Performance-Aware Scheduling and Resource Allocation for Large-Scale Structural Simulations

Seismic performance assessment requires thousands of nonlinear response-history analyses, typically organized as incremental dynamic analysis (IDA) campaigns that share clusters with other users. These campaigns differ from generic batch workloads: every ground motion induces a chain of analyses whose next intensity level is needed only if the current one does not collapse, runtimes are governed by record length, model cost and intensity, and memory requests are sized for the worst case. We measured 810 real OpenSees analyses of four archetype frames under the 22 FEMA P695 far-field records and show that chain parallelism decays sharply as records collapse, that history-based runtime predictors are uncorrelated with actual runtimes, and that static memory requests exceed footprints by a median factor of 2.28. We propose PASS, a cross-task scheduler that combines an intensity-conditioned runtime and memory model, an online collapse-hazard estimator yielding the survival-weighted remaining work of every chain in constant time, a bound-guided critical-chain priority with aged shortest-job-first ordering, probability-gated and preemptible speculative execution of future intensity levels on idle cores, and memory right-sizing with out-of-memory feedback. In trace-driven experiments against FIFO, EASY backfilling, shortest-job-first with history-based and exact runtimes, static partitioning, Tetris and critical-path-first scheduling, PASS shortens the makespan of IDA batches by 33.9% and reduces the mean waiting time of online campaigns by 89.6% relative to the best baseline, raises core utilization from at most 57.8% to 74.6–79.5%, and requires about 142µs per scheduling decision. On 1024-core clusters, where every baseline is bounded by the longest chain, PASS reduces the makespan by 56%. Ablation, sensitivity and robustness experiments quantify the contribution of each component and show that the gains persist under runtime noise, heterogeneous nodes and bursty arrivals.

Authors

Institutions

Publication Details

Journal
Transactions on Environment Energy and Earth Sciences
Published
2026-11-13
DOI
https://doi.org/10.62051/np19jp25
Primary Topic
Distributed and Parallel Computing Systems
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Performance-Aware Scheduling and Resource Allocation for Large-Scale Structural Simulations

Teng Dai
Transactions on Environment Energy and Earth Sciences
Distributed and Parallel Computing Systems
article

Performance-Aware Scheduling and Resource Allocation for Large-Scale Structural Simulations

Teng Dai
article en

Abstract

Seismic performance assessment requires thousands of nonlinear response-history analyses, typically organized as incremental dynamic analysis (IDA) campaigns that share clusters with other users. These campaigns differ from generic batch workloads: every ground motion induces a chain of analyses whose next intensity level is needed only if the current one does not collapse, runtimes are governed by record length, model cost and intensity, and memory requests are sized for the worst case. We measured 810 real OpenSees analyses of four archetype frames under the 22 FEMA P695 far-field records and show that chain parallelism decays sharply as records collapse, that history-based runtime predictors are uncorrelated with actual runtimes, and that static memory requests exceed footprints by a median factor of 2.28. We propose PASS, a cross-task scheduler that combines an intensity-conditioned runtime and memory model, an online collapse-hazard estimator yielding the survival-weighted remaining work of every chain in constant time, a bound-guided critical-chain priority with aged shortest-job-first ordering, probability-gated and preemptible speculative execution of future intensity levels on idle cores, and memory right-sizing with out-of-memory feedback. In trace-driven experiments against FIFO, EASY backfilling, shortest-job-first with history-based and exact runtimes, static partitioning, Tetris and critical-path-first scheduling, PASS shortens the makespan of IDA batches by 33.9% and reduces the mean waiting time of online campaigns by 89.6% relative to the best baseline, raises core utilization from at most 57.8% to 74.6–79.5%, and requires about 142µs per scheduling decision. On 1024-core clusters, where every baseline is bounded by the longest chain, PASS reduces the makespan by 56%. Ablation, sensitivity and robustness experiments quantify the contribution of each component and show that the gains persist under runtime noise, heterogeneous nodes and bursty arrivals.

Transactions on Environment Energy and Earth SciencesVol. 7
University of Illinois Urbana-Champaign (US)
Industry, innovation and infrastructure
Distributed and Parallel Computing Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon