Publications

Preprint arXiv 2026

COMPASS: A Unified Decision-Intelligence System for Navigating Performance Trade-off in HPC

Ankur Lahiry, Banooqa Banday, Yugesh Bhattarai, Tanzima Z Islam, and Mohammad Zaeed

Decision Intelligence HPC Explainability Uncertainty Ranking

Decision-intelligence framework for HPC configuration optimization using uncertainty-aware ranking and explainable recommendations.

Key Contribution: Introduced an uncertainty-aware decision-intelligence framework for HPC with up to 100× faster training and 80× faster inference than generative baselines.

Scaled to 1.3B samples (126 GB) for high-volume HPC traces.
Balanced performance, cost, and reliability in a single recommendation pipeline.
Delivered explainable outputs with uncertainty-aware ranking for system decisions.

Workshop Euro-par 2025

A Distributed Framework for Causal Modeling of Performance Variability in GPU Traces

Ankur Lahiry, Ayush Pokharel, Banooqa Banday, Seth Ockerman, Amal Gueroudji, Mohammad Zaeed, Tanzima Z Islam, and Line Pouchard

GPU Causal Inference HPC Performance Variability

Distributed causal analysis framework for GPU traces that identifies root causes of performance variability in HPC workloads.

Key Contribution: Framed large-scale GPU performance variability analysis as a distributed causal inference problem for root-cause discovery.

Modeled performance variability in GPU traces with causal structure rather than correlation alone.
Scaled analysis across large trace collections for HPC workflows.
Targeted root-cause identification for recurring performance anomalies.

Poster SSDBM 2025

Scalable GPU Performance Variability Analysis Framework

Ankur Lahiry, Ayush Pokharel, Seth Ockerman, Amal Gueroudji, Line Pouchard, and Tanzima Z Islam

GPU Distributed Systems HPC Trace Analysis

Distributed GPU trace-analysis framework that reduces memory overhead and improves scalability for large HPC log pipelines.

Key Contribution: Improved GPU trace-analysis scalability by 67% through distributed partitioning and parallel processing.

Reduced analysis time for repeated GPU trace workloads.
Lowered memory overhead on large trace datasets.
Enabled faster detection of stalls, variability, and bottlenecks.

Preprint arXiv 2025

WANDER: An Explainable Decision-Support Framework for HPC

Ankur Lahiry, Banooqa Banday, Yugesh Bhattarai, and Tanzima Z Islam

Explainable AI HPC Anomaly Detection Graph Learning

Explainable HPC decision-support framework that turns complex system logs into interpretable signals for early anomaly detection.

Key Contribution: Unified graph-based modeling and explainable AI for earlier and more interpretable diagnosis of HPC performance anomalies.

Converted complex HPC logs into actionable diagnostic signals.
Improved interpretability for anomaly detection decisions.
Supported earlier diagnosis of unusual large-scale system behavior.

Research Paper COMPSAC'24 2024

Reimagine Application Performance as a Graph: Novel Graph-Based Method for Performance Anomaly Classification in High-Performance Computing

Chase Phelps, Ankur Lahiry, Tanzima Z Islam, and Line C Pouchard

Graph Learning HPC Anomaly Detection Performance Analytics

Graph-based application-performance representation for anomaly classification in HPC using graph neural networks.

Key Contribution: Showed that graph-structured performance modeling improves anomaly classification in large-scale HPC systems.

Represented tasks and resources as structured graph relationships.
Applied graph neural networks to HPC anomaly classification.
Improved expressiveness over flat performance feature views.

Research Paper ICMLA'23 2023

Novel Representation Learning Technique Using Graphs for Performance Analytics

Tarek Ramadan, Ankur Lahiry, and Tanzima Z Islam

Graph Learning Performance Prediction HPC Representation Learning

Graph representation learning method that transforms tabular HPC performance data into graphs for performance prediction.

Key Contribution: Converted tabular HPC performance data into graph structure to improve representation learning for regression tasks.

Mapped tabular performance records into graph form for richer dependencies.
Applied graph neural network reasoning to performance analytics.
Targeted execution-time prediction and related regression problems.

Poster SC'23 2023

Graph Based Anomaly Detection in Chimbuko: Feasible or Fallible?

Chase Phelps, Ankur Lahiry, Tanzima Z Islam, and Christopher Kelly

HPC Anomaly Detection Graph Learning Performance Analytics

Evaluation of graph-based anomaly detection for Chimbuko performance analytics in large-scale supercomputing applications.

Key Contribution: Assessed the feasibility of graph-based deep learning for anomaly classification inside the Chimbuko HPC analytics framework.

Tested graph-based anomaly classification in an operational HPC analytics pipeline.
Focused on monitoring efficiency in large-scale supercomputing applications.
Surfaced practical trade-offs of graph methods for real HPC anomaly workloads.