GPU usage monitoring (CUDA)

For Intel GPU’s there exists the intel-gpu-tools from http://intellinuxgraphics.org/ project, which brings the command intel_gpu_top (amongst other things). It is similar to top and htop, but specifically for the Intel GPU.

   render busy:  18%: ███▋                                   render space: 39/131072
bitstream busy:   0%:                                     bitstream space: 0/131072
  blitter busy:  28%: █████▋                                blitter space: 28/131072

          task  percent busy
           GAM:  33%: ██████▋                 vert fetch: 0 (0/sec)
          GAFS:   3%: ▋                       prim fetch: 0 (0/sec)
            VS:   0%:                      VS invocations: 559188 (150/sec)
            SF:   0%:                      GS invocations: 0 (0/sec)
            VF:   0%:                           GS prims: 0 (0/sec)
            DS:   0%:                      CL invocations: 186396 (50/sec)
            CL:   0%:                           CL prims: 186396 (50/sec)
           SOL:   0%:                      PS invocations: 8191776208 (38576436/sec)
            GS:   0%:                      PS depth pass: 8158502721 (38487525/sec)
            HS:   0%:                      
            TE:   0%:                      
          GAFM:   0%:                      
           SVG:   0%:                      

For OS X

Including Mountain Lion

iStat Menus

Excluding Mountain Lion

atMonitor

The last version of atMonitor to support GPU related features is atMonitor 2.7.1.

– and the link to 2.7.1 delivers 2.7b.

For the more recent version of the app, atMonitor – FAQ explains:

To make atMonitor compatible with MacOS 10.8 we have removed all GPU related features.

I experimented with 2.7b a.k.a. 2.7.1 on Mountain Lion with a MacBookPro5,2 with NVIDIA GeForce 9600M GT. The app ran for a few seconds before quitting, it showed temperature but not usage:

For linux, use nvidia-smi -l 1 will continually give you the gpu usage info, with in refresh interval of 1 second.

For completeness, AMD has two options:

1. fglrx (closed source drivers).

    $ aticonfig --odgc --odgt
   

2. mesa (open source drivers), you can use RadeonTop.

    $ sudo apt-get -y install radeontop; radeontop
   

View your GPU utilization, both for the total activity percent and individual blocks.

for nvidia on linux i use the following python script which uses an optional delay and repeat like iostat and vmstat

https://gist.github.com/matpalm/9c0c7c6a6f3681a0d39d

$ gpu_stat.py 1 2
{"util":{"PCIe":"0", "memory":"10", "video":"0", "graphics":"11"}, "used_mem":"161", "time": 1424839016}
{"util":{"PCIe":"0", "memory":"10", "video":"0", "graphics":"9"}, "used_mem":"161", "time":1424839018}

nvidia-smi does not work on some linux machines (returns N/A for many properties). You can use nvidia-settings instead (this is also what mat kelcey used in his python script).

nvidia-settings -q GPUUtilization -q useddedicatedgpumemory

You can also use:

watch -n0.1 "nvidia-settings -q GPUUtilization -q useddedicatedgpumemory"

for continuous monitoring.

Recently I have written a simple command-line utility called gpustat (which is a wrapper of nvidia-smi) : please take a look at https://github.com/wookayin/gpustat.

For Linux, I use this HTOP like tool that I wrote myself. It monitors and gives an overview of the GPU temperature as well as the core / VRAM / PCI-E & memory bus usage. It does not monitor what’s running on the GPU though.

gmonitor

I have a GeForce 1060 GTX video card and I found that the following command give me info about card utilization, temperature, fan speed and power consumption:

$ nvidia-smi --format=csv --query-gpu=power.draw,utilization.gpu,fan.speed,temperature.gpu

You can see list of all query options with:

$ nvidia-smi --help-query-gpu

I have had processes terminate (probably killed or crashed) and continue to use resources, but were not listed in nvidia-smi. Usually these processes were just taking gpu memory.

If you think you have a process using resources on a GPU and it is not being shown in nvidia-smi, you can try running this command to double check. It will show you which processes are using your GPUs.

sudo fuser -v /dev/nvidia*

This works on EL7, Ubuntu or other distributions might have their nvidia devices listed under another name/location.

The following function appends information such as PID, user name, CPU usage, memory usage, GPU memory usage, program arguments and run time of processes that are being run on the GPU, to the output of nvidia-smi:

function better-nvidia-smi () {
    nvidia-smi
    join -1 1 -2 3 
        <(nvidia-smi --query-compute-apps=pid,used_memory 
                     --format=csv 
          | sed "s/ //g" | sed "s/,/ /g" 
          | awk 'NR<=1 {print toupper($0)} NR>1 {print $0}' 
          | sed "/[NotSupported]/d" 
          | awk 'NR<=1{print $0;next}{print $0| "sort -k1"}') 
        <(ps -a -o user,pgrp,pid,pcpu,pmem,time,command 
          | awk 'NR<=1{print $0;next}{print $0| "sort -k3"}') 
        | column -t
}

Example output:

$ better-nvidia-smi
Fri Sep 29 16:52:58 2017
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 378.13                 Driver Version: 378.13                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GT 730      Off  | 0000:01:00.0     N/A |                  N/A |
| 32%   49C    P8    N/A /  N/A |    872MiB /   976MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+
|   1  Graphics Device     Off  | 0000:06:00.0     Off |                  N/A |
| 23%   35C    P8    17W / 250W |    199MiB / 11172MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0                  Not Supported                                         |
|    1      5113    C   python                                         187MiB |
+-----------------------------------------------------------------------------+
PID   USED_GPU_MEMORY[MIB]  USER    PGRP  %CPU  %MEM   TIME      COMMAND
9178  187MiB                tmborn  9175  129   2.6    04:32:19  ../path/to/python script.py args 42

Glances has a plugin which shows GPU utilization and memory usage.

http://glances.readthedocs.io/en/stable/aoa/gpu.html

Uses the nvidia-ml-py3 library: https://pypi.python.org/pypi/nvidia-ml-py3

You can use nvtop, it’s similar to htop but for NVIDIA GPUs. Link: https://github.com/Syllo/nvtop

This script is more readable and is designed for easy mods and extensions.

You can replace gnome-terminal with your favorite terminal window program.

#! /bin/bash

if [ "$1" = "--guts" ]; then
    echo; echo "    ctrl-c to gracefully close"
    f "$a"
    f "$b"
    exit 0; fi

# easy to customize here using "nvidia-smi --help-query-gpu" as a guide
a='--query-gpu=pstate,memory.used,utilization.memory,utilization.gpu,encoder.stats.sessionCount'
b='--query-gpu=encoder.stats.averageFps,encoder.stats.averageLatency,temperature.gpu,power.draw'
p=0.5    # refresh period in seconds
s=110x9  # view port as width_in_chars x line_count

c="s/^/    /; s/, +/t/g"
t="`echo '' |tr 'n' 't'`"
function f() { echo; nvidia-smi --format=csv "$1" |sed -r "$c" |column -t "-s$t" "-o   "; }
export c t a b; export -f f
gnome-terminal --hide-menubar --geometry=$s -- watch -t -n$p "`readlink -f "$0"`" --guts

#

License: GNU GPLv2, TranSeed Research

You can use

nvidia-smi pmon -i 0

to monitor every process in GPU 0. including compute/graphic mode, sm usage, memory usage, encoder usage, decoder usage.

I didn’t see it in the available answers (except maybe in a comment), so I thought I’d add that you can get a nicer refreshing nvidia-smi with watch. This refreshes the screen with each update rather than scrolling constantly.

watch -n 1 nvidia-smi

for one second interval updates. Replace the 1 with whatever you want, including fractional seconds:

watch -n 5 nvidia-smi
watch -n 0.1 nvidia-smi

Recently, I have written a monitoring tool called nvitop, the interactive NVIDIA-GPU process viewer.

It is written in pure Python and is easy to install.

Install from PyPI:

pip3 install --upgrade nvitop

Install the latest version from GitHub:

pip3 install git+https://github.com/XuehaiPan/nvitop.git#egg=nvitop

Run as a resource monitor:

nvitop

nvitop will show the GPU status like nvidia-smi but with additional fancy bars and history graphs.

For the processes, it will use psutil to collect process information and display the USER, %CPU, %MEM, TIME and COMMAND fields, which is much more detailed than nvidia-smi. Besides, it is responsive for user inputs in monitor mode. You can interrupt or kill your processes on the GPUs.

nvitop comes with a tree-view screen and an environment screen:

In addition, nvitop can be integrated into other applications. For example, integrate into PyTorch training code:

import os
from nvitop.core import host, CudaDevice, HostProcess, GpuProcess
from torch.utils.tensorboard import SummaryWriter

device = CudaDevice(0)
this_process = GpuProcess(os.getpid(), device)
writer = SummaryWriter()
for epoch in range(n_epochs):

    # some training code here
    # ...

    this_process.update_gpu_status()
    writer.add_scalars(
        'monitoring',
        {
            'device/memory_used': float(device.memory_used()) / (1 << 20),  # convert bytes to MiBs
            'device/memory_percent': device.memory_percent(),
            'device/memory_utilization': device.memory_utilization(),
            'device/gpu_utilization': device.gpu_utilization(),

            'host/cpu_percent': host.cpu_percent(),
            'host/memory_percent': host.virtual_memory().percent,

            'process/cpu_percent': this_process.cpu_percent(),
            'process/memory_percent': this_process.memory_percent(),
            'process/used_gpu_memory': float(this_process.gpu_memory()) / (1 << 20),  # convert bytes to MiBs
            'process/gpu_sm_utilization': this_process.gpu_sm_utilization(),
            'process/gpu_memory_utilization': this_process.gpu_memory_utilization(),
        },
        global_step
    )

See https://github.com/XuehaiPan/nvitop for more details.

Note: nvitop is dual-licensed by the GPLv3 License and Apache-2.0 License. Please feel free to use it as a dependency for your own projects. See Copyright Notice for more details.

you can use "GPU Dashboards in Jupyter Lab"

Introduction
NVDashboard is an open-source package for the real-time visualization of NVIDIA GPU metrics in interactive Jupyter Lab environments. NVDashboard is a great way for all GPU users to monitor system resources. However, it is especially valuable for users of RAPIDS, NVIDIA’s open-source suite of GPU-accelerated data-science software libraries.ref

To monitor GPU usage in real-time, you can use the nvidia-smi command with the –loop option on systems with NVIDIA GPUs. Open a terminal and run the following command:

nvidia-smi --query-gpu=timestamp,name,utilization.gpu,utilization.memory,memory.total,memory.free,memory.used --format=csv --loop=1

This command will display GPU usage information in real-time with a refresh interval of 1 second (you can change the interval by modifying the value after –loop=). The displayed information includes timestamp, GPU name, GPU utilization, memory utilization, total memory, free memory, and used memory.