skills.rest
primatrix
primatrixOFFICIAL
Explore all skills by this author →
GitHub

comm-analysis

Analyze communication primitives in TPU pretraining profiles for bandwidth utilization and overlap.

5|Updated Feb 4, 2026
Search Tags:#performance optimization#communication analysis#pretraining#tpu#profile analysis
One-click install
npx skills add https://github.com/primatrix/skills --skill comm-analysis
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: comm-analysis
Source: https://github.com/primatrix/skills/tree/main/plugins/tpu-perf/skills/comm-analysis
Command: npx skills add https://github.com/primatrix/skills --skill comm-analysis

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires xplane_pb2, hlo_pb2, op_stats_pb2, protobuf, yaml, and includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill analyzes communication primitives in TPU pretraining profiles, providing insights into communication bandwidth utilization and overlap with computation. It helps identify performance bottlenecks and optimize TPU utilization.

Core Features & Use Cases

  • Communication Analysis: List, attribute, and analyze communication primitives (async/sync, TC/SC, P2P) with rich attributes like bandwidth, latency, and overlap.
  • Per-axis Bandwidth Utilization: Compute and compare NCCL bus bandwidth and peak ICI link bandwidth for each axis, with utilization percentage.
  • Compute/Comm Overlap: Calculate and report compute-to-communication overlap for each step.
  • Critical-path Attribution: Identify and report the critical-path exposure for each communication primitive.
  • Use Case: Analyze a TPU pretraining profile to identify high-bandwidth communication primitives and optimize data placement or collective communication patterns for better performance.

Quick Start

Use the comm-analysis skill to analyze the communication primitives in the profile located at /tmp/tensorboard/tensorboard/plugins/profile/dp8_fsdp128.

Frequently Asked Questions about comm-analysis

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I analyze communication primitives in a TPU pretraining profile?▼

To analyze communication primitives in a TPU pretraining profile, use a tool that parses xplane.pb, hlo_proto.pb, and op_stats.pb files. It calculates bandwidth utilization, compute overlap, and critical-path exposure to identify performance bottlenecks.

What is critical-path attribution for TPU communication operations?▼

Critical-path attribution for TPU communication identifies the exact exposure of each communication primitive on the critical path. It analyzes pretraining profiles to pinpoint sync or async operations that directly delay overall execution.

How do I calculate NCCL bus bandwidth and peak ICI link utilization per axis?▼

To calculate NCCL bus bandwidth and peak ICI link utilization per axis, compare observed communication bandwidth against a mesh-spec YAML topology. This yields a precise utilization percentage for each topology axis.

Do I need a mesh-spec YAML file to measure TPU communication overlap with computation?▼

A mesh-spec YAML file is not required to measure basic compute and communication overlap, but it is necessary to resolve peak ICI link bandwidth and accurately calculate per-axis bandwidth utilization percentages.

Why does my TPU pretraining profile show poor bandwidth utilization for collective operations?▼

Poor TPU pretraining bandwidth utilization often stems from unoptimized data placement or inefficient collective communication patterns. Analyzing individual primitive latency and overlap helps identify these specific bottlenecks.

Can I use this comm-analysis tool to optimize data placement for TPU pretraining?▼

Yes, comm-analysis identifies high-bandwidth communication primitives and reports their compute overlap. These insights directly guide data placement adjustments and collective pattern optimizations for better TPU performance.

Related Skills

Explore More Software Engineering Skills→
primatrixprimatrix

memory-profile

Analyze TPU pretraining profiles to determine HBM peak occupancy and buffer allocation.

Official
Advanced
camoneartcamoneart
5

microservices-patterns

Design microservice boundaries, contracts, and inter-service communication patterns for distributed systems.

Community
Advanced
flashinfer-aiflashinfer-ai
264

extract-kernel-definitions

Extract GPU kernel schemas from SGLang models into deduplicated Definition JSON files.

Official
Advanced
primatrixprimatrix

compute-breakdown

Analyze TPU pretraining compute efficiency from XSpace/XPlane profile data.

Official
Advanced