WisIO

WisIO converts large HPC I/O traces into explanations of where bottlenecks occur and what causes them. It examines the same trace through time, process, and file views, then uses metric-driven classification and an extensible reasoning engine to connect bottlenecks with root causes. The ICS 2025 paper reports up to 805x greater bottleneck coverage than a single view and up to 11x faster analysis than the tools it compared against.

Analysis Model​

File-level summaries can miss interactions that emerge only across processes or over time. Fixed detection rules also become brittle as workflows and storage systems change. WisIO separates measurement from interpretation: metrics classify anomalous records, while rules explain the behaviors associated with each finding.

The framework processes traces in five stages:

1. Trace collection reads Darshan DXT, DFTracer, or Recorder output.
2. View generation organizes the same events by time, process, and file.
3. Classification scores metric-level bottlenecks within each view.
4. Reasoning attaches one or more observed behaviors to each finding.
5. Diagnosis presents the bottleneck, its severity, and its likely causes.

Published Results​

WisIO (opens in a new tab) processes multi-perspective views in parallel and distributes analysis of multi-terabyte traces. The ICS 2025 paper reports up to 805x greater bottleneck coverage than single-perspective analysis. Its classification engine processed 340,000 bottlenecks per second and its reasoning engine about 35,000 per second. Across five HPC workloads, it analyzed traces up to 11x faster and identified up to 144x more bottlenecks than the tools it compared against.

IOMax (opens in a new tab) provides the out-of-core analysis layer for large traces. On Lawrence Livermore National Laboratory's Lassen supercomputer, the PDSW 2023 paper reports 8.6x faster queries, an 11x lower memory footprint, up to 11.4x faster data slicing, and up to 7x faster end-to-end analysis.

Montage: Interacting Workflow Stages​

For Montage, WisIO analyzed 12.3 million I/O operations from 11,400 processes and 19,600 files in 48 seconds. It found 495 distinct bottlenecks with 1,735 associated reasons. Seven of the workflow's applications issued very small reads, averaging 76 bytes.

CM1: Metadata Contention​

For CM1, WisIO analyzed 27,500 operations from 1,200 processes and 775 files in 17 seconds. It found 75 bottlenecks with 176 reasons. The time view showed all ranks issuing concurrent open calls during initialization, before the first rank on each node (32 ranks) read the configuration files; 31 of those ranks, all but rank 0, issued only reads.

1000 Genomes: Analysis at Trace Scale​

For 1000 Genomes, WisIO processed 646 million operations from 2,700 processes and 21.2 million files in 24 minutes. It identified 430,000 bottlenecks and 560,000 associated reasons. The analysis connected small writes and metadata activity within mutation_overlap to several distinct root causes.

Extensibility​

New perspectives can use the same classification and reasoning pipeline. New rules can add domain or system knowledge without changing the metric-driven detector.

WisIO v0.2.2 (opens in a new tab), released in September 2026, is the current public version of the analysis tool. Subsequent work with Lawrence Livermore National Laboratory on ML-driven I/O intent extraction continues in DeepIO.