Skip to main content
GPFS is a shared file storage service for high-performance computing and AI workloads. You can mount a GPFS file system to CES instances or CKS cluster nodes to store and share datasets, model weights, training outputs, and log files. This document helps you decide whether to use GPFS and understand the storage types, capacity metering, and usage limits that you need to consider before creating a file system.

​
Use cases

GPFS is suitable for scenarios that require shared data access through file system semantics, such as:
  • Multiple CES instances or CKS cluster nodes need to access the same data.
  • Training jobs need to read datasets and write model weights, checkpoints, or training outputs.
  • Inference, batch processing, or data processing jobs need to share input files and output results.
If your workloads mainly read and write data through object APIs, OSS may be a better fit.

​
Storage types

GPFS provides two storage types: Performance and Capacity. If a Performance file system requires more than 500 GB/s of bandwidth, or if you need to increase the number of inodes, contact support for evaluation.

​
Storage capacity and billing

GPFS file systems support two billing methods:
  • Pay-as-you-go: billed based on the billing duration and provisioned capacity.
  • Subscription: prepaid for a fixed period.
The capacities in the GPFS storage type table are usable capacities. GPFS file systems may use replicas, erasure coding, or other mechanisms to protect data at the underlying layer. Because the underlying configuration may vary, the capacity and usage that you see on compute nodes by running df or du may not exactly match the usable capacity and usage displayed in the console. For example, in CES and CKS scenarios, the file system may use a replica configuration. If you create an 11.5 TiB file system, df may show approximately 23 TiB. After you write a 1 GiB file, du may show that the file occupies 2 GiB.

​
Quotas and limits

GPFS file systems use thick provisioning. You specify the capacity when you create a file system. Later scale-up or scale-down operations adjust the target capacity of the file system. The minimum capacity and scaling increment depend on the storage type. The same storage type may vary across regions and storage clusters.

​
Performance recommendations

For stable performance, we recommend that you keep file system usage below 95%. This usage includes:
  • Usable space usage.
  • Inode usage.
If file system usage is too high, I/O performance may decrease.

​
Data protection

GPFS protects data at the underlying layer by using replicas, erasure coding, or other mechanisms. To further improve data security, you still need to back up important data by yourself. After you delete a file system or delete data from a file system, the data cannot be recovered. Before you delete data or a file system, make sure that the data you want to keep has been backed up or migrated.

​
Runtime environment restrictions

The following restrictions apply to the runtime environment of CES instances or CKS cluster nodes after mounting. If you need to modify these configurations, contact support first.

​
Port requirements

Keep the following ports and protocols used by GPFS available: If the network port of an IP address used by GPFS changes, for example from eth0 to br0, restart the GPFS client on the node.

​
API rate limits

If you manage GPFS file systems by using APIs, follow the API rate limits. Storage management APIs: CES management service APIs: