nvidia-smi（英伟达 SMI）

Skill: Deep Dive into 'nvidia-smi'

Name: nvidia-smi-mastery Description: Provides a comprehensive guide to using the nvidia-smi command for monitoring, managing, and troubleshooting NVIDIA GPU devices, from basic status checks to advanced scripting. Keywords: ["nvidia-smi", "gpu monitoring", "nvml", "cuda", "deep learning", "gpu troubleshooting", "system administration"]

Deep Dive into 'nvidia-smi'

Objective To transform users from basic nvidia-smi observers into power users capable of advanced GPU diagnostics, performance tuning, and automated monitoring.

Core Concept: The GPU's Dashboard

nvidia-smi (NVIDIA System Management Interface) is the primary command-line utility for monitoring and managing NVIDIA GPUs. It acts as a "dashboard," providing real-time data on your GPU's health and performance.

• Under the Hood: nvidia-smi is a client for the NVIDIA Management Library (NVML), a C library that provides the interface to the GPU driver. This means you can also use NVML bindings in languages like Python (pynvml) to build custom monitoring tools.
• Primary Use Cases:

- Monitoring GPU utilization and temperature during model training. - Diagnosing "CUDA out of memory" (OOM) errors by identifying processes consuming VRAM. - Troubleshooting performance bottlenecks like thermal throttling. - Managing multi-GPU environments.

The Output: A Field Guide

Running nvidia-smi provides a table of information. Here's how to interpret the key fields:

• GPU: The ID of the GPU (e.g., 0, 1) in a multi-GPU system.
• Fan, Temp, Perf: Fan speed (%), GPU temperature (°C), and performance state (P0 is max performance, P12 is idle).
• Pwr:Usage/Cap: Current power draw versus the board's power limit (TDP).
• Memory-Usage: The most critical field for many. It shows Used / Total VRAM.
• GPU-Util: The percentage of time the GPU's compute cores were busy. Note the distinction between high VRAM usage and high GPU utilization.
• Processes: A list of PIDs and process names currently using the GPU, along with their individual memory consumption. This is invaluable for finding "zombie" processes.

Essential Commands and Workflows

Basic Monitoring

• The Standard View: nvidia-smi
• Continuous Monitoring: Instead of a one-time snapshot, use watch or the built-in loop option for real-time updates.

- watch -n 1 nvidia-smi (Updates every 1 second) - nvidia-smi -l 1 (Native loop, updates every 1 second)

Targeted Queries

• List GPUs: nvidia-smi -L provides a clean list of all detected GPUs and their UUIDs.
• Focus on a Single GPU: nvidia-smi -i 0 restricts the view to GPU 0.
• Get Detailed Info: nvidia-smi -i 0 -q gives a verbose, detailed report for GPU 0, including clock speeds, temperature limits, and ECC status.
• Filter Information: Use the -d flag to display only specific categories.

- nvidia-smi -q -d MEMORY,POWER shows only memory and power data.

Process Management

• Find the Culprit: When you get an OOM error, the main table's "Processes" section is your first stop.
• Clean Output for Scripting: nvidia-smi --query-compute-apps=pid,process_name,used_memory --format=csv provides a clean CSV output of all compute processes, which is perfect for parsing in scripts.

Advanced Diagnostics and Troubleshooting

Scenario 1: Diagnosing "CUDA out of memory"

1. Check Baseline: Run nvidia-smi before starting your task to see how much VRAM is already occupied by other processes (e.g., a display manager).
2. Identify Process: If an OOM occurs, use the process list to find the PID of the job that consumed all the memory.
3. Kill the Process: Use kill <PID> to terminate the process and free up the VRAM.

Scenario 2: Troubleshooting Performance Drops If your training loop suddenly slows down, use continuous monitoring (watch -n 0.5 nvidia-smi) and look for:

• Thermal Throttling: A sudden spike in Temp followed by a change in Perf state (e.g., from P2 to P8) indicates the GPU is reducing its clock speed to cool down.
• Power Capping: If Pwr:Usage is constantly hitting the Cap, the GPU might be limited by its power budget.

Daemon Mode (dmon) For a more compact, columnar output ideal for long-term monitoring, use the daemon mode.

• nvidia-smi dmon -s pumt -d 1
• -s pumt: Specifies the metrics to show: power, utilization, memory, temperature.
• -d 1: Sets the update interval to 1 second.

Automation and Scripting

You can use nvidia-smi in shell scripts or Python to automate monitoring and logging.

Python Scripting with pynvml Since nvidia-smi is a wrapper for NVML, you can use the pynvml library for more direct and faster access in Python.

import pynvml
import time

pynvml.nvmlInit()
device_count = pynvml.nvmlDeviceGetCount()

print(f"Found {device_count} GPU(s)")

try:
    while True:
        for i in range(device_count):
            handle = pynvml.nvmlDeviceGetHandleByIndex(i)
            name = pynvml.nvmlDeviceGetName(handle).decode('utf-8')
            temp = pynvml.nvmlDeviceGetTemperature(handle, pynvml.NVML_TEMPERATURE_GPU)
            power = pynvml.nvmlDeviceGetPowerUsage(handle) / 1000.0  # Convert mW to W
            mem_info = pynvml.nvmlDeviceGetMemoryInfo(handle)
            util = pynvml.nvmlDeviceGetUtilizationRates(handle)

            print(f"[{i}] {name}")
            print(f"  Temp: {temp}°C")
            print(f"  Power: {power:.2f}W")
            print(f"  Memory: {mem_info.used // 1024**2} / {mem_info.total // 1024**2} MB")
            print(f"  GPU Util: {util.gpu}%")
            print("-" * 20)
        time.sleep(2)
except KeyboardInterrupt:
    print("\
Monitoring stopped.")
    pynvml.nvmlShutdown()