Multi-Fidelity Candidate Screening with a Simulator-Faithful Decoder for Risk-Aware Cloud Workflow Scheduling
Year: 2026
Type: software
Abstract: LIWSA-MF2 is a research codebase for evaluating workflow schedulers under multi-objective and risk-aware settings. It includes the LIWSA method, the guided **LIWSA+** search (The novelty), simulator-faithful scheduling, competing multi-objective algorithms, benchmark workflows, and scripts to regenerate reports, tables, and plots. The package contains **20 Pegasus workflow definitions** from five workflow families. It includes reference experiment outputs and the scripts needed to reproduce the analyses. ## 🧭 Research focus Workflow scheduling is a multi-objective problem: a schedule that minimizes completion time may cost more, while a cheaper schedule may take longer or be more vulnerable to uncertain execution times and slow virtual machines. This work studies a Pareto-based search approach for exploring those trade-offs, then examines whether more faithful simulation and multi-fidelity evaluation can make the search results more reliable and efficient. ### 🧠 Methodology 1. **Represent a schedule.** A candidate maps each workflow task, in topological order, to a virtual machine. 2. **Optimize competing objectives.** The search evaluates makespan and execution cost together, preserving non-dominated alternatives rather than collapsing them into a single weighted score. 3. **Explore alternative assignments.** LIWSA uses density-adaptive solitary and social search moves. LIWSA+ combines these with canonical evolutionary, critical-path, cost-directed, and estimate-guided moves. 4. **Preserve and adapt.** LIWSA+ uses Pareto-rank and crowding-based population survival, maintains an archive of evaluated non-dominated schedules, and adapts variation-operator selection based on observed contributions during the run. 5. **Evaluate fidelity and risk.** A simulator-faithful decoder models WorkflowSim scheduling behavior, including stage-in, per-VM input-file caching, and FCFS execution. Risk-aware experiments evaluate schedules across perturbed task lengths and machine speeds; CVaR can emphasize poor makespan outcomes in the scenario tail. 6. **Test multi-fidelity screening.** When full scenario evaluation is expensive, candidate pools can be screened using nominal decoder values, online ridge models, or boosted trees before selected candidates receive expensive evaluation. Random-selection controls and equal-budget comparisons are included. ### ✨ Research contributions examined The repository implements and evaluates the following methodological elements. These describe the scope of this work; they are not a claim that each element is individually unprecedented: - **A density-adaptive, Pareto-based LIWSA search** for discrete task-to-machine assignments, together with a population-based LIWSA+ extension. - **Search operators informed by schedule structure:** critical-chain moves target tasks contributing to the current makespan, while cost moves consider machine price per unit of work. - **Online adaptive operator selection** and an anytime archive of non-dominated solutions, with ablations to examine component effects. - **A simulator-faithful scheduling decoder** validated against full WorkflowSim, reducing discrepancies between search-time estimates and simulator-measured outcomes. - **Risk-aware and multi-fidelity evaluation** combining perturbed scenarios, CVaR objectives, candidate pre-screening, and learned surrogate models. - **A reproducible comparative study** against NSGA-II, SPEA2, MOEA/D, HEFT, and Min-Min across workflow families, workflow sizes, pricing settings, and algorithm ablations.
Source: Code Ocean
Author: Mohammed Alaa Ala'anzy
Institutions: Shandong University, University of Southern Denmark
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225620862