Analysis Process

The cluster performance is affected by multiple factors, such as the AI processor type, network, communication algorithm, and communication configuration. For performance problems, the Profiling tool is used for performance analysis. The main analysis process is as follows:

1. Collect full profiling data. For details, see Profiling Data Collection.
2. Identify the performance bottleneck of the entire network and perform further analysis and optimization based on the dispatch and execution phases of communication operators. For details, see Profiling Data Analysis.

This section focuses on the identification of HCCL-related profiling information and the analysis of common cases. For more performance tuning cases, see Communication Tuning Solutions in General Performance Issue Troubleshooting Guide. After collecting full profiling data, analyze it by referring to MindStudio Insight User Guide.