Cloud Parallel File Storage:Cache acceleration

Important notices

• Release status: The EFC cache feature is in invitational preview.

• Billing: EFC cache is free during the invitational preview.

• Data reliability: The cache does not guarantee high data availability. Cached data may be lost due to machine corruption, node replacement, or decommissioning. Back up important training data promptly. Do not use the cache as the sole data store.

Use cases

Inference

Model files must load onto every GPU server quickly, ideally within one minute. As GPU count grows, storage performance requirements increase proportionally. EFC distributed cache on local GPU disks and memory decouples storage performance from model loading speed, enabling rapid model loading at scale.

Training

Training workloads read datasets repeatedly. Client-side caching improves read performance, increases GPU utilization, and boosts overall training speed.

Key capabilities

• High-speed client-side cache

  EFC uses compute node memory and local disks to build single-node and distributed caches. This accelerates dataset and checkpoint read/write operations and reduces remote storage access latency.

• P2P model distribution

  EFC uses peer-to-peer (P2P) technology for high-concurrency loading and distribution of large-scale models. It leverages the high-speed network between GPU nodes to accelerate parallel reads of hot data. Aggregate throughput scales linearly with compute node count, supporting hundreds to thousands of nodes. The following figure shows how P2P works.

  

• Full-path hardware acceleration

  EFC supports Remote Direct Memory Access (RDMA) networks and VSC hardware-software integrated paths. Multi-link load balancing distributes requests across multiple backend servers, delivering low latency and high throughput under high concurrency.

• POSIX compatibility and cloud-native integration

  EFC is designed for containers and fully compatible with POSIX file interfaces. It provides container-level fencing, Quality of Service (QoS) capabilities, and application-oriented access control.

Limits

Get started

• Platform for AI (PAI) (Lingjun resources)

  EFC cache uses local disks on PAI Lingjun resources to build a distributed read cache. It accelerates data reads by reducing reliance on remote storage and improves application performance. For details, see Use EFC cache to accelerate tasks on the PAI platform.

• Lingjun bare metal

  Console-based activation is not yet available for Lingjun bare metal. To request a trial, submit a ticket.