Overview

ℹ️ Introduction

MindStudio Monitor (msMonitor) is an online performance monitoring and dynamic collection tool for Ascend cluster scenarios. Built on dynolog (Meta CPU-GPU monitoring system) and msPTI (MindStudio Profiler Tools Interface), it supports capabilities such as npu-monitor, nputrace, and Monitor API.

Supported framework Profiler tools: Ascend PyTorch Profiler and MindSpore Profiler

As shown in the preceding figure, the core components of msMonitor are described as follows:

Component	Purpose	Documentation
Dynolog daemon	Server-side daemon process, responsible for receiving dyno requests and triggering monitoring and collection.	dynolog
Dyno CLI	Client-side command-line entry for issuing npu-monitor and nputrace commands.	dyno
msPTI Monitor	msPTI-based collection module, responsible for obtaining and reporting performance data.	-

Product Support

[!NOTE]

For specific Ascend product models, see Ascend Product Models.

Product Type	Whether Supported
Atlas 350 Accelerator Card	√
Atlas A3 Training Series/Atlas A3 Inference Series	√
Atlas A2 Training Series/Atlas A2 Inference Series	√
Atlas 200I/500 A2 Inference Product	√
Atlas Inference Series	×
Atlas Training Series	×

⚙️ Features

msMonitor provides the following core features:

Feature Name	Feature Description	Documentation
npu-monitor	A lightweight resident daemon that continuously monitors the latency of key operators, suitable for online observation of performance fluctuations.	npu-monitor
nputrace	Dynamically triggers performance data collection and parsing on the framework, CANN, and Device sides without interrupting running tasks.	nputrace
Monitor API	Provides Python APIs for collecting performance data such as compute operators, communication operators, APIs, Runtime APIs, and Mstx.	Monitor API

[!NOTE]

Due to underlying resource limitations, npu-monitor and nputrace cannot be enabled simultaneously.