cloudscaleKnowledge Base

The scratch disk is located directly in the same physical host where the virtual GPU server runs, ensuring minimal latency and maximum performance. This contrasts with standard NVMe SSD and Bulk volumes, which are located in our storage clusters and are therefore accessible from all compute hosts, but have noticeably higher latency due to being accessed over the network.

When scaling GPU servers or during maintenance work by us, the contents of the scratch disk – along with the associated, powered-off GPU server – can be migrated to other hardware, which does, however, take some time. Otherwise, the scratch disk is tied to the respective GPU server and cannot be detached from it, attached to another server, or retained independently of the server.

However, it is possible to connect a standard NVMe SSD volume to a GPU server. If the GPU server is deleted, this volume can remain independent in the respective project and be attached to another or future server. Depending on the use case or the required disk performance, you may then choose to work directly on this volume. If the minimal latency of a scratch disk is required during operation, its contents can be backed up to the standard NVMe SSD volume before the GPU server is deleted, and a new scratch disk can be pre-populated from this volume in the future (even in an automated manner, e.g., via cloud-init).

Important: Even during normal operation, you should expect that the contents of the scratch disk may no longer be available after a technical issue (and that, for example, a new, empty scratch disk will be provisioned on replacement hardware). Depending on the use case, it may therefore be advisable to continuously back up the contents of the scratch disk to another storage location.