Analysis Process

Cluster performance is steered by multiple factors such as AI processor type, network, communication algorithm, and communication configuration. For performance issues, use profiling for analysis as follows:

1. Collect full profile data. For details, see Profile Data Collection.

2. Identify the bottleneck of the overall cluster performance, and perform further analysis and optimization based on the different stages of communication operator dispatch and execution. For details, see Profile Data Analysis.

This section focuses on HCCL-related profile data identification and analysis approaches for common cases. For more performance tuning cases, see "Solutions for TopN Performance Issues > Communication Tuning Solution" in General Performance Issue Troubleshooting Guide. After collecting full profile data, refer to MindStudio Insight Tool User Guide to analyze the profile data.