File systems

The Denvr Cloud is pre-integrated with network-attached file storage systems to preserve data sets, model checkpoints, and system output.

High Performance Storage

Features:

• Petabyte-scale network-attached Flash SSD with configurable data protection

• Up to 5 GB/s read throughput

• Encrypted at rest and delivered to GPUs over tenant-private networking

Primary use:

• Working space for very large datasets that are copied in and out for local Cache

• Application file systems, databases, and large capacity/high IOPS requirements

• Faster version of User Storage but is typically used for pre-processing and streaming large datasets in and out of Cache

User Storage

Features:

• Petabyte-scale RAID-protected storage, optimized for low cost, parallel I/O, and persistent storage

• Encrypted at rest and delivered to GPUs over tenant-private networking

• Up to 1 GB/s of read throughput

Primary use:

• Used for home directory files, application code and configuration, and datasets

• Clients may optionally prefer to keep data in their own on-prem storage systems and leverage Denvr for caching

• GPUs can process data directly from User Storage, but may benefit from Cache storage depending on your use

Cache Storage

Features:

• Up to 16 TB of direct-attached NVMe SSD per bare metal node

• Lowest latency and up to 12 GB/s read throughput

• Cache is non-persistent and is freed when applications are stopped

Primary uses:

• Volumes provide local storage for model training data, checkpoint files, and intermediate results data

• Data loaders should copy training data into the Cache Storage to ensure full GPU utilization and not bottleneck performance on slower network-attached storage