GPFS overview

GPFS is a shared file storage system designed 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.

​

Use cases

GPFS is suitable when workloads need to share data through a file system, for example:

• Multiple CES instances or CKS cluster nodes need to access the same data.
• Training tasks need to read datasets and write model weights, checkpoints, or training outputs.
• Inference, batch processing, or data processing tasks need to share input files and output results.

​

Core concepts

Creating a mount service, binding a file system, and adding clients prepare the access environment. Compute nodes can access data through a local directory only after the file system is mounted to that directory.

​

Usage workflow

1. Choose a specification based on your data volume, throughput requirements, and region, then create a GPFS file system.
2. Prepare a mount service: reuse a suitable service or create one, then bind the file system and add clients as needed. See Manage GPFS mount services.
3. Mount the file system or one of its subdirectories to a local directory on the target compute nodes.
4. Read and write data through the mount directory, and monitor the file system to track capacity, performance, and status.
5. Resize the file system as your data volume changes. Unmount it when access is no longer needed, and delete it when you no longer need to retain it.

​

Specification types

GPFS provides two specification types: performance and capacity. Contact technical support for an assessment if you need to change the maximum bandwidth or increase the inode limit.

​

Capacity and billing

GPFS file systems support two billing modes:

• Pay-as-you-go: Billed based on actual usage duration and configured capacity.
• Subscription: Prepaid for a fixed term.

All capacities in the GPFS specifications table are usable capacities. The underlying file system may use replicas or erasure coding to protect data. Depending on the configuration, capacity and usage reported by df or du on compute nodes may differ from the usable capacity and usage shown in the console. For example, when replication is enabled, df may report a capacity greater than the usable capacity shown in the console, and du may report 2 GiB in use after you write a 1 GiB file.

​

Capacity and quantity limits

GPFS file systems use thick provisioning. You must specify the capacity when creating a file system. Expanding or shrinking the file system adjusts its target capacity. The starting capacity and expansion increment depend on the storage hardware model. Specifications may vary across regions and storage clusters.

​

Performance and utilization recommendations

For stable performance, keep both storage space and inode utilization below 95%. High file system utilization can degrade I/O performance.

​

Data protection

GPFS protects data through mechanisms such as replication and erasure coding. You still need to back up important data yourself for additional protection. After you delete a file system or data in a file system, the data cannot be recovered. Before performing the deletion, ensure that any data you need to retain 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. Contact technical support before changing these configurations.

​

Port requirements

Ensure that the following ports and protocols used by GPFS are available: If the IP address used by GPFS is moved to a different network interface, for example, from eth0 to br0, restart the GPFS client on the node.

​

API call rate limits

If you manage GPFS file systems through the API, comply with the call rate limits. Storage management API: CES management service API: