Unified Component API: Supporting Standardized Middleware Integration and Flexible Interoperability in Cyberinfrastructure

Unified Component API: Supporting Standardized Middleware Integration and Flexible Interoperability in Cyberinfrastructure 1st Nick Cliffel The Ohio State University Columbus, Ohio, USA [email protected] 2nd Swathi Vallabhajosyula The Ohio State University Columbus, Ohio, USA [email protected] 3rd Jingxin Wang Rutgers university Piscataway, NJ, USA [email protected] 4th Zhao Zhang Rutgers university) Piscataway, NJ, USA [email protected] 5th Rajiv Ramnath The Ohio State University Columbus, Ohio, USA [email protected] Abstract—Modern cyberinfrastructure supports an increasingly complex ecosystem of scientific workflows that span on-field sensors, edge computing devices, high-performance computing (HPC) centers, and hybrid cloud environments. Within the NSF ICICLE AI Institute, we are developing tools across project thrusts such as AI for Cyberinfrastructure (AI4CI) and Cyberinfrastructure for AI (CI4AI), guided by user-driven domains like digital agriculture and animal ecology to enable an edgecentric continuum. However, as these modules are developed by independent research groups in collaboration with domain scientists, they often lack interoperability within and across thrusts. To address these challenges, we propose a unified reference API framework as an initial step toward integrating these components into a cohesive system, starting with middleware tools developed within the AI4CI thrust. We present the design and implementation of the Unified Component API (UCA), which adopts a narrow waist architecture to unify Intelligent Middleware Technologies (IMT) components with standardized lifecycle management methods (Start,Stop, Terminate, and Restart), along with service invocations such as InferenceService using HARPRuntimeInference and GPUDistributedInference. This design enables plug-andplay integration tailored to specific execution scenarios. We demonstrate these capabilities through the Smart Scheduler, a modular component for AI-enhanced HPC job provisioning that orchestrates auxiliary tools such as workload profilers, data generators, and estimators. Our results show how UCA simplifies user interactions and supports rapid development and testing of alternative implementations, enabling extensible and scalable cyberinfrastructure systems across diverse scientific domains. I. INTRODUCTION Modern cyberinfrastructure is no longer limited to an integrated set of advanced computing and storage systems, This work was supported by the National Science Foundation’s - AI Institute for Intelligent Cyberinfrastructure with Computational Learning in the Environment(ICICLE) and SGX3 A Center of Excellence to Extend Access, Expand the Community, and Exemplify Good Practices for CI Through Science Gateways under grant agreements OAC-2112606, and OAC2231406. visualization tools, and high-speed networks combined with software to support scientific discovery. Instead, it extends from the edge to the center, comprising heterogeneous systems with diverse compute and memory resources distributed across this emerging ecosystem, while also addressing energy and power constraints, as illustrated in Figure 1. In this context, ICICLE (Intelligent Cyberinfrastructure with Computational Learning in the Environment) [1], an NSF AI Institute, is pioneering the development of intelligent, distributed cyberinfrastructure that integrates edge devices with high-performance computing systems. ICICLE focuses on advancing AI-driven tools and workflows to support scientific and societal applications. The goals of this infrastructure are broad and include: 1) accommodating modern AI workflows that begin with edge sensors for gathering data, continue with storing and processing this data at centralized facilities, building models, and deploying these models back to edge or near-edge devices such as laptops connected to sensor networks; 2) making cyberinfrastructure more accessible and interactive for non-computing specialists, including domain experts such as ecologists and farmers; and 3) optimizing the edge-to-center continuum for varied objectives such as code optimization for faster processing at data centers, parallelization for efficient training, energy-aware model selection for edge inference with larger models, and providing user interface gateways that offer insights into both the gathered data and generated models to help users understand cost, energy, and other resource constraints. The ICICLE project is organized around multiple research thrusts, each addressing broader objectives at the intersection of artificial intelligence (AI) and cyberinfrastructure (CI). Ini- tially, the project focused on foundational research questions such as “What foundational AI models can democratize AI for diverse domains?”,“What cyberinfrastructure optimizations are required to accommodate modern AI workflows (CI4AI)?”, “What AI-centric models could be developed to improve CI and resource optimization (AI4CI)?”, and “What use cases can guide the design of AI-driven CI, such as solving scientific problems in domains like digital agriculture and animal ecology?”. Over time, these research efforts have resulted in the development of a range of tools and frameworks, including smart schedulers, compiler optimizers, resource estimators, cyberinfrastructure knowledge graphs, and other software components. The current software stack is illustrated in the reference architecture shown in Figure 2. Over time, these research efforts have resulted in the development of a range of tools and frameworks, including smart schedulers [3], compiler optimizations [5], resource estimators [6], cyberinfrastructure knowledge graphs [4], and other software components. The current software stack is illustrated in the reference architecture shown in Figure 2. A. Problem Since most of the code was developed in close collaboration with domain scientists, different AI4CI and CI4AI teams partnered with domain science thrusts to create research software stacks. However, these tools were built by independent groups using varied programming languages and software architectures, resulting in multiple management and communication interfaces. This fragmentation has led to high learning curves for new users and a lack of uniform interfaces or standard mechanisms for accessing the services provided by these tools. Fig. 1. Modern cyberinfrastructure architecture spanning on-field sensors, edge devices, hybrid cloud, and HPC systems to support scientific computing, data science, and AI workflows [1]. B. Solution To address these challenges, the Unified Component API (UCA) project was proposed. The UCA framework aims to unify the management of Intelligent Middleware Technologies (IMT) components by standardizing interfaces for starting, stopping, and communicating with IMT components. As part of this effort, a Python SDK was developed to serve as a gateway, enabling both domain scientists and other ICICLE thrusts to interact with middleware tools in a structured and simplified manner without needing to deploy or learn each individual component’s API. In this paper, we describe our approach for synergizing these diverse components through the development of a basic framework with sample modules. Our goal is to achieve scalability and support plug-and-play capabilities as key design objectives for future integration efforts. II. DESIGN AND DEVELOPMENT OF THE UNIFIED COMPONENT API: FRAMEWORK ARCHITECTURE AND IMPLEMENTATION The development of the Unified Component API (UCA) followed a structured, multi-stage process to unify and streamline the management of Intelligent Middleware Technologies (IMT) tools. The methodology consisted of five stages: (1) identifying relevant IMT components, (2) gathering background information and defining use cases via one-to-one interviews, (3) designing a standardized interface for component lifecycle management with Start,Stop,Terminate, and Restart functionalities for administration and datacentric communication, enabling plug-and-play between different models designed for the same research question (for example, “estimate resources for AI training job X” could involve black-box execution time monitoring estimators or white-box cost-based estimators that can seamlessly swap based on their performance or user requirements within the framework), (4) creating class diagrams for individual components, and (5) implementation of the unified framework. In Stage 1, we reviewed all available IMT tools to identify key components for inclusion in the unified framework. This resulted in the selection of five primary components: Smart Scheduler, Smart Compiler, RefDNN with the DNN Training Estimator, Cyberinfrastructure Knowledge Network (CKN), and Model Commons. In Stage 2, we conducted interviews with the developers of each selected component and created a Google Form to gather detailed information about their functionality, use cases, and technical requirements. This process helped us document the general input and output expectations for each tool. For instance, the Smart Scheduler utilizes AI-driven predictions to enhance HPC job provisioning, while the Smart Compiler leverages AI models for improved compiler optimizations and scalability. RefDNN estimates HPC metrics, including time, cost, and memory feasibility, for DNN workloads, while the DNN Training Estimator predicts training times for deep neural networks on HPC systems. CKN monitors and manages the cloud-to-edge computing continuum. Ultimately, Model Commons serves as a centralized platform for downloading, uploading, and searching models across various platforms. Fig. 2. Reference architecture of the Unified Component API (UCA). The design follows a narrow waist paradigm, where consumer services interact with producer services through standardized Data APIs and a common Data Schema. This enables seamless communication and integration across heterogeneous middleware components, ranging from intelligent middleware tools (e.g., RefDNN, DNN Estimator, Schedule Manager) to system middleware (e.g., TAPIS, Kubernetes) and HPC infrastructure (e.g., TACC, SDSC, OSC). In Stage 3, we analyzed the architectural patterns of all components and proposed a reference architecture for unification. A “narrow waist” design paradigm (depicted in Figure 2) was adopted, where the consumer service represents the user or service invoking the IMT component, and the producer service represents the IMT component itself. Data APIs on the producer side define the required inputs, while data APIs on the consumer side specify the expected outputs. The data schema standardizes the communication between producers and consumers. Based on this architecture, we defined a common interface, IMTComponent, comprising three core functions: Start,Stop,Terminate, and Restart. Each component must first be initialized with the necessary parameters to validate proper formatting and completeness. The Start function initializes the component and may return a dictionary of results, Terminate shuts down the component and its associated services, and Restart reinitializes the component either by sequentially calling Terminate and Start, or by executing component-specific restart logic. In Stage 4, we designed class diagrams for each component, ensuring they conform to the IMTComponent interface. These diagrams formalized the relationships between component-specific methods and the standardized lifecycle management functions. The class diagram for the Smart Scheduler is discussed in detail later in this paper. In Stage 5, we implemented the designed class diagrams and integrated them into a Python SDK that serves as the user gateway for managing IMT tools. This SDK simplifies interaction with IMT components, enabling researchers to operate middleware tools without needing to understand their individual APIs or deployment details. III. USE CASE: SMART SCHEDULER AND PLUG-AND-PLAY DESIGN The IMTComponent interface, along with the class diagram for the Smart Scheduler implementing this interface, is depicted in Figure 3. The SmartScheduler interface itself extends the IMTComponent interface. The TrainingSmartScheduler and InferenceSmartScheduler classes implement SmartScheduler, requiring them to also implement the methods defined in both IMTComponent and SmartScheduler. Training and inference were separated into two distinct classes to allow their logic to be tailored specifically for their respective workflows. The ProfileWorkload,DataGenerator, ModelTrainer,Inference, InformationManagementFramework, and ServiceManager classes are used by the Smart Scheduler to perform the necessary functionality for generating HPC schedules to submit jobs via TAPIS [7]. As these classes implement the IMTComponent interface, the Smart Scheduler can fully manage their lifecycle. For example, if the Smart Scheduler spawns processes such as ProfileWorkload and DataGenerator, invoking the Terminate function on the Smart Scheduler will call the Terminate functions of each spawned process, ensuring clean shutdowns. Having these classes implement IMTComponent also simplifies interactions with them. The Smart Scheduler can Fig. 3. Class diagram illustrating the Smart Scheduler architecture and its supporting components. Interfaces such as IMTComponent are implemented across key classes (e.g., ProfileWorkload,DataGenerator,ModelTrainer, and Inference) to enable lifecycle management and seamless orchestration of HPC workflows. start these services directly by calling their Start function and can communicate with them consistently, just as the IMTComponent interface standardizes user interactions with all IMT components. Plug-and-Play Capability: The reference architecture and IMTComponent class were designed with plug-and-play as a core priority. For example, if a developer rewrites the implementation of a dependency such as ProfileWorkload, DataGenerator,ModelTrainer, or Inference, the Smart Scheduler itself does not need to be modified, nor do client calls to the Smart Scheduler. The new class only needs to conform to the functions defined in IMTComponent and the corresponding functional interface, allowing the Smart Scheduler implementation to remain unchanged. This design also promotes research flexibility. Researchers can develop alternative implementations of components and quickly test their impact without altering the component’s core logic. For instance, to optimize the TrainingSmartScheduler, a researcher could create a new implementation of the ModelTrainer interface and test it directly within the framework. This capability supports rapid experimentation and evaluation, reinforcing the system’s extensibility and modularity. REFERENCES [1] Dhabaleswar K. (DK) Panda, Vipin Chaudhary, Eric Fosler-Lussier, Raghu Machiraju, Amit Majumdar, Beth Plale, Rajiv Ramnath, Ponnuswamy Sadayappan, Neelima Savardekar, and, Karen Tomko, ”Creating intelligent cyberinfrastructure for democratizing AI”, EEE AI Magazine, March 2024. [2] M. Vallabhajosyula, R. Ramnath, ”Towards Characterizing DNNs to Estimate Training Time using HARP (HPC Application Resource (runtime) Predcitor)”, Practice and Experience in Advanced Research Computing, pp. 483-485, September 2023. [3] M. Vallabhajosyula, S. Budhya, and R. Ramnath, ”Reference implementation of smart scheduler: a CI-aware, AI-driven scheduling framework for HPC workloads”, Practice and Experience in Advanced Research Computing, pp. 1-4, July 2024. [4] S. Withana and B. Plale, ”CKN: An Edge AI Distributed Framework”, 2023 IEEE 19th Int’l Conf on e-Science (e-Science), pp. 1–10. IEEE, October 2023. [5] Romero Rosas, Miguel Angel, Miguel Angel Torres Sanchez, and Rudolf Eigenmann. ”Should AI Optimize Your Code? A Comparative Study of Classical Optimizing Compilers Versus Current Large Language Models.” Proceedings of the 2025 Supercomputing Asia Conference. 2025. [6] Vallabhajosyula, Manikya Swathi, and Rajiv Ramnath. ”Insights from the HARP Framework: Using an AI-Driven Approach for Efficient Resource Allocation in HPC Scientific Workflows.” Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good. 2023. 341-344. [7] Stubbs, Joe, et al. ”Tapis: An API platform for reproducible, distributed computational research.” Advances in Information and Communication: Proceedings of the 2021 Future of Information and Communication Conference (FICC), Volume 1. Springer International Publishing, 2021.