WANDER Framework Offers Explainable Decision Support for HPC Tuning

Global: WANDER Framework Offers Explainable Decision Support for HPC Tuning

Researchers Ankur Lahiry, Banooqa Banday, Yugesh Bhattarai, and Tanzima Z. Islam introduced WANDER, a decision‑support framework for high‑performance computing (HPC) configuration, in a paper submitted to arXiv on June 4, 2025 and revised through December 25, 2025.

Background and Motivation

WANDER aims to address the complexity of HPC systems, which expose numerous interdependent configuration knobs affecting runtime, resource consumption, power usage, and variability. Existing tools predict outcomes but lack structured exploration and explanation capabilities.

Framework Overview

The framework combines predictive modeling with counterfactual analysis, allowing users to generate alternative configurations that align with specific performance goals and constraints. It presents suggestions through a composite trade‑off score that accounts for prediction uncertainty, causal consistency, and similarity to historical feature distributions.

Composite Trade‑off Scoring

By integrating causal models, WANDER evaluates the relationship between configuration features and target metrics, improving the trustworthiness of recommendations. The system also ranks alternatives based on how closely they match the user’s desired trade‑offs.

Experimental Validation

Experiments conducted on multiple HPC datasets demonstrated that WANDER can produce human‑readable, interpretable configuration suggestions that guide users toward achieving performance objectives while maintaining resource efficiency.

Implications for HPC Optimization

The authors emphasize that WANDER represents the first system to unify prediction, exploration, and explanation for HPC tuning under a single query interface, potentially streamlining the optimization workflow for researchers and practitioners.

The paper, listed under the Performance (cs.PF) and Artificial Intelligence (cs.AI) categories, is available on arXiv with DOI https://doi.org/10.48550/arXiv.2506.04049.

This report is based on information from arXiv, licensed under See original source. Source attribution required.