CLUSTER-SMI

The same as nvidia-smi but for multiple machines.

Run cluster-smi and the output should be something like

Additional information are available, when using

user@host $ cluster-smi -h

Usage of cluster-smi:
  -n string
        match node-names with regex for display information (if not specified, all nodes will be shown) (default ".")
  -p	verbose process information
  -t	show time of events

Monitoring Modes

This repository contains two versions: cluster-smi-local, cluster-smi.

Local (cluster-smi-local)

cluster-smi-local is the same as nvidia-smi but provides more verbose process information with the flag -p:

user@host $ cluster-smi-local -p

Thu Jan 18 21:44:51 2018
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+
| Node          | Gpu                   | Memory-Usage                  | GPU-Util | PID   | User     | Command       | GPU Mem   | Runtime                 |
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+
| node01        | 0:TITAN Xp            |  4477 MiB / 12189 MiB ( 36 %) |   39 %   |  5641 | john     | smokeparticle | 4465 MiB  |  1 d 21 h 58 min 19 sec |
|               | 1:TITAN Xp            |     0 MiB / 12189 MiB (  0 %) |    0 %   |       |          |               |           |                         |
|               | 2:TITAN Xp            |  4477 MiB / 12189 MiB ( 36 %) |   37 %   | 15963 | john     | smokeparticle | 4465 MiB  |  1 d 10 h 36 min 53 sec |
|               | 3:TITAN Xp            |  9930 MiB / 12189 MiB ( 81 %) |   94 %   | 10501 | john     | smokeparticle | 4465 MiB  |  1 d 19 h 30 min 27 sec |
|               |                       |                               |          | 10200 | jane     | caffe         | 5465 MiB  |  2 d 11 h 01 min  7 sec |
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+

Cluster (cluster-smi)

cluster-smi displays all information from cluster-smi-local but for multiple machines at the same time.

On each machine you want to monitor you need to start cluster-smi-node. They are sending information from the nvidia-driver to a cluster-smi-router, which further distributes these information to client (cluster-smi) when requested. Only the machines running cluster-smi-node require CUDA dependencies.

You might be interested as well in cluster-top for CPUS.

Installation

All steps below are used to test possible changes to this codebase. See the dockerfile and specific steps to compile this project in the provided files used by CI.

Requirements + Dependencies

• CUDA (just for cluster-smi-node.go)
• ZMQ (4.0.1)

Unfortunately, ZMQ can only be dynamically linked (libzmq.so) to this repository and you need to build it separately by

# compile ZMQ library for c++
cd /path/to/your_lib_folder
wget http:/files.patwie.com/mirror/zeromq-4.1.0-rc1.tar.gz
tar -xf zeromq-4.1.0-rc1.tar.gz
cd zeromq-4.1.0
./autogen.sh
./configure
./configure --prefix=/path/to/your_lib_folder/zeromq-4.1.0/dist
make
make install

Finally:

export PKG_CONFIG_PATH=/path/to/your_lib_folder/zeromq-4.1.0/dist/lib/pkgconfig/:$PKG_CONFIG_PATH
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/your_lib_folder/zeromq-4.1.0/dist/lib

Edit the CFLAGS, LDFLAGS in file nvvml/nvml.go to match your setup.

Compiling

You need to copy one config-file

user@host $ cp config.example.go config.go

To obtain a portable small binary, I suggest to directly embed the configuration settings (ports, ip-addr) into the binary as compile-time constants. This way, the app is fully self-contained (excl. libzmq.so) and does not require any configuration-files. This can be done by editing config.go:

...
c.RouterIp = "127.0.0.1"
c.Tick = 3
c.Timeout = 180
c.Ports.Nodes = "9080"
c.Ports.Clients = "9081"
...

Otherwise, you can specify the environment variable CLUSTER_SMI_CONFIG_PATH pointing to a yaml file (example in cluster-smi.example.yml).

Then run

cd proc
go install
cd ..
make all

Run

1. start cluster-smi-node at different machines having GPUs
2. start cluster-smi-router at a specific machine (machine with ip-addr: cluster_smi_router_ip)
3. use cluster-smi like nvidia-smi

Make sure, the machines can communicate using the specifiec ports (e.g., ufw allow 9080, 9081)

Use systemd

To ease the use of this app, I suggest to add the cluster-smi-node into a systemd-service. An example config file can be found here. The steps would be

# add new entry to systemd
sudo cp docs/cluster-smi-node.example.service /etc/systemd/system/cluster-smi-node.service
# edit the path to cluster-smi-node
sudo nano /etc/systemd/system/cluster-smi-node.service
# make sure you can start and stop the service (have a look at you cluster-smi client)
sudo service cluster-smi-node start
sudo service cluster-smi-node stop
# register cluster-smi-node to start on reboots
sudo systemctl enable cluster-smi-node.service

# last, start the service
sudo service cluster-smi-node start

Show all information

user@host $ cluster-smi -p -t

Thu Jan 18 21:44:51 2018
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+--------------------------+
| Node          | Gpu                   | Memory-Usage                  | GPU-Util | PID   | User     | Command       | GPU Mem   | Runtime                 | Last Seen                |
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+--------------------------+
| node00        | 0:TITAN Xp            |     0 MiB / 12189 MiB (  0 %) |    0 %   |       |          |               |           |                         | Thu Jan 18 21:44:49 2018 |
|               | 1:TITAN Xp            |     0 MiB / 12189 MiB (  0 %) |    0 %   |       |          |               |           |                         |                          |
|               | 2:TITAN Xp            |     0 MiB / 12189 MiB (  0 %) |    0 %   |       |          |               |           |                         |                          |
|               | 3:TITAN Xp            |     0 MiB / 12189 MiB (  0 %) |    0 %   |       |          |               |           |                         |                          |
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+--------------------------+
| node01        | 0:TITAN Xp            |  4477 MiB / 12189 MiB ( 36 %) |   39 %   |  5641 | john     | smokeparticle | 4465 MiB  |  1 d 21 h 58 min 19 sec | Thu Jan 18 21:44:50 2018 |
|               | 1:TITAN Xp            |     0 MiB / 12189 MiB (  0 %) |    0 %   |       |          |               |           |                         |                          |
|               | 2:TITAN Xp            |  4477 MiB / 12189 MiB ( 36 %) |   37 %   | 15963 | john     | smokeparticle | 4465 MiB  |  1 d 10 h 36 min 53 sec |                          |
|               | 3:TITAN Xp            |  9930 MiB / 12189 MiB ( 81 %) |   94 %   | 10501 | john     | smokeparticle | 4465 MiB  |  1 d 19 h 30 min 27 sec |                          |
|               |                       |                               |          | 10200 | jane     | caffe         | 5465 MiB  |  2 d 11 h 01 min  7 sec |                          |
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+--------------------------+
| node02        | 0:GeForce GTX 1080 Ti |  9352 MiB / 11172 MiB ( 83 %) |   61 %   |  9368 | doe      | python        | 2325 MiB  |  9 h 52 min  9 sec      | Thu Jan 18 21:44:49 2018 |
|               |                       |                               |          |  9434 | doe      | python        | 2339 MiB  |  9 h 51 min 48 sec      |                          |
|               |                       |                               |          |  9461 | doe      | python        | 2339 MiB  |  9 h 51 min 40 sec      |                          |
|               |                       |                               |          |  9503 | doe      | python        | 2339 MiB  |  9 h 51 min 31 sec      |                          |
|               | 1:GeForce GTX 1080 Ti |  9352 MiB / 11172 MiB ( 83 %) |   34 %   |  9621 | doe      | python        | 2339 MiB  |  9 h 49 min 13 sec      |                          |
|               |                       |                               |          |  9644 | doe      | python        | 2325 MiB  |  9 h 49 min  7 sec      |                          |
|               |                       |                               |          |  9670 | doe      | python        | 2339 MiB  |  9 h 49 min  1 sec      |                          |
|               |                       |                               |          |  9751 | doe      | python        | 2339 MiB  |  9 h 48 min 51 sec      |                          |
|               | 2:GeForce GTX 1080 Ti |  9366 MiB / 11172 MiB ( 83 %) |   15 %   |  9857 | doe      | python        | 2339 MiB  |  9 h 47 min 44 sec      |                          |
|               |                       |                               |          |  9868 | doe      | python        | 2339 MiB  |  9 h 47 min 37 sec      |                          |
|               |                       |                               |          |  9911 | doe      | python        | 2339 MiB  |  9 h 47 min 27 sec      |                          |
|               |                       |                               |          |  9983 | doe      | python        | 2339 MiB  |  9 h 47 min 19 sec      |                          |
|               | 3:GeForce GTX 1080 Ti |  9340 MiB / 11172 MiB ( 83 %) |   33 %   | 10144 | doe      | python        | 2325 MiB  |  9 h 43 min 30 sec      |                          |
|               |                       |                               |          | 10168 | doe      | python        | 2339 MiB  |  9 h 43 min 23 sec      |                          |
|               |                       |                               |          | 10192 | doe      | python        | 2339 MiB  |  9 h 43 min 17 sec      |                          |
|               |                       |                               |          | 10220 | doe      | python        | 2327 MiB  |  9 h 43 min 10 sec      |                          |
+---------------+-----------------------+-------------------------------+----------+-------+----------+---------------+-----------+-------------------------+--------------------------+

Explanation:

For flag -p

• Node is hostname of the machine
• Gpu lists all devices
• Memory-Usage lists used GPU memory
• Gpu-Util current GPU utilization
• PID process id of processes with active cuda context
• User owner (username) of PID
• Command command which is running
• Command command which is running
• GPU Mem use memory for only that particular process
• Runtime total time the process is already running

For flag -t

• Last Seen timestamp of last message of the machine status

For flag -n

• filter all nodes and only show those matching the given regEx

For flag -u

• filter all information and only show those related to given user

If the last message is older than 3 minutes, this list will print a "Timeout/Offline" in this line. Make sure the machine is running. Any new message will remove this note.