HPC Render

A Blender extension to render frames and animations on a high performance compute cluster using SLURM.

Setup

• SSH Host: This is the thing you write when you SSH into your cluster.
• Blender path (cluster): This is the absolute path to your Blender installation on the cluster.
• Remote project dir: This is where you want your Blender project (and render outputs) to go on the supercomputer when they're being rendered. It will copy the Blender file and render to ./renders when it starts rendering.
• CUDA module: This is the CUDA module installation that you load. To find available installations on your cluster, SSH into your supercomputer cluster and run module avail cuda. Then, select one that looks reasonable from the list.
• SLURM partition: This is the partition of the node that will be called to run your project. You can find available partitions with sinfo -s, but verify that your selected partition has GPUs if needed.

Usage

Note that you can't render a single frame across multiple nodes, since that would be very inefficient. Instead, the multi-node setup uses a queue system to render frames of an animation on the different nodes by opening a Blender instance to render a batch of frames from the queue.

• Nodes: The number of nodes you want to request. It's usually best to set this to 1, unless you are rendering a whole bunch of stuff. For each node you add, it multiplies the following parameters (e.g., 2 nodes with 4 GPUs would be 8 GPUs total).
• Batch size: When more than one node is requested, the batch size determines how many frames a given Blender instance will render from the animation frame queue. For optimal performance, the batch size should be the number of frames divided by the number of nodes.
• GPUs: The number of GPUs you want on a node.
• Tasks: The number of parallel threads you want on a node.
• CPU cores per task: The number of cores on a given thread of a given node.
• Memory (GB): The ammount of memory (in gigabytes) allocated for a given node.
• Time Limit: The time limit you want to set on the render. You can find the maximum time limit for a given partition or configuration by running sinfo -o "%P %l".

You can verify your configuration by consulting the available configurations on your cluster using sinfo -N -l. When rendering using many GPUs, it is ideal to reduce tile size to improve load balancing performance. This can be done in Render Properties > Performance > Memory.

Benchmarks

Test Scene:

• 30 frames, 3 nodes, 4 GPUs per node: 55 seconds. GPU unknown (probably teslas or something)
• 30 frames, local: 1 hr 57 mins. Single RTX2060 Super