CETOFS: A High-Performance File System with Host-Server Collaboration for Remote Storage
summary
CETOFS implements a userspace-kernel collaborative architecture for NVMe-over-RDMA remote SSDs, moving the data path into user space and offloads permission checking, concurrency control, failure-atomic IO to the storage server. It runs on RDMA and Optane-class SSDs, implements block-granularity owner maps, request-group creation and merging, and transaction-aware batching. Compared to kernel Ext4 and a userspace FS it cuts single-thread read/write latency by up to 52% and raises concurrent throughput up to 19×, demonstrating the prototype works and scales.
AI-written summary · about
chapters (13)
- 0:00CETOFS introduction and motivation
- 1:00Remote SSD access workflow
- 2:44Software overhead in remote storage
- 3:48Scalability with concurrent access
- 4:51Atomic IO over remote storage
- 5:53CETOFS architecture design
- 6:55Permission checking in CETOFS
- 8:20Concurrency control with grouping
- 10:52Merging request groups
- 11:34Atomic IO and transactions
- 11:56Evaluation setup and metrics
- 12:40Results: Kernel overhead and scalability
- 13:02Conclusion and future EFS work
related talks
17mFisc: A Large-scale Cloud-native-oriented File System
34mNVMeoF in Ceph - What's New and Performance Best Practices
14mFlexTOE: Flexible TCP Offload with Fine-Grained Parallelism
30mOrion: A Distributed File System for Non-Volatile Main Memory and RDMA-Capable Networks
39mSecured I/O in the Cloud with NVMe-Over-TCP
16mNo Provisioned Concurrency: Fast RDMA-codesigned Remote Fork for Serverless Computing
49mPetabyte CephFS for Satellite Data Processing
17mCo-Designing Traffic Control with NVMe-oF for Disaggregated Storage: A Comparative Study
