CPFS for Lingjun is built for high-throughput, low-latency scenarios such as AI training, inference, and HPC. It provides two network channels for compute resources to access the file system. This topic is the entry point for the mount access documentation set. It describes how to choose a channel, the standard method for determining a compute-storage affinity domain, and an index of mount entry points for each compute resource.
Network channels and selection
The channel used is determined by two factors together: the type of compute resource, and whether the resource is in the same compute-storage affinity domain as CPFS.
Channel | Protocol | Applicable resources | Trigger condition |
VSC + RDMA direct mount | RDMA | Lingjun AI computing resources (bare metal / PAI Lingjun / ACK Lingjun / ACS Lingjun) | Same compute-storage affinity domain as CPFS |
VPC mount point | TCP | General-purpose computing resources + Lingjun resources across compute-storage affinity domains | Always used for general-purpose resources; used for Lingjun when crossing affinity domains, or as a fallback |
VSC + RDMA direct mount: Mounts directly over the RDMA network through the VSC (Virtual Storage Controller), providing the best latency and bandwidth. Available only when the Lingjun resource and the file system are in the same compute-storage affinity domain.
VPC mount point: Accessed over the TCP protocol through the mount point domain name that the file system exposes within the VPC (in the format
cpfs-***-vpc-***.<Region>.cpfs.aliyuncs.com). Applicable to general-purpose computing resources and to Lingjun resources accessing across compute-storage affinity domains.
For container scenarios, we recommend also setting mountpointAutoSwitch: "true", so that the CSI driver automatically switches to the VPC channel if the RDMA channel becomes abnormal or the workload crosses affinity domains.How to determine whether Lingjun and CPFS are in the same compute-storage affinity domain
This section describes the determination method shared by all mounting topics.
Call the DescribeFileSystemAssociatedHpnZones OpenAPI operation, passing in the target CPFS file system ID, to obtain the list of Lingjun HPNZones that are in the same compute-storage affinity domain as the file system.
Query the HPNZone of the Lingjun resource you plan to use through the Lingjun OpenAPI or the corresponding platform console (PAI, ACK, or ACS).
If the Lingjun HPNZone matches an entry in the list, the resources are in the same compute-storage affinity domain. If not, they are in different affinity domains.
Always use the API response as the authoritative result; do not determine this solely based on whether the HPNZone values are the same. Compute and storage resources with the same cluster number belong to the same compute-storage affinity domain, and resources with different cluster numbers may still belong to the same affinity domain.
Mount entry points by compute resource
Compute scenario | Documentation |
PAI (DLC/DSW/EAS) task mounting | |
Mounting for ACS clusters | |
Mounting for ECS/EGS instances | |
VPC mount point resource management (create/view/delete) | |
VSC OpenAPI programmatic access |