Vulkan is a low-level graphics API used to program the GPU to perform graphics rendering and other tasks. It is a modern, high-performance replacement of the legacy OpenGL API. It is supported by all desktop and mobile GPU vendors, and support is included in all recent graphics drivers, although you may need to install the small Vulkan API loader manually.
Vulkan provides the following improvements over OpenGL in the backend renderer:
Vulkan is designed specifically for speed. The Vulkan renderer is often 2 times faster than OpenGL, and performance will improve in the future.
OpenGL does not provide direct access to GPU memory, making it difficult to manage memory utilization. Vulkan requires the application to manage GPU memory directly, and you can use the same memory for multiple, nonoverlapping tasks, which allows more work with less memory. Interaction with Cuda is more direct, because Cuda can reference and use the same memory as Vulkan with virtually no overhead.
Vulkan is more multi-thread friendly than OpenGL. It is designed to give the application far more control of task scheduling on the GPU. This will enable a more efficient end-to-end rendering process by providing opportunities to overlap multiple render requests at different stages.
HEAVY.AI recommends using the Vulkan renderer unless you have hit a blocking issue with Vulkan. Although the Vulkan driver is reliable and stable, it has not yet had wide external usage compared to its OpenGL counterpart and may be prone to unforeseen configuration issues; see Troubleshooting below for guidance on known issues. Having Vulkan enabled helps us to find and address additional issues while a reliable fallback exists to ensure a smooth, less risky transition.
Follow these steps if you receive error messages when using Vulkan.
You might see the following error when trying to start a rendering-enabled server for the first time:
error while loading shared libraries: libvulkan.so.1: cannot open shared object file: No such file or directory
If you receive this error, you need to install the Vulkan API loader:
CentOS: sudo yum install vulkan
Ubuntu: sudo apt install libvulkan1 --no-install-recommends
You might experience a server crash or get errors on startup like the following:
Oct 21 17:09:33 ip-xxx-xxx-xx-xx.us-east-2.compute.internal omnisci_server[61162]: LLVM ERROR: inconsistency in registered CommandLine option
If this occurs, HEAVY.AI recommends setting the VK_ICD_FILENAMES to the Nvidia Driver Manifest variable before starting the server.
Ubuntu/Debian: export VK_ICD_FILENAMES=$(ls /usr/share/vulkan/icd.d/nvidia*)
CentOS/Redhat: export VK_ICD_FILENAMES=$(ls /etc/vulkan/icd.d/nvidia*)
Although /usr/share/vulkan/icd.d and /etc/vulkan/icd.d are the most commonly used paths, the Vulkan loader looks in all the following locations:
/etc/vulkan/icd.d
/usr/local/etc/vulkan/icd.d
/usr/local/share/vulkan/icd.d
/usr/share/vulkan/icd.d
$HOME/.local/share/vulkan/icd.d
If the server has been installed as a service, check the name of the Nvidia icd json file with the following command:
Ubuntu/Debian: ls /usr/share/vulkan/icd.d/nvidia*
CentOS/Redhat: ls /etc/vulkan/icd.d/nvidia*
Then, add the output as an Environment to the HEAVY.AI server Service section with the following command:
sudo vi /usr/lib/systemd/system/heavydb.service
Ubuntu/Debian
[Service]
Environment="VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json"
CentOS/Redhat
[Service]
Environment="VK_ICD_FILENAMES=/etc/vulkan/icd.d/nvidia_icd.json"
Then, run sudo systemctl daemon-reload.
As a last resort, you can remove the package mesa-vulkan-driver from the OS.
CentOS: sudo yum remove mesa-vulkan-drivers
Ubuntu: sudo apt remove mesa-vulkan-drivers
For a summary of how to install the Vulkan loader on various Linux distributions, see https://linuxconfig.org/install-and-test-vulkan-on-linux.
If you run the server in a Docker container and you see an error similar to the following in your logs, even with the proper drivers installed, your NVIDIA Docker runtime may be out of date:
Backend rendering disabled: Error creating Vulkan instance: Vulkan Error: VK_ERROR_INCOMPATIBLE_DRIVER
To resolve, try upgrading the nvidia-docker2 package on your system.