Resource-Aware HPC
Doctoral Dissertation · The Ohio State University

AI-Driven Resource Optimization for High-Performance Computing

A cyberinfrastructure-aware framework that derives a job's resource request from the application and the target system, rather than asking the user to guess it, and revises it as the job runs at the HPC center, while at the edge it selects the model variant that fits the field device before deployment.

Manikya Swathi Vallabhajosyula · CSE
Advisor: Dr. Rajiv Ramnath
Committee: Dr. D.K. Panda, Dr. Hari Subramoni, Dr. Suren Byna
Defense: June 11, 2026 · NSF ICICLE · Tapis · OSC
The cost of guessing

Human-driven provisioning wastes measured, persistent resources

Before a job runs, users hand-estimate wall-time, memory, and accelerators. Across leadership-class systems the result is the same shape of waste, and it is documented, not anecdotal.

~15%
of GPU-hours fully idle
NERSC Perlmutter, one-month study
4%
median host-CPU use on GPU nodes
NERSC Perlmutter
20.6%
mean GPU use, 40%+ nodes idle
OSC Pitzer, week of May 1, 2026
2–3×
wall-time over-requested by users
queue-wait literature
30–50%
of GPU-hours unused in batch ML
cost-aware scheduling studies
What goes wrong

Recurring failure patterns

The waste is not random misconfiguration. It follows from asking people to predict machine behavior on evolving infrastructure that now spans the edge-to-center continuum.

At the center
Wall-time over-estimation. Users request 2–3× their actual need, so reservations hold idle and queues lengthen for everyone.
GPU underutilization in AI jobs. Accelerators sit idle during data loading, gradient sync, and hyperparameter search.
Reactive resubmission. Jobs that exceed a request are killed and re-queued, doubling effective wait and discarding partial work.
Static scheduling. Once dispatched, a classical scheduler cannot re-tune a job from observed performance.
Heterogeneous stages. Multi-stage pipelines over-provision to the worst-case stage everywhere.
At the edge
Blind edge deployment. Models reach field devices without knowing whether they fit the power and memory budget or meet the accuracy the science needs. A poor match is discovered only after a wasted deployment.
Challenge to component

A direct mapping

Each failure pattern has one primary component that answers it, with a few drawing support from others. The Reference Architecture composes them into a single closed loop.

Failure patternAnswered byWith support from
Wall-time over-estimationHARPiScheduler
GPU underutilization in AI jobsDNN EstimatorHARP
Reactive resubmissionHARPiScheduler
Static scheduling, no runtime adaptationiSchedulerRules Engine
Heterogeneous workflow stagesHARPCI Database
Constant CI / middleware updatesCI DatabaseRules Engine
Blind edge model deploymentML Field PlannerCKN telemetry
Abstract

High-performance computing underpins modern scientific discovery, yet researchers struggle to use it efficiently. Before submitting a job, users must estimate wall-time, memory, node counts, accelerator needs, and scheduling configurations, usually by hand. The result is chronic overprovisioning, idle accelerators, long queue waits, premature job termination, and repeated resubmission. The problem deepens with AI workloads and extends to the edge, where energy budgets and on-device feasibility decide whether a model can run on field hardware.

This dissertation answers with an AI-driven, cyberinfrastructure-aware framework for resource estimation and adaptive workflow orchestration. It builds estimation, validation, scheduling, monitoring, and edge planning as programmable components that call one another through well-defined endpoints, which also positions the work as a foundation for future agentic systems. The framework contributes six named components, composed into a reference architecture with an animal-ecology workflow as the worked example: edge camera-trap detection that degrades in the field triggers a center-side retraining run that the framework estimates, validates, schedules, and redeploys.