Clone this repo on an Nvidia GPU equipped machine with at least 2 GPUs (g2.8xlarge instances will do nicely).
Then run:
./build.sh
./run.sh
./deploy-tasks.sh
This will launch 2 marathon applications. The first application runs
nvidia-smi every 60 seconds without a docker container, and the second one
runs nvidia-smi every 60 seconds inside an nvidia/cuda container.
1. Set up port forwarding in case your are running docker on a remote machine and want to browse all of the output locally:
ssh -NT -L 8080:localhost:8080 -L 5050:localhost:5050 -L 5051:localhost:5051 <remote-ip>
2. Navigate to marathon UI:
http://<your-ip>:8080
Verify that the tasks are running.
3. Go to the Mesos UI and look at the stdout of the tasks:
http://<your-ip>:5050
4. If they say the following, you've got problems:
Failed to initialize NVML: Unknown Error
If they print something, like the following, you are good to go:
+------------------------------------------------------+
| NVIDIA-SMI 352.79 Driver Version: 352.79 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla M60 Off | 0000:04:00.0 Off | 0 |
| N/A 31C P8 14W / 150W | 34MiB / 7679MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+