Vertex AI Workbench is a managed Jupyter Notebook service of Google Cloud. It allows you to choose a wide range of configurations such as GPU types, disk size, and environment (which Docker image to compute on). The Docker image options include Kaggle Python ¹, so Workbench is one of the best (paid) alternatives when you run out of the GPU quota of the Kaggle Notebook.

Is Shared Memory Too Small?

However, in contrast to Kaggle Notebook’s 5.5 GB, Workbench provides only 64 MB of shared memory (shm) by default.

$ df -h
Filesystem      Size  Used Avail Use% Mounted on
overlay          99G   27G   68G  28% /
tmpfs            64M     0   64M   0% /dev
tmpfs           1.9G     0  1.9G   0% /sys/fs/cgroup
shm              64M     0   64M   0% /dev/shm
/dev/sdb         98G   12K   98G   1% /home/jupyter
/dev/sda1        99G   27G   68G  28% /etc/hosts
tmpfs           1.9G     0  1.9G   0% /proc/acpi
tmpfs           1.9G     0  1.9G   0% /sys/firmware

When you are using PyTorch, this often leads to fatal errors in DataLoader, such as:

DataLoader worker (pid xxx) is killed by signal: Bus error.

ERROR: Unexpected bus error encountered in worker. This might be caused by insufficient shared memory (shm)

Setting num_workers to zero solves these errors themselves, but it’s not a real solution because it sacrifices the speed of the DataLoder process.

The real solution is to execute the docker run command with either of the following two options:

• --shm-size=5.5gb (ref)
• --ipc=host (ref)

But can you do this in Workbench? The launching command docker run is hidden by GUI.

Solution: Specify in Metadata Pane

When you create a new notebook in Workbench GUI, you’ll see an optional pane for setting some metadata. In this pane, you can pass the option --shm-size=5.5gb or --ipc=host with the key container-custom-params, as shown in the screenshot below.

If you create a notebook with this metadata, you’ll get an instance with sufficient shm.

$ df -h
Filesystem      Size  Used Avail Use% Mounted on
overlay          99G   27G   68G  28% /
tmpfs            64M     0   64M   0% /dev
tmpfs           1.9G     0  1.9G   0% /sys/fs/cgroup
shm             5.5G     0  5.5G   0% /dev/shm
/dev/sdb         98G   12K   98G   1% /home/jupyter
/dev/sda1        99G   27G   68G  28% /etc/hosts
tmpfs           1.9G     0  1.9G   0% /proc/acpi
tmpfs           1.9G     0  1.9G   0% /sys/firmware