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.

This paper introduces Heimdall, a highly accurate and efficient machine learning-powered I/O admission policy for flash storage, designed to operate in a black-box manner. We make domain-specific innovations in various ML stages by introducing accurate period-based labeling, 3-stage noise filtering, in-depth feature engineering, and fine-grained tuning, which together improve the decision accuracy from 67% up to 93%. We perform various deployment optimizations to reach a sub-μs inference latency and a small, 28KB, memory overhead. With 500 unbiased random experiments derived from production traces, we show Heimdall delivers 15-35% lower average I/O latency compared to the state of the art and up to 2x faster to a baseline. Heimdall is ready for user-level, in-kernel, and distributed deployments.

Environmental stressors such as temperature, humidity, vibration, and radiation can severely impact the performance and reliability of SSDs, particularly in edge, automotive, aerospace, and datacenter deployments. Capturing sensor data in the field and conducting accelerated lab experiments are challenging, as they are time-consuming, resource-intensive, and often destructive to hardware. Specialized setups, such as thermal chambers or vibration rigs, are also required, which is why few studies explore this area, and current storage management techniques like RAID, tiering, and deduplication do not consider environmental factors. Models to capture these impacts would open new research opportunities across various fields. However, accurately modeling these effects remains challenging due to, (1) the limited availability of experimental data, (2) the complex, domino-like impact of historical exposure, (3) the interrelated nature of environmental factors, such as temperature and humidity, which exhibit correlation, (4) different response of each type of NAND flash memory TLC, MLC, and SLC to environmental factors, and (5) the difficulty that analytical and simple machine learning models face in generalizing across devices, environments, and unseen combinations of stressors. We believe that LLMs may offer a transformative alternative to this complex problem, with embedded domain knowledge and reasoning capabilities, to facilitate prompt-based natural language interaction. We propose a hybrid framework that combines Chain-of-Thought prompting and Retrieval-Augmented Generation to guide LLMs using physical principles and prior experiments. It enables interpretable "what-if" analysis of SSD behavior under environmental changes. Our results show that the LLM can effectively model the impact of temperature, humidity, and vibration on SSD performance, producing tail latency and bandwidth predictions with minimal error. The code and data are available on GitHub at https://github.com/Damrl-lab/SSD_LLM.

Accurate weather forecasting across time scales is critical for anticipating and mitigating the impacts of climate change. Recent data-driven methods based on deep learning have achieved significant success in the medium range, but struggle at longer subseasonal-to-seasonal (S2S) horizons due to error accumulation in their autoregressive approach. In this work, we propose OmniCast, a scalable and skillful probabilistic model that unifies weather forecasting across timescales. OmniCast consists of two components: a VAE model that encodes raw weather data into a continuous, lower-dimensional latent space, and a diffusion-based transformer model that generates a sequence of future latent tokens given the initial conditioning tokens. During training, we mask random future tokens and train the transformer to estimate their distribution given conditioning and visible tokens using a per-token diffusion head. During inference, the transformer generates the full sequence of future tokens by iteratively unmasking random subsets of tokens. This joint sampling across space and time mitigates compounding errors from autoregressive approaches. The low-dimensional latent space enables modeling long sequences of future latent states, allowing the transformer to learn weather dynamics beyond initial conditions. OmniCast performs competitively with leading probabilistic methods at the medium-range timescale while being 10× to 20× faster, and achieves state-of-the-art performance at the subseasonal-to-seasonal scale across accuracy, physics-based, and probabilistic metrics. Furthermore, we demonstrate that OmniCast can generate stable rollouts up to 100 years ahead. Code and model checkpoints are available at https://github.com/tung-nd/omnicast.

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.

Information theoretic quantities play a central role in machine learning. The recent surge in the complexity of data and models has increased the demand for accurate estimation of these quantities. However, as the dimension grows the estimation presents significant challenges, with existing methods struggling already in relatively low dimensions. To address this issue, in this work, we introduce REMEDI for efficient and accurate estimation of differential entropy, a fundamental information theoretic quantity. The approach combines the minimization of the cross-entropy for simple, adaptive base models and the estimation of their deviation, in terms of the relative entropy, from the data density. Our approach demonstrates improvement across a broad spectrum of estimation tasks, encompassing entropy estimation on both synthetic and natural data. Further, we extend important theoretical consistency results to a more generalized setting required by our approach. We illustrate how the framework can be naturally extended to information theoretic supervised learning models, with a specific focus on the Information Bottleneck approach. It is demonstrated that the method delivers better accuracy compared to the existing methods in Information Bottleneck. In addition, we explore a natural connection between REMEDI and generative modeling using rejection sampling and Langevin dynamics.

Weather forecasting is a fundamental problem for anticipating and mitigating the impacts of climate change. Recently, data-driven approaches for weather forecasting based on deep learning have shown great promise, achieving accuracies that are competitive with operational systems. However, those methods often employ complex, customized architectures without sufficient ablation analysis, making it difficult to understand what truly contributes to their success. Here we introduce Stormer, a simple transformer model that achieves state-of-the art performance on weather forecasting with minimal changes to the standard transformer backbone. We identify the key components of Stormer through careful empirical analyses, including weather-specific embedding, randomized dynamics forecast, and pressure-weighted loss. At the core of Stormer is a randomized forecasting objective that trains the model to forecast the weather dynamics over varying time intervals. During inference, this allows us to produce multiple forecasts for a target lead time and combine them to obtain better forecast accuracy. On WeatherBench 2, Stormer performs competitively at short to medium-range forecasts and outperforms current methods beyond 7 days, while requiring orders-of-magnitude less training data and compute. Additionally, we demonstrate Stormer’s favorable scaling properties, showing consistent improvements in forecast accuracy with increases in model size and training tokens. Code and checkpoints are available at https://github.com/tung-nd/stormer.

The information bottleneck framework provides a systematic approach to learning representations that compress nuisance information in the input and extract semantically meaningful information about predictions. However, the choice of a prior distribution that fixes the dimensionality across all the data can restrict the flexibility of this approach for learning robust representations. We present a novel sparsity-inducing spike-slab categorical prior that uses sparsity as a mechanism to provide the flexibility that allows each data point to learn its own dimension distribution. In addition, it provides a mechanism for learning a joint distribution of the latent variable and the sparsity, and hence it can account for the complete uncertainty in the latent space. Through a series of experiments using in-distribution and out-of-distribution learning scenarios on the MNIST, CIFAR-10, and ImageNet data, we show that the proposed approach improves accuracy and robustness compared to traditional fixed-dimensional priors, as well as other sparsity induction mechanisms for latent variable models proposed in the literature.

We present the Memristor-Spikelearn simulator (open-sourced), which is capable of incorporating detailed mem-ristor and circuit models in simulation to enable thorough study of synaptic plasticity in spiking neural networks under realistic device and circuit behaviors. Using this simulator, we demonstrate that: (1) a detailed device model is essential for simulating synaptic plasticity workloads, because results obtained using a simplified model can be misleading (e.g., it can overestimate test accuracy by up to 21.9%); (2) detailed simulation helps to determine the proper range of conductance values to represent weights, which is critical in order to achieve the desired accuracy -energy tradeoff (e.g., increasing the conductance values by$10\times$can increase accuracy from 70% to 83% at the price of$20\times$higher energy); and (3) detailed simulation also helps to determine an optimized circuit structure, which is another important design parameter that can yield different accuracy -energy tradeoffs.

Distributed data storage services tailored to specific applications have grown popular in the high-performance computing (HPC) community as a way to address I/O and storage challenges. These services offer a variety of specific interfaces, semantics, and data representations. They also expose many tuning parameters, making it difficult for their users to find the best configuration for a given workload and platform. To address this issue, we develop a novel variational-autoencoder-guided asynchronous Bayesian optimization method to tune HPC storage service parameters. Our approach uses transfer learning to leverage prior tuning results and use a dynamically updated surrogate model to explore the large parameter search space in a systematic way. We implement our approach within the DeepHyper open-source framework, and apply it to the autotuning of a high-energy physics workflow on Argonne's Theta supercomputer. We show that our transfer-learning approach enables a more than 40 x search speedup over random search, compared with a 2.5 x to 10 x speedup when not using transfer learning. Additionally, we show that our approach is on par with state-of-the-art autotuning frameworks in speed and outperforms them in resource utilization and parallelization capabilities.

I/O efficiency is crucial to productivity in scientific computing, but the growing complexity of HPC systems and applications complicates efforts to understand and optimize I/O behavior at scale. Data-driven machine learning-based I/O throughput models offer a solution: they can be used to identify bottlenecks, automate I/O tuning, or optimize job scheduling with minimal human intervention. Unfortunately, current state-of-the-art I/O models are not robust enough for production use and underperform after being deployed. We analyze four years of application, scheduler, and storage system logs on two leadership-class HPC platforms to understand why I/O models underperform in practice. We propose a taxonomy consisting of five categories of I/O modeling errors: poor application and system modeling, inadequate dataset coverage, I/O contention, and I/O noise. We develop litmus tests to quantify each category, allowing researchers to narrow down failure modes, enhance I/O throughput models, and improve future generations of HPC logging and analysis tools.

With the growing complexity of high-performance computing (HPC) systems, achieving high performance can be difficult because of I/O bottlenecks. We analyze multiple years' worth of Darshan logs from the Argonne Leadership Computing Facility's Theta supercomputer in order to understand causes of poor I/O throughput. We present Gauge: a data-driven diagnostic tool for exploring the latent space of supercomputing job features, understanding behaviors of clusters of jobs, and interpreting I/O bottlenecks. We find groups of jobs that at first sight are highly heterogeneous but share certain behaviors, and analyze these groups instead of individual jobs, allowing us to reduce the workload of domain experts and automate I/O performance analysis. We conduct a case study where a system owner using Gauge was able to arrive at several clusters that do not conform to conventional I/O behaviors, as well as find several potential improvements, both on the application level and the system level.

Training Convolutional Neural Network (CNN) is a computationally intensive task, requiring efficient parallelization to shorten the execution time. Considering the ever-increasing size of available training data, the parallelization of CNN training becomes more important. Data-parallelism, a popular parallelization strategy that distributes the input data among compute processes, requires the mini-batch size to be sufficiently large to achieve a high degree of parallelism. However, training with large batch size is known to produce a low convergence accuracy. In image restoration problems, for example, the batch size is typically tuned to a small value between 16 ~ 64, making it challenging to scale up the training. In this paper, we propose a parallel CNN training strategy that gradually increases the mini-batch size and learning rate at run-time. While improving the scalability, this strategy also maintains the accuracy close to that of the training with a fixed small batch size. We evaluate the performance of the proposed parallel CNN training algorithm with image regression and classification applications using various models and datasets.

Supervised learning is a promising approach for modeling the performance of applications running on large HPC systems. A key assumption in supervised learning is that the training and testing data are obtained under the same conditions. However, in production HPC systems these conditions might not hold because the conditions of the platform can change over time as a result of hardware degradation, hardware replacement, software upgrade, and configuration updates. These changes could alter the data distribution in a way that affects the accuracy of the predictive performance models and render them less useful; this phenomenon is referred to as concept drift. Ignoring concept drift can lead to suboptimal resource usage and decreased efficiency when those performance models are deployed for tuning and job scheduling in production systems. To address this issue, we propose a concept-drift-aware predictive modeling approach that comprises two components: (1) an online Bayesian changepoint detection method that can automatically identify the location of events that lead to concept drift in near-real time and (2) a moment-matching transformation inspired by transfer learning that converts the training data collected before the drift to be useful for retraining.

Storage system performance modeling is crucial for efficient use of heterogeneous shared resources on leadership-class computers. Variability in application performance, particularly variability arising from concurrent applications sharing I/O resources, is a major hurdle in the development of accurate performance models. We adopt a deep learning approach based on conditional variational auto encoders (CVAE) for I/O performance modeling, and use it to quantify performance variability. We illustrate our approach using the data collected on Edison, a production supercomputing system at the National Energy Research Scientific Computing Center (NERSC). The CVAE approach is investigated by comparing it to a previously proposed sensitivity-based Gaussian process (GP) model. We find that the CVAE model performs slightly better than the GP model in cases where training and testing data come from different applications, since CVAE can inherently leverage the whole data from multiple applications whereas GP partitions the data and builds separate models for each partition. Hence, the CVAE offers an alternative modeling approach that does not need pre-processing; it has enough flexibility to handle data from a wide variety of applications without changing the inference approach.

Storage resources in high-performance computing are shared across all user applications. Consequently, storage performance can vary markedly, depending not only on an application's workload but also on what other activity is concurrently running across the system. This variability in storage performance is directly reflected in overall execution time variability, thus confounding efforts to predict job performance for scheduling or capacity planning. I/O variability also complicates the seemingly straightforward process of performance measurement when evaluating application optimizations. In this work we present a methodology to measure I/O contention with more rigor than in prior work. We apply statistical techniques to gain insight from application-level statistics and storage-side logging. We examine different correlation metrics for relating system workload to job I/O performance and identify an effective and generally applicable metric for measuring job I/O performance. We further demonstrate that the system-wide monitoring granularity can directly affect the strength of correlation observed. Insufficient granularity and measurements can hide the correlations between application I/O performance and system-wide I/O activity.