Oscar Hernandez
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Oscar Hernandez has a total of 219 co-authors. Here are the top co-authors by number of shared publications:
- Wael R. Elwasif
Oak Ridge National Laboratory
9 shared publications - Aaron Welch
Oak Ridge National Laboratory, Oak Ridge, USA
7 shared publications - Stephen Poole
Los Alamos National Laboratory, Los Alamos, USA
6 shared publications - Vivek Sarkar
Georgia Institute of Technology, Atlanta, USA
5 shared publications - William Fredie Godoy
Oak Ridge National Laboratory, Oak Ridge, USA
4 shared publications - Akihiro Hayashi
Georgia Institute of Technology, Atlanta, USA
3 shared publications
Latest Publications
- Open Access
- 19 August 2026
- Published by MDPI AG in Preprints.org
Early Experiences Using AI Software Agents on OLCF Systems: From Installation and Environments to Simulation Workflows for Molecular and Materials Sciences
Scientific workflows on large computing clusters increasingly combine massively parallel simulation engines, machine-learned models, and theory development with the technical aspects of installation, testing and debugging, batch scheduling, and designing architecture-specific GPU-accelerated software. On leadership-class systems, these workflows can be difficult to deploy successfully for production science due to compounding problems: compilation, installation, and package compatibility; accelerator runtime issues; scheduler behavior; and many others. We report early experiences using OpenAI's Codex and Anthropic's Claude Code as AI assistants for molecular simulation workflow development on OLCF resources. Our results show several important capabilities. With not much more than the user guide for the Frontier supercomputer, both assistants were able to help resolve complex build and deployment paths on this nonstandard HPC system. The overall speedup for completion of these types of difficult technical tasks, over the time it would take an unassisted human computational researcher, can approach several orders of magnitude. However, the amount of work that must be verified, due to problems with AI reliability, faithfulness and reproducibility, also grows accordingly, which in the case of HPC production workflows, can become unmanageable. This result supports the recently recognized shift from a production bottleneck to a verification bottleneck for AI in academic pursuits, including science and mathematics, and helps us to map out requirements for agentic AI for HPC-based computational science in this new shifting paradigm. Under expert supervision and with file-based provenance, AI coding systems could reduce the translation cost between scientific intent and machine-specific execution, as long as scientific truth can be codified and agentic workflows can be verified with auditable artifacts.
- 19 August 2026
- Published by MDPI AG in Preprints.org
Open AccessEarly Experiences Using AI Software Agents on OLCF Systems: From Installation and Environments to Simulation Workflows for Molecular and Materials Sciences
Scientific workflows on large computing clusters increasingly combine massively parallel simulation engines, machine-learned models, and theory development with the technical aspects of installation, testing and debugging, batch scheduling, and designing architecture-specific GPU-accelerated software. On leadership-class systems, these workflows can be difficult to deploy successfully for production science due to compounding problems: compilation, installation, and package compatibility; accelerator runtime issues; scheduler behavior; and many others. We report early experiences using OpenAI's Codex and Anthropic's Claude Code as AI assistants for molecular simulation workflow development on OLCF resources. Our results show several important capabilities. With not much more than the user guide for the Frontier supercomputer, both assistants were able to help resolve complex build and deployment paths on this nonstandard HPC system. The overall speedup for completion of these types of difficult technical tasks, over the time it would take an unassisted human computational researcher, can approach several orders of magnitude. However, the amount of work that must be verified, due to problems with AI reliability, faithfulness and reproducibility, also grows accordingly, which in the case of HPC production workflows, can become unmanageable. This result supports the recently recognized shift from a production bottleneck to a verification bottleneck for AI in academic pursuits, including science and mathematics, and helps us to map out requirements for agentic AI for HPC-based computational science in this new shifting paradigm. Under expert supervision and with file-based provenance, AI coding systems could reduce the translation cost between scientific intent and machine-specific execution, as long as scientific truth can be codified and agentic workflows can be verified with auditable artifacts.
- Open Access
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 620-627
Performance Analysis of Conveyors: Memory Dominates?
Small-message aggregation is critical for scaling irregular, communication intensive applications in high-performance computing. In this paper, contrary to conventional wisdom, we present the first systematic study showing that memory contention, not network bandwidth, is the dominant bottleneck in message aggregation runtimes. Using the state-of-the-art conveyors library as our reference implementation, we conducted extensive experiments on HPC systems featuring Slingshot 11 and InfiniBand interconnects, scaling to 16k cores (256 nodes) and processing 10s–100s GB of data. Our measurements reveal that interference between user data and aggregation buffers drives LLC miss rates to 77%, inflating memory costs by 2–3× over the algorithmic baseline. Consequently, we advocate for dedicated near-memory subsystems to improve the scalability and performance of message aggregation runtimes. This paper also demonstrates up to an order of magnitude higher latency for conveyor termination compared to a traditional HPC barrier, and it examines the impact of communication context isolation and the critical challenge of programmability.
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 620-627
Open AccessPerformance Analysis of Conveyors: Memory Dominates?
Small-message aggregation is critical for scaling irregular, communication intensive applications in high-performance computing. In this paper, contrary to conventional wisdom, we present the first systematic study showing that memory contention, not network bandwidth, is the dominant bottleneck in message aggregation runtimes. Using the state-of-the-art conveyors library as our reference implementation, we conducted extensive experiments on HPC systems featuring Slingshot 11 and InfiniBand interconnects, scaling to 16k cores (256 nodes) and processing 10s–100s GB of data. Our measurements reveal that interference between user data and aggregation buffers drives LLC miss rates to 77%, inflating memory costs by 2–3× over the algorithmic baseline. Consequently, we advocate for dedicated near-memory subsystems to improve the scalability and performance of message aggregation runtimes. This paper also demonstrates up to an order of magnitude higher latency for conveyor termination compared to a traditional HPC barrier, and it examines the impact of communication context isolation and the critical challenge of programmability.
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 238-252
Toward a Unified GPU-Aware OpenSHMEM Specification
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 238-252
Open AccessToward a Unified GPU-Aware OpenSHMEM Specification
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 217-228
Heading Towards Energy Savings or Just Performance: Case-Study on Message Aggregation-Based Runtimes
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 217-228
Open AccessHeading Towards Energy Savings or Just Performance: Case-Study on Message Aggregation-Based Runtimes
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 84-99
Accounting for the Invisible Watt: Bridging Node-Level and Cabinet-Level Power
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 84-99
Open AccessAccounting for the Invisible Watt: Bridging Node-Level and Cabinet-Level Power
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 32-43
Are We There Yet? Evaluating AMDFlang Fortran Compiler Readiness on Production Workloads
Fortran applications constitute a significant portion of large-scale HPC workloads deployed at many large computational centers. The continued viability of such applications depends on the availability of a robust development environment that can target the increasingly heterogeneous node architectures of modern systems. The amdflang Fortran compiler is AMD’s next-generation compiler that targets x86 CPUs with OpenMP offloading capabilities to AMD GPUs. Based on the upstream LLVM flang compiler, amdflang was officially released as part of ROCm 7 and is currently being deployed as part of the standard AMD programming environment on HPE Cray systems. Given the importance of this transition, this work explores the current capabilities and performance of amdflang, using a wide spectrum of applications from different domains that are representative of the Fortran HPC workloads commonly deployed at computational centers. We assess the capability of amdflang as it relates to base Fortran language support, as well as OpenMP on the CPU and GPU offloading to AMD devices. We discuss the changes and effort needed to port existing codes to use amdflang, and an initial performance evaluation of the ported codes. For a few cases, we also explore the compilation times. We take a breadth-first approach to exploration to provide a guide to the compiler developers, computational centers, and end users as they consider migrating their codes to this new compiler.
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 32-43
Open AccessAre We There Yet? Evaluating AMDFlang Fortran Compiler Readiness on Production Workloads
Fortran applications constitute a significant portion of large-scale HPC workloads deployed at many large computational centers. The continued viability of such applications depends on the availability of a robust development environment that can target the increasingly heterogeneous node architectures of modern systems. The amdflang Fortran compiler is AMD’s next-generation compiler that targets x86 CPUs with OpenMP offloading capabilities to AMD GPUs. Based on the upstream LLVM flang compiler, amdflang was officially released as part of ROCm 7 and is currently being deployed as part of the standard AMD programming environment on HPE Cray systems. Given the importance of this transition, this work explores the current capabilities and performance of amdflang, using a wide spectrum of applications from different domains that are representative of the Fortran HPC workloads commonly deployed at computational centers. We assess the capability of amdflang as it relates to base Fortran language support, as well as OpenMP on the CPU and GPU offloading to AMD devices. We discuss the changes and effort needed to port existing codes to use amdflang, and an initial performance evaluation of the ported codes. For a few cases, we also explore the compilation times. We take a breadth-first approach to exploration to provide a guide to the compiler developers, computational centers, and end users as they consider migrating their codes to this new compiler.
- 24 November 2025
- Published by Springer Nature
- p. 177-190
Characterizing GPU Energy Usage in Exascale-Ready Portable Science Applications
- 24 November 2025
- Published by Springer Nature
- p. 177-190
Characterizing GPU Energy Usage in Exascale-Ready Portable Science Applications
- 24 November 2025
- Published by Springer Nature
- p. 231-244
Power-Capping Metric Evaluation for Improving Energy Efficiency in HPC Applications
- 24 November 2025
- Published by Springer Nature
- p. 231-244
Power-Capping Metric Evaluation for Improving Energy Efficiency in HPC Applications
- Open Access
- 8 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-8
Comprehensive Evaluation of LLMs in HPC Code Performance Optimization
- 8 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 1-8
Open AccessComprehensive Evaluation of LLMs in HPC Code Performance Optimization