Agentic systems, in which diverse agents cooperate to tackle challenging problems, are exploding in popularity in the AI community. However, existing agentic frameworks take a relatively narrow view of agents, apply a centralized model, and target conversational, cloud-native applications (e.g., LLM-based AI chatbots). In contrast, scientific applications require myriad agents be deployed and managed across diverse cyberinfrastructure. Here we introduce Academy, a modular and extensible middleware designed to deploy autonomous agents across the federated research ecosystem, including HPC systems, experimental facilities, and data repositories. To meet the demands of scientific computing, Academy supports asynchronous execution, heterogeneous resources, high-throughput data flows, and dynamic resource availability. It provides abstractions for expressing stateful agents, managing inter-agent coordination, and integrating computation with experimental control. We present microbenchmark results that demonstrate high performance and scalability in HPC environments. To explore the breadth of applications that can be supported by agentic workflow designs, we also present case studies in materials discovery, astronomy, decentralized learning, and information extraction in which agents are deployed across diverse HPC systems.

Data distribution across different facilities offers benefits such as enhanced resource utilization, increased resilience through replication, and improved performance by processing data near its source. However, managing such data is challenging due to heterogeneous access protocols, disparate authentication models, and the lack of a unified coordination framework. This paper presents DynoStore, a system that manages data across heterogeneous storage systems. At the core of DynoStore are data containers, an abstraction that provides standardized interfaces for seamless data management, irrespective of the underlying storage systems. Multiple data container connections create a cohesive wide-area storage network, ensuring resilience using erasure coding policies. Furthermore, a load-balancing algorithm ensures equitable and efficient utilization of storage resources. We evaluate DynoStore using benchmarks and realworld case studies, including the management of medical and satellite data across geographically distributed environments. Our results demonstrate a 10 % performance improvement compared to centralized cloud-hosted systems while maintaining competitive performance with state-of-the-art solutions such as Redis and IPFS. DynoStore also exhibits superior fault tolerance, withstanding more failures than traditional systems.

Failures in Task-based Parallel Programming (TBPP) can severely degrade performance and result in incomplete or incorrect outcomes. Existing failure-handling approaches, including reactive, proactive, and resilient methods such as retry and checkpointing mechanisms, often apply uniform retry mechanisms regardless of the root cause of failures, failing to account for the unique characteristics of TBPP frameworks such as heterogeneous resource availability and task-level failures. To address these limitations, we propose Wrath, a novel systematic approach that categorizes failures based on the unique layered structure of TBPP frameworks and defines specific responses to address failures at different layers. Wrath combines a distributed monitoring system and a resilient module to collaboratively address different types of failures in real time. The monitoring system captures execution and resource information, reports failures, and profiles tasks across different layers of TBPP frameworks. The resilient module then categorizes failures and responds with appropriate actions, such as hierarchically retrying failed tasks on suitable resources. Evaluations demonstrate that Wrath significantly improves TBPP robustness, tripling the task success rate and maintaining an application success rate of over 90 % for resolvable failures. Additionally, Wrath can reduce the time to failure by$20 \%-50 \%$, allowing tasks that are destined to fail to be identified and fail more quickly.

The exponential growth of data necessitates distributed storage models, such as peer-to-peer systems and data federations.While distributed storage can reduce costs and increase reliability, the heterogeneity in storage capacity, I/O performance, and failure rates of storage resources makes their efficient use a challenge.Further, node failures are common and can lead to data unavailability and even data loss.

Second-order optimization methods have been developed to enhance convergence and generalization in deep neural network (DNN) training compared to first-order methods like Stochastic Gradient Descent (SGD). However, these methods face challenges in distributed settings due to high communication overhead. Gradient compression, a technique commonly used to accelerate communication for first-order approaches, often results in low communication reduction ratios, decreased model accuracy, and/or high compression overhead when applied to second-order methods. To address these limitations, we introduce a novel gradient compression method for second-order optimizers called COMPSO. This method effectively reduces communication costs while preserving the advantages of second-order optimization. COMPSO employs stochastic rounding to maintain accuracy and filters out minor gradients to improve compression ratios. Additionally, we develop GPU optimizations to minimize compression overhead and performance modeling to ensure end-to-end performance gains across various systems. Evaluation of COMPSO on different DNN models shows that it achieves a compression ratio of 22.1×, reduces communication time by 14.2×, and improves overall performance by 1.9×, all without any drop in model accuracy.

Workflow and serverless frameworks have empowered new approaches to distributed application design by abstracting compute resources. However, their typically limited or one-size-fits-all support for advanced data flow patterns leaves optimization to the application programmer—optimization that becomes more difficult as data become larger. The transparent object proxy, which provides wide-area references that can resolve to data regardless of location, has been demonstrated as an effective low-level building block in such situations. Here we propose three high-level proxy-based programming patterns—distributed futures, streaming, and ownership—that make the power of the proxy pattern usable for more complex and dynamic distributed program structures. We motivate these patterns via careful review of application requirements and describe implementations of each pattern. We evaluate our implementations through a suite of benchmarks and by applying them in three meaningful scientific applications, in which we demonstrate substantial improvements in runtime, throughput, and memory usage.

Serverless computing has revolutionized application development and deployment by abstracting infrastructure management, allowing developers to focus on writing code. To do so, serverless platforms dynamically create execution environments, often using containers. The cost to create and deploy these environments is known as "cold start" latency, and this cost can be particularly detrimental to scientific computing workloads characterized by sporadic and dynamic demands. We investigate methods to mitigate cold start issues in scientific computing applications by pre-installing Python packages in container images. Using data from Globus Compute and Binder, we empirically analyze cold start behavior and evaluate four strategies for building containers, including fully pre-built environments and dynamic, on-demand installations. Our results show that pre-installing all packages reduces initial cold start time but requires significant storage. Conversely, dynamic installation offers lower storage requirements but incurs repetitive delays. Additionally, we implemented a simulator and assessed the impact of different warm times, finding that moderate warm times significantly reduce cold starts without the excessive overhead of maintaining always-hot states.

Task-based execution frameworks, such as parallel programming libraries, computational workflow systems, and function-as-a-service platforms, enable the composition of distinct tasks into a single, unified application designed to achieve a computational goal and abstract the parallel and distributed execution of those tasks on arbitrary hardware. Research into these task executors has accelerated as computational sciences increasingly need to take advantage of parallel compute and/or heterogeneous hardware. However, the lack of evaluation standards makes it challenging to compare and contrast novel systems against existing implementations. Here, we introduce TaPS, the Task Performance Suite, to support continued research in distributed task executor frameworks. TaPS provides (1) a unified, modular interface for writing and evaluating applications using arbitrary execution frameworks and data management systems and (2) an initial set of reference synthetic and real-world science applications. We discuss how the design of TaPS supports the reliable evaluation of frameworks and demonstrate TaPS through a survey of benchmarks using the provided reference applications.

Deep learning methods are transforming research, enabling new techniques, and ultimately leading to new discoveries. As the demand for more capable AI models continues to grow, we are now entering an era of Trillion Parameter Models (TPM), or models with more than a trillion parameters---such as Huawei's PanGu-Σ. We describe a vision for the ecosystem of TPM users and providers that caters to the specific needs of the scientific community. We then outline the significant technical challenges and open problems in system design for serving TPMs to enable scientific research and discovery. Specifically, we describe the requirements of a comprehensive software stack and interfaces to support the diverse and flexible requirements of researchers.

Advances in networks, accelerators, and cloud services encourage programmers to reconsider where to compute---such as when fast networks make it cost-effective to compute on remote accelerators despite added latency. Workflow and cloud-hosted serverless computing frameworks can manage multi-step computations spanning federated collections of cloud, high-performance computing (HPC), and edge systems, but passing data among computational steps via cloud storage can incur high costs. Here, we overcome this obstacle with a new programming paradigm that decouples control flow from data flow by extending the pass-by-reference model to distributed applications. We describe ProxyStore, a system that implements this paradigm by providing object proxies that act as wide-area object references with just-in-time resolution. This proxy model enables data producers to communicate data unilaterally, transparently, and efficiently to both local and remote consumers. We demonstrate the benefits of this model with synthetic benchmarks and real-world scientific applications, running across various computing platforms.

Scaling deep neural network training to more processors and larger batch sizes is key to reducing end-to-end training time; yet, maintaining comparable convergence and hardware utilization at larger scales is challenging. Increases in training scales have enabled natural gradient optimization methods as a reasonable alternative to stochastic gradient descent and variants thereof. Kronecker-factored Approximate Curvature (K-FAC), a natural gradient method, preconditions gradients with an efficient approximation of the Fisher Information Matrix to improve per-iteration progress when optimizing an objective function. Here we propose a scalable K-FAC algorithm and investigate K-FAC’s applicability in large-scale deep neural network training. Specifically, we explore layer-wise distribution strategies, inverse-free second-order gradient evaluation, and dynamic K-FAC update decoupling, with the goal of preserving convergence while minimizing training time. We evaluate the convergence and scaling properties of our K-FAC gradient preconditioner, for image classification, object detection, and language modeling applications. In all applications, our implementation converges to baseline performance targets in 9–25% less time than the standard first-order optimizers on GPU clusters across a variety of scales.

Kronecker-factored Approximate Curvature (K-FAC) has recently been shown to converge faster in deep neural network (DNN) training than stochastic gradient descent (SGD); however, K-FAC's larger memory footprint hinders its applicability to large models. We present KAISA, a K-FAC-enabled, Adaptable, Improved, and ScAlable second-order optimizer framework that adapts the memory footprint, communication, and computation given specific models and hardware to improve performance and increase scalability. We quantify the tradeoffs between memory and communication cost and evaluate KAISA on large models, including ResNet-50, Mask R-CNN, U-Net, and BERT, on up to 128 NVIDIA A100 GPUs. Compared to the original optimizers, KAISA converges 18.1--36.3% faster across applications with the same global batch size. Under a fixed memory budget, KAISA converges 32.5% and 41.6% faster in ResNet-50 and BERT-Large, respectively. KAISA can balance memory and communication to achieve scaling efficiency equal to or better than the baseline optimizers.

FanStore is a shared object store that enables efficient and scalable neural network training on supercomputers. By providing a global cache layer on node-local burst buffers using a compressed representation, it significantly enhances the processing capability of deep learning (DL) applications on existing hardware. In addition, FanStore allows POSIX-compliant file access to the compressed data in user space. We investigate the tradeoff between runtime overhead and data compression ratio using real-world datasets and applications, and propose a compressor selection algorithm to maximize storage capacity given performance constraints. We consider both asynchronous (i.e., with prefetching) and synchronous I/O strategies, and propose mechanisms for selecting compressors for both approaches. Using FanStore, the same storage hardware can host 2–13× more data for example applications without significant runtime overhead. Empirically, our experiments show that FanStore scales to 512 compute nodes with near linear performance scalability.

Training neural networks with many processors can reduce time-to-solution; however, it is challenging to maintain convergence and efficiency at large scales. The Kroneckerfactored Approximate Curvature (K-FAC) was recently proposed as an approximation of the Fisher Information Matrix that can be used in natural gradient optimizers. We investigate here a scalable K-FAC design and its applicability in convolutional neural network (CNN) training at scale. We study optimization techniques such as layer-wise distribution strategies, inverse-free second-order gradient evaluation, and dynamic K-FAC update decoupling to reduce training time while preserving convergence. We use residual neural networks (ResNet) applied to the CIFAR10 and ImageNet-1k datasets to evaluate the correctness and scalability of our K-FAC gradient preconditioner. With ResNet-50 on the ImageNet-1k dataset, our distributed K-FAC implementation converges to the 75.9% MLPerf baseline in 18-25% less time than does the classic stochastic gradient descent (SGD) optimizer across scales on a GPU cluster.