SymetryML supports GPU and Multi-GPU projects for hardware-accelerated machine learning on wide datasets. The SymetryML Docker image includes all required CUDA libraries and native libraries pre-installed — no manual GPU software setup is needed inside the container.
To use GPU acceleration, the host machine must have NVIDIA drivers and the NVIDIA Container Toolkit installed.
Prerequisites
A working SymetryML Docker installation. Refer to the Installation Guide for setup instructions.
A SymetryML license that allows GPU or Multi-GPU projects.
NVIDIA GPU with Compute Capability >= 3.5 on the host.
NVIDIA drivers installed on the host.
NVIDIA Container Toolkit installed on the host.
Host Setup
Step 1 — Verify NVIDIA Drivers
Run nvidia-smi on the host to confirm that NVIDIA drivers are installed and your GPU is detected:
$ nvidia-smi
Fri May 26 11:32:04 2023
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.47.03 Driver Version: 510.47.03 CUDA Version: 11.6 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla V100-SXM2... On | 00000000:00:1E.0 Off | 0 |
| N/A 41C P0 27W / 300W | 0MiB / 16384MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
If nvidia-smi is not found or does not show your GPU, install the appropriate NVIDIA drivers for your host OS before continuing.
To run SymetryML with GPU support, use the runtime: nvidia option and GPU environment variables in your docker-compose.yml. Refer to the GPU Support section of the Installation Guide for the full Docker Compose example.
The key settings are:
Setting
Value
Description
runtime
nvidia
Enables GPU access in the container
NVIDIA_VISIBLE_DEVICES
all
Exposes all host GPUs to the container
LD_LIBRARY_PATH
/usr/local/cuda/lib64:/opt/symetry/nativelib
CUDA and native library paths
SymetryML Memory Requirements
Please consult the Technical Requirements for more information on memory requirements for various project sizes. Note that with SymetryML, project size is determined by the number of attributes, not the number of rows.
SymetryML Configuration and GPU
The symetry-rest.txt configuration file contains properties that control SymetryML GPU behavior. The following table lists the properties relevant to using SymetryML with NVIDIA GPUs.
Property
Description
rtlm.option.rtlm.gpu.matrix.minsize
Minimum matrix size to use GPU. Matrix operations like multiplication, inversion, etc are used when SymetryML builds models. Other operation like PCA and SVD also can leverage GPU. Recommended values : 512
rtlm.option.rtlm.gpu.update.minsize
Minimum size to use GPU when updating a SymetryML project. Recommended values:64 to 128
rtlm.mgpu.num.gpus
The maximum number of GPUs that can be used on a server in a MultiGPU project.
rtlm.mgpu.runon.one
Enabling that specifies that a MultiGPU project can run on a server with only 1 GPU. '1' enables and '0' disables. Default: '0'
