witty-profiler Skills

The skills directory is intended for agent-facing integrations in witty-profiler.

Its purpose is to expose the collection framework through simple, reusable skills so agents can trigger profiling workflows and consume structured results without dealing with low-level collector details.

Available Skills

1. dataflow-topology-restore

Purpose: Reconstruct data flow topologies from witty-profiler (Anansi) system graphs.

Use Cases:

• Analyze system topology and communication paths
• Identify NCCL/HCCL communication patterns
• Detect cross-NUMA access patterns
• Understand NPU/GPU data paths

2. hotspot-thread-discovery

Purpose: Identify performance hotspot threads and processes in AI training systems.

Use Cases:

• Analyze CPU usage patterns and detect hotspot threads
• Investigate NUMA affinity issues
• Analyze context switches and contention
• Identify performance bottlenecks in multi-threaded AI workloads
• Diagnose compute, communication, memory, and synchronization bottlenecks

3. bottleneck-identification

Purpose: Systematic methodology for diagnosing performance bottlenecks in AI training infrastructures using the 7-layer bottleneck framework.

Use Cases:

• Comprehensive system performance diagnosis
• Identify bottlenecks across all layers (Compute, Memory, Interconnect, Network, Storage, Control Plane, Data Plane)
• Pattern matching for known bottleneck types
• Generate actionable bottleneck diagnosis reports
• Provide prioritized optimization recommendations

Status

This part of the project is still under development.