GPUTraceAnomalyScoring

Overview

GPUTraceAnomalyScoring scores every canonical span against dense peer-group baselines. The scoring policy is explicit and replaceable: duration deviation and error-rate deviation have independent weights, duration can be slow-only or two-sided, and the caller owns the threshold.

At a glance

Question	Answer
Problem	Score canonical spans against explicit peer-group duration and error baselines.
Reads / writes	Reads span metrics, group IDs, and baselines; writes per-span scores, anomaly mask, and summary.
Ownership	Public inputs and outputs are caller-owned; scratch storage is graph-owned transient memory.
Output contract	Source-aligned scores/mask plus a compact maximum and validation summary.
Expected work	Linear scoring plus bounded summary reduction.
Chunks	Preserves declared views and source identity; it does not implicitly concatenate or repack chunks.
Conditions / budgets	Supports maximumRowsPerPass and resumable publication of complete generations.
Neighborhood	peer baselines + trace columns → GPUTraceAnomalyScoring → filter, color, label, or aggregation.

CostSelected span count and reduction; small summaries do not remove per-span scoring work.

Common mistakeDo not present policy scores as causal explanations.

When to use

Use this contributor after a grouping or comparison stage has produced explicit peer baselines. It is intended for explainable policy scores, not model training or an embedded observability policy.

Usage

import {GPUTraceAnomalyScoring} from '@luma.gl/experimental/gpu-trace';

new GPUTraceAnomalyScoring({
  groupIndices,
  durations,
  errorMask,
  baselineDurationMeans,
  baselineDurationStandardDeviations,
  baselineErrorRates,
  threshold: 3,
  maximumRowsPerPass: 262_144,
  output: {scores, anomalyMask, summary}
}).addToGraph(graph);

Baselines may come from a saved trace, a deployment cohort, or another GPU scheduling aggregation. The primitive does not embed one observability product's grouping or anomaly semantics.

Score

For each valid span, the default slow-only score is:

durationWeight × max((duration - peerMean) / max(peerDeviation, floor), 0)
+ errorWeight × abs(errorValue - peerErrorRate)

Set durationMode: 'two-sided' to use the absolute duration z-score. Invalid group indices, durations, or baselines produce a zero score, stay out of the mask, and set a validation flag.

Inputs and outputs

View	Meaning
scores	Nonnegative policy score per canonical span
anomalyMask	One when the score reaches threshold
summary[0]	Number of thresholded anomalies
summary[1]	Maximum score encoded as positive f32 bits
summary[2]	Lowest canonical index attaining the maximum score
summary[3]	Invalid-group, invalid-duration, invalid-baseline, and overflow flags

The per-span outputs remain GPU-resident for rendering, filtering, label priority, or follow-up aggregation. A UI can sample only the four-word summary.

Scoring, summary reduction, and stable maximum selection publish invocation bounds to the command graph. maximumRowsPerPass optionally turns each source chunk and full-output reduction into smaller resumable nodes. Applications can combine that with compiled.createExecution(...), submit one bounded step at a time, and cancel between steps without exposing partial results as a completed analysis.

Execution and ownership

The contributor adds graph work only. Source columns and public outputs are caller-owned; graph scratch may use transient allocations. Submission, cancellation, and publication remain with the application.

Capacity, validation, and failure behavior

Invalid groups, durations, and baselines are excluded and reported through summary flags. Output capacity is fixed by the canonical span count compiled into the graph.

Performance

Work is linear in the selected span count plus a bounded summary reduction. Use maximumRowsPerPass and resumable execution for large global analyses.

Limitations

Peer construction and baseline policy are intentionally outside this contributor. Scores identify statistical deviation under the supplied policy; they do not establish causality.

Comparison workflow

Use stable operation dictionary IDs as groupIndices, produce baseline columns from the reference trace, and score the current trace without CPU row materialization. The same score and mask columns can feed GPUTraceAggregation, histograms, semantic colors, and selection workflows. Compose GPUTraceComparison in the same command graph when the UI also needs group-level count, duration, and error-rate deltas.