Public Abstract

DE-SC0025603: Dynamic Space-Time Memory Curation for Traceable Wafer-Scale Agent-Based Models

Award Status: Active

Institution: Michigan State University, East Lansing, MI
UEI: R28EKN92ZTZ9
DUNS: 193247145

Most Recent Award Date: 09/12/2024
Number of Support Periods: 1
PM: Perumalla, Kalyan

Current Budget Period: 09/01/2024 - 08/31/2025
Current Project Period: 09/01/2024 - 08/31/2026
PI: Dolson, Emily

Supplement Budget Period: N/A

Public Abstract

Dynamic Space-Time Memory Curation for Traceable Wafer-Scale Agent-Based Models Luis Zaman, University oF Michigan (Principal Investigator) Emily Dolson, Michigan State University (Co-Investigator) Agent-Based Models or Simulations (ABMS) play a key role in informing policymakers, where accomplishing macro scale policy objectives depends on policy decisions that influence microscale interactions between individual actors. In sectors ranging from agriculture to public health, Parallel Discrete Event Simulation (PDES) approaches have enabled highly impactful work with ABMS tracking hundreds of thousands of agents across thousands of locations. Emerging AI/ML accelerator hardware, such as the Cerebras Wafer-Scale Engine (WSE) and Graphcore Intelligence Processing Unit (IPU), could enable an order of magnitude increase in the size of ABMS, increasing their utility in domains where greater scale and detail are necessary to capture modeled systems' emergent behavior. However, with up to hundreds of thousands of available compute cores, these accelerators are also highly distributed, memory-constrained, and bandwidth-limited, which restricts their usefulness for traditional PDES/ABMS workloads. This project aims to overcome these limitations by establishing a framework of "inferential observability," through which the fidelity and specificity of extracted data may be tuned to balance hardware constraints within the tolerances of modeling objectives. To this end, we will extend space-time memory algorithms to develop a versatile, efficient, and robust interface between on-device PDES/ABMS experiments and dynamic, inquiry-driven data extraction. We will use national-scale infectious disease ABMS on the WSE and IPU platforms as a testbed application to characterize performance and data quality characteristics of our inferential observability-based approaches. Through incorporation of detailed within-individual dynamics of infection at unprecedented scales, the computational power afforded by hardware accelerators will enable ABMS experiments testing the role of public health policy in shaping pathogen evolution. Our proposed methodology is applicable across research domains, as the scale-up of PDES/ABMS is a current bottleneck for much of complex systems research. Inferential observability is ultimately designed to better align data storage and extraction to experimental objectives by facilitating dynamic, context-driven analyses. In addition, bandwidth and memory savings enabled by these methods stand to benefit other classes of hardware systems, including traditional cluster computing. To facilitate the uptake of AI/ML accelerator capabilities across broader High-Performance Computing (HPC) and ABMS/PDES communities, we will make publicly available plug-and-play Python, C++, Cerebras Software Language (CSL), and Poplar implementations of developed algorithms, as well as extensible evo-epidemiological ABMS/PDES frameworks tailored to the WSE and IPU platforms. As an outcome, this work will enable scientists across diverse domains to apply ABMS/PDES methodology in ways that were previously intractable due to scaling concerns.

Dynamic Space-Time Memory Curation for Traceable Wafer-Scale Agent-Based Models
Luis Zaman, University oF Michigan (Principal Investigator)
Emily Dolson, Michigan State University (Co-Investigator)

Agent-Based Models or Simulations (ABMS) play a key role in informing policymakers, where accomplishing macro scale policy objectives depends on policy decisions that influence microscale interactions between individual actors. In sectors ranging from agriculture to public health, Parallel Discrete Event Simulation (PDES) approaches have enabled highly impactful work with ABMS tracking hundreds of thousands of agents across thousands of locations. Emerging AI/ML accelerator hardware, such as the Cerebras Wafer-Scale Engine (WSE) and Graphcore Intelligence Processing Unit (IPU), could enable an order of magnitude increase in the size of ABMS, increasing their utility in domains where greater scale and detail are necessary to capture modeled systems' emergent behavior. However, with up to hundreds of thousands of available compute cores, these accelerators are also highly distributed, memory-constrained, and bandwidth-limited, which restricts their usefulness for traditional PDES/ABMS workloads. This project aims to overcome these limitations by establishing a framework of "inferential observability," through which the fidelity and specificity of extracted data may be tuned to balance hardware constraints within the tolerances of modeling objectives. To this end, we will extend space-time memory algorithms to develop a versatile, efficient, and robust interface between on-device PDES/ABMS experiments and dynamic, inquiry-driven data extraction. We will use national-scale infectious disease ABMS on the WSE and IPU platforms as a testbed application to characterize performance and data quality characteristics of our inferential observability-based approaches. Through incorporation of detailed within-individual dynamics of infection at unprecedented scales, the computational power afforded by hardware accelerators will enable ABMS experiments testing the role of public health policy in shaping pathogen evolution.

Our proposed methodology is applicable across research domains, as the scale-up of PDES/ABMS is a current bottleneck for much of complex systems research. Inferential observability is ultimately designed to better align data storage and extraction to experimental objectives by facilitating dynamic, context-driven analyses. In addition, bandwidth and memory savings enabled by these methods stand to benefit other classes of hardware systems, including traditional cluster computing. To facilitate the uptake of AI/ML accelerator capabilities across broader High-Performance Computing (HPC) and ABMS/PDES communities, we will make publicly available plug-and-play Python, C++, Cerebras Software Language (CSL), and Poplar implementations of developed algorithms, as well as extensible evo-epidemiological ABMS/PDES frameworks tailored to the WSE and IPU platforms. As an outcome, this work will enable scientists across diverse domains to apply ABMS/PDES methodology in ways that were previously intractable due to scaling concerns.