CANN/hcomm HCCL可视化分析指南

HVRM Insight User Guide

【免费下载链接】hcommHCOMM（Huawei Communication）是HCCL的通信基础库，提供通信域以及通信资源的管理能力。项目地址: https://gitcode.com/cann/hcomm

1. Overview

HVRM Insight is a distributed operator visualization analysis tool for unified display and linked analysis of graph structures, memory snapshots, and verification error information during the execution of HCCL collective communication operators. The tool includes three pages:

2. Installation and Startup

2.1 Prerequisites

HVRM Insight depends on data files output by the Checker plugin. Before use, ensure Checker has been fully executed and the following two data output switches are enabled in the Checker configuration file:

The Checker configuration file is located at/path/to/hccl_vm_install/plugin/checker/manifest.jsonby default.

"setting": { "enable_insight_dump": true, "enable_memory_snapshot_dump": true }

• enable_insight_dump: When enabled, Checker outputs the graph structure and verification result files required by Insight during execution.
• enable_memory_snapshot_dump: When enabled, Checker outputs memory snapshot files for each rank during execution.

Both switches default tofalse. If not enabled, Insight will not be able to read the complete analysis data.

2.2 Data Directory Structure

The tool reads datasets from thedata/directory. Each dataset is an independent subdirectory:

data/ └── <dataset_name>/ ├── manifest.json # Dataset metadata (required) ├── graph/ # Graph structure files (DAG) ├── memory/ # Memory snapshot files └── validation/ └── issues.msgpack # Verification error records (optional)

When optional directories are missing, the corresponding page displays an empty state.

2.3 Method 1: Via hccl-vm Plugin (Recommended)

# Install and start the backend service hccl-vm plugin install insight # After successful installation, the terminal outputs the access URL (default `http://localhost:8000`). Open it in your browser. # View the access URL again hccl-vm plugin run @insight # Uninstall the plugin (also stops the backend service) hccl-vm plugin uninstall @insight

2.4 Method 2: Local Direct Startup

# Enter the insight plugin directory python3 serve.py

The backend listens onhttp://localhost:8000by default, serving both the frontend page and the data API.

2.5 Verify Successful Startup

After opening the URL in the browser, if you can see the dataset list on the Dashboard page, startup was successful. If the page is blank, check whether thedata/directory exists and contains themanifest.jsonfile.

3. Scenario Demonstrations

3.1 Scenario 1: Select Dataset and View Overview (Dashboard)

Browse all datasets, click to select a target dataset, view operator information and statistics, then enter the correlation analysis page.

3.2 Scenario 2: Linked Analysis of Memory and Task Graph (MemView)

Through the linked display of the memory timeline and DAG task graph, locate memory operations for a specific step, view node details, and trace data sources.

3.3 Scenario 3: View Errors and Navigate to the Problem Site (Analytic)

View error details on the Analytic page or the error list on the left side of MemView, then one-click navigate to the associated DAG node or memory context.

4. Page Details

4.1 Dashboard Page

Dashboard is the starting point for using the tool. The page is divided into three areas from left to right:

• Left: Operator summary and Rank tree. The Rank tree supports selection/deselection, and the selection is synchronized to all pages.
• Center: Dataset list. Click a row to select a dataset; the left and right panels refresh simultaneously.
• Right: Details panel for the selected dataset, showing statistics and navigation buttons.

After selecting a dataset, detailed statistics are displayed on the right:

The bottom of the details panel provides two navigation entry points:

• Enter Correlation View: Navigate to MemView with the current dataset and rank selection.
• View Diagnostics: Navigate to Analytic to view all verification errors for this dataset.

4.2 MemView Page

MemView displays the memory timeline and DAG task graph in the same view for linked analysis. It is the core page for troubleshooting.

Page Layout:

• Left: Rank selection tree, issue list, search panel.
• Top: Memory timeline — shows the buffer state evolution for each rank by step.
• Bottom: DAG task graph — displays task nodes and dependencies in Rank/Stream swimlanes.
• Right: Details panel — shows detailed information for the selected step or node.

4.3 Core Coordination Mechanism: Step, Rank, and Task Interplay

• Select a step on the timeline → DAG highlights the corresponding node synchronously.
• Click a node in the DAG → the right panel switches to node details.
• Click a navigation button on the right → the timeline locates the corresponding step synchronously.

The right details panel has two modes, automatically switching based on whether a DAG node is selected:

• Step Detail Mode: Displays the memory operation list (Task Memory Ops) and Buffer slice details (Buffer Layout) for the current step. Each slice in Buffer Layout can be expanded to view data source semantics, with click navigation support.
• Node Detail Mode: Displays node attributes, parent/child node relationships, cross-stage node mapping, and memory operations associated with the node.

Other Features:

• Bottom Player: Supports browsing by step order, with a draggable slider for quick jumping. Playback speed is 1 second/step.
• Focus View: When a step is selected, the DAG automatically enters a partial view showing only nodes related to the current step and their context. Suitable for narrowing down when nodes are dense.
• Subgraph Browsing: When a node contains a subgraph, clickOpen Subgraphon the right to view a finer-grained task structure.
• Stage Switching: Switch graph stages via the top dropdown (e.g.,input_graph/input_task_queues).
• Search: Supports searching DAG nodes by taskId / taskType / notifyId. Clicking a result locates the node.

4.4 Analytic Page

Analytic centrally displays all errors found during verification. The page is divided into three areas:

• Left: Dataset and rank selection.
• Center: Error list, each showing title, severity label (color-coded), key field summary, and original error code.
• Right: Details panel for the selected error, including basic error information, involved ranks, associated node information, slice details (precise location of data inconsistencies), supplementary information, and raw JSON.

The top of the details panel provides navigation buttons:

• View memView: Navigate to MemView, locating the memory step associated with the error.
• View dagView: Navigate to MemView, directly locating the DAG node associated with the error.

Navigation automatically carries the dataset, rank, and node location information, so no re-selection is needed after arriving at MemView.

5. Common Usage Flows

5.1 Starting Analysis from Dashboard

1. Click the target dataset in the dataset list.
2. Select the ranks to view in the Rank tree on the left.
3. Confirm the dataset's file count and error statistics on the right.
4. ClickEnter Correlation Viewto enter MemView.

5.2 Locating Memory Issues in MemView

1. Click the target step on the timeline (or use the bottom player to jump).
2. View the step's memory operations in the rightTask Memory Opssection.
3. Select a rank and buffer type inBuffer Layout, expand a slice to view the data source.
4. ClickLocate Taskor the semantic card to jump to the corresponding DAG node.

5.3 Reverse Tracing Memory from a DAG Node

1. Click the target node in the DAG.
2. View the node'sMemory Opson the right.
3. ClickJump to Stepto return to the timeline, or click parent/child nodes to continue tracing along the dependency chain.

5.4 Quick Location from Errors

1. Click an error in the Analytic page or the error list on the left of MemView.
2. View the associated node and slice information in the details.
3. ClickView memVieworView dagViewto jump to the problem site.

6. Usage Tips

• First confirm the dataset and rank range in Dashboard, then enter MemView for analysis.
• When investigating memory issues, useTask Memory OpsandBuffer Layouttogether.
• When investigating node dependency relationships, first select the node in the DAG, then use the parent/child relationship and node mapping on the right to navigate.
• When DAG nodes are too dense, first use Focus View to narrow down, then expand gradually.
• The error list on the left of MemView allows quick browsing and error node location without leaving the current page.

7. Terminology

【免费下载链接】hcommHCOMM（Huawei Communication）是HCCL的通信基础库，提供通信域以及通信资源的管理能力。项目地址: https://gitcode.com/cann/hcomm