A Unifying Framework for Provisioning and Executing Computational Tools across Heterogeneous Computing Environments

Abstract

The landscape of distributed scientific computing is characterized by a high degree of heterogeneity, ranging from local workstations to high-performance clusters and cloud environments. Historically, this fragmentation forced users to manually access and navigate diverse execution interfaces or rely on disconnected, ad-hoc solutions to run computational tasks. This direct dependency on infrastructure implementation details compromises portability and forces users to manage the mechanics of where and how execution occurs, rather than focusing on the scientific tool itself.

In this thesis, we address these challenges by proposing a contract-based framework that enables the execution of computational tools across heterogeneous computing environments through a shared API. At the core of the framework is the Tool Contract, a formal abstraction that explicitly separates the semantic specification of a tool from its concrete runtime realization. Interaction with the framework is mediated through a unified RESTful API that manages the complete execution lifecycle, allowing users to submit executions, monitor their progress, and retrieve results without requiring direct access to or knowledge of the underlying execution environment. This interaction model is supported by a distributed architecture comprising a central Coordinator for global orchestration, environment-specific Launchers that adapt the framework to heterogeneous infrastructures, and a portable Agent responsible for tool execution and artifact handling.

We validated the proposed framework through the implementation of a functional prototype and a systematic empirical evaluation. The results indicate that the system functions as intended and allows users to interact with tools uniformly and independently of the computing environments in which they are ultimately executed. The performance analysis shows that the architectural overhead introduced by the abstraction remains in the millisecond range. Under the evaluated conditions, this level of overhead appears compatible with a broad range of workloads, including both short-lived executions and long-running scientific computations. Overall, the results suggest that the proposed architecture addresses the interoperability challenges defined in the initial use case by enabling users to reason about and invoke tools independently of their concrete computing environments, while the underlying system transparently manages interaction with heterogeneous infrastructure.