Orion: A Distributed File System for Non-Volatile Main Memory and RDMA-Capable Networks
summary
Orion is a kernel file system that unifies POSIX I/O and RDMA networking to expose distributed NVMM as byte-addressable storage, using one-sided RDMA for data and two-sided verbs for metadata. It inherits NOVA’s per-inode log, implements tail-check and speculative log-commit to persist within ~500 CPU cycles, and supports replication over RDMA. A Linux prototype (about 15,000 lines of kernel code) using Mellanox InfiniBand achieves up to 14M tail-checks/sec with 8 clients, 1.3–100× better than prior distributed systems and 68–85% of local NVMM throughput.
AI-written summary · about
chapters (9)
- 0:00Introduction to Orion
- 3:04Design Motivation
- 6:16Orion Architecture
- 7:51File System Implementation
- 10:01Core Operations
- 12:47Metadata Optimization
- 17:22Evaluation Results
- 19:09Conclusion and Future Work
- 21:07Q&A Session
related talks
20mRDMA at Cloud Scale: The OCI Experience
18mFileMR: Rethinking RDMA Networking for Scalable Persistent Memory
14mODRP: On-Demand Remote Paging with Programmable RDMA
15mNaos: Serialization-free RDMA networking in Java
28mPersistent Memory plus RDMA, new age remote device
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs
16mAchieving Wire-Latency Storage Systems by Exploiting Hardware ACKs
12mOrion: Google's Software-Defined Networking Control Plane