Visualize the Invisible: Exposing Causal Layers in GPU Performance Analysis through Milestone Abstractions
Authors
Yuanhuan Deng (Brandeis University), Katherine E. Isaacs (The University of Utah), Yifan Sun (William & Mary)
Presentation
- Session
- The more the merrier
- Time
- Tuesday, Nov 10, 11:00 – 11:12 (US/Eastern) · session 10:00 – 11:30
- Location
- Hall America north
Keywords
GPU performance analysis, visual analytics, causal reasoning, execution tracing, milestone abstraction
Abstract
Visual analytics has the potential to provide critical support in computer architecture design by helping architects interpret execution behavior, but current tools are limited to exposing what happened during execution rather than why, relying on user expertise rather than visual evidence to identify true causes of performance issues. Expertise can help analysts formulate hypotheses, but analysts cannot validate those hypotheses through visualization when causal factors are not visible. We design a visual analytics system that makes causal factors, such as task blocking and underlying mechanisms, visible. At the core of our approach is a conceptual abstraction, the Milestone, a data primitive that simultaneously marks the resolution of prior blocking conditions and assigns reasons to waiting states. Milestones convert execution traces from records of activity into records of causation. We further design three complementary visualizations that leverage Milestones, enabling us to show summaries of blocking patterns, analyze how patterns evolve over time, and navigate causal relationships across parent-subtask hierarchies. We evaluate our approach through a study with GPU architecture researchers, showing that the abstraction and resulting visualizations enable participants to validate hypotheses about an architecture's performance directly from visual evidence. Finally, we discuss how the Milestone abstraction can be adapted to support causal analysis in other expertise-intensive domains.
For Practitioners
This paper is most directly relevant to computer architects and GPU architecture researchers who use simulators to diagnose performance bottlenecks, and to the simulator and profiling-tool developers who build the tracing infrastructure they depend on. Beyond that specific community, the work speaks to performance engineers and observability practitioners working with execution traces in distributed systems, databases, and compilers, where the same gap exists between recording that something was slow and recording why. Finally, visualization designers working in any expertise-intensive domain will find the diagnostic framing applicable to their own design studies. Practitioners can adopt the Milestone primitive directly as an instrumentation pattern: insert a record at each point where a task makes forward progress, capturing both what became available and what had been blocking until that moment. This requires modest changes to an existing tracer and yields blocking-reason data that standard start/end tracing discards. The accompanying visualizations (blocking summaries, temporal blocking-reason analysis, and a hierarchy-consolidating detail view) offer reusable designs for reading that data.