflows.

Summary Parallel I/O is a critical technique for moving data between compute and storage subsystems of supercomputers. With massive amounts of data produced or consumed by compute nodes, high‐performant parallel I/O is essential. I/O benchmarks play an important role in this process; however, there is a scarcity of I/O benchmarks representative of current workloads on HPC systems. Toward creating representative I/O kernels from real‐world applications, we have created h5bench , a set of I/O kernels that exercise hierarchical data format version 5 (HDF5) I/O on parallel file systems in numerous dimensions. Our focus on HDF5 is due to the parallel I/O library's heavy usage in various scientific applications running on supercomputing systems. The various tests benchmarked in the h5bench suite include I/O operations (read and write), data locality (arrays of basic data types and arrays of structures), array dimensionality (one‐dimensional arrays, two‐dimensional meshes, three‐dimensional cubes), I/O modes (synchronous and asynchronous). In this paper, we present the observed performance of h5bench executed along several of these dimensions on existing supercomputers (Cori and Summit) and pre‐exascale platforms (Perlmutter, Theta, and Polaris). h5bench measurements can be used to identify performance bottlenecks and their root causes and evaluate I/O optimizations. As the I/O patterns of h5bench are diverse and capture the I/O behaviors of various HPC applications, this study will be helpful to the broader supercomputing and I/O community.

Modern scientific workflows couple simulations with AI-powered analytics by frequently exchanging data to accelerate time-to-science to reduce the complexity of the simulation planes. However, this data exchange is limited in performance and portability due to a lack of support for scientific data formats in AI frameworks. We need a cohesive mechanism to effectively integrate at scale complex scientific data formats such as HDF5, PnetCDF, ADIOS2, GNCF, and Silo into popular AI frameworks such as TensorFlow, PyTorch, and Caffe. To this end, we designed Stimulus, a data management library for ingesting scientific data effectively into the popular AI frameworks. We utilize the StimOps functions along with StimPack abstraction to enable the integration of scientific data formats with any AI framework. The evaluations show that Stimulus outperforms several large-scale applications with different use-cases such as Cosmic Tagger (consuming HDF5 dataset in PyTorch), Distributed FFN (consuming HDF5 dataset in TensorFlow), and CosmoFlow (converting HDF5 into TFRecord and then consuming that in TensorFlow) by 5.3 x, 2.9 x, and 1.9 x respectively with ideal I/O scalability up to 768 GPUs on the Summit supercomputer. Through Stimulus, we can portably extend existing popular AI frameworks to cohesively support any complex scientific data format and efficiently scale the applications on large-scale supercomputers.

Modern-era high performance computing (HPC) systems are providing multiple levels of memory and storage layers to bridge the performance gap between fast memory and slow disk-based storage system managed by Lustre or GPFS. Several of the recent HPC systems are equipped with SSD and NVMe-based storage that is attached locally to compute nodes. A few systems are providing an SSD-based “burst buffer” intermediate storage layer that is accessible by all compute nodes as a single file system. Although these hardware layers are intended to reduce the latency gap between memory and disk-based long-term storage, how to utilize them has been left to the users. High-level I/O libraries, such as HDF5 and netCDF, can potentially take advantage of the node-local storage as a cache for reducing I/O latency from capacity storage. However, it is challenging to use node-local storage in parallel I/O especially for a single shared file. In this paper, we present an approach to integrate node-local storage as transparent caching or staging layers in a high-level parallel I/O library without placing the burden of managing these layers on users. We designed this to move data asynchronously between the caching storage layer and a parallel file system to overlap the data movement overhead in performing I/O with compute phases. We implement this approach as an external HDF5 Virtual Object Layer (VOL) connector, named Cache VOL. HDF5 VOL is a layer of abstraction in HDF5 that allows intercepting the public HDF5 application programming interface (API) and performing various optimizations to data movement after the interception. Existing HDF5 applications can use Cache VOL with minimal code modifications. We evaluated the performance of Cache VOL in HPC applications such as VPIC-10, and deep learning applications such as ImageNet and CosmoFlow. We show that using Cache VOL, one can achieve higher observed I/O performance, more scalable and stable I/O compared to direct I/O to the parallel file system, thus achieving faster time-to-solution in scientific simulations. While the caching approach is implemented in HDF5, the methods are applicable in other high-level I/O libraries.

Deep learning has been shown as a successful method for various tasks, and its popularity results in numerous open-source deep learning software tools. Deep learning has been applied to a broad spectrum of scientific domains such as cosmology, particle physics, computer vision, fusion, and astrophysics. Scientists have performed a great deal of work to optimize the computational performance of deep learning frameworks. However, the same cannot be said for I/O performance. As deep learning algorithms rely on big-data volume and variety to effectively train neural networks accurately, I/O is a significant bottleneck on large-scale distributed deep learning training. This study aims to provide a detailed investigation of the I/O behavior of various scientific deep learning workloads running on the Theta supercomputer at Argonne Leadership Computing Facility. In this paper, we present DLIO, a novel representative benchmark suite built based on the I/O profiling of the selected workloads. DLIO can be utilized to accurately emulate the I/O behavior of modern scientific deep learning applications. Using DLIO, application developers and system software solution architects can identify potential I/O bottlenecks in their applications and guide optimizations to boost the I/O performance leading to lower training times by up to 6.7x.