Choosing a storage type
Overview
ECI offers three storage types. Picking the right type for the use case lets you optimize both performance and cost.
Comparison
| Item | Block storage | Object storage | Parallel file system (PFS) |
|---|---|---|---|
| Access method | Mounted to a VM as a disk | HTTP API (S3-compatible) | Mounted simultaneously by many VMs |
| Concurrent access | One VM at a time | Unlimited (API) | Multiple VMs read and write at once |
| Performance | Fast (local-disk-class) | Slow (API latency) | Very fast (dedicated VLAN, SR-IOV) |
| Primary use | OS disks, training checkpoints | Dataset storage, backup | Shared storage for distributed training |
| DR (disaster recovery) | Supported (separate price) | Not supported | Not supported |
| After VM deletion | Data preserved | Data preserved | Data preserved |
How to choose
Storing OS and data on a single VM → block storage
Mounts to the VM like a disk. Good for OS disks and training checkpoints.
Storing or sharing large datasets → object storage
Accessed via an S3-compatible API (aws-cli, rclone). Good for sharing the same data across multiple VMs or exchanging data with external systems.
Shared storage for multi-node distributed training → parallel file system
Multiple VMs can read and write the same files simultaneously. Used in large-scale distributed training for checkpoint and log sharing.
Next steps
- Block storage: persistent disks attached directly to a VM
- Object storage: S3-compatible bucket storage
- Parallel file system: shared storage for multi-node training