Everyone Loves File: File Storage Service (FSS) in Oracle Cloud Infrastructure
Talkat USENIX ATC '19 · on Wednesday, 2019-07-10· at 11:40· USENIX ATC Track I: Grand Ballroom I–VI·
summary
Oracle FSS implements a managed NFS by building a pipelined Paxos replicated backend: a scalable block store offering linearizable multipage limited-size transactions (the system supports a 15-page MPSC atomic primitive) and a scalable B-tree that maps keys to pages using self-validating nodes. Snapshots use versioned key-value pairs, presentation servers are stateless, extents use 5-way replication, and a Paxos-replicated hash table bootstraps config. Throughput follows a simple model (bandwidth/5 with queuing losses, practical running ~1/3 peak); unloaded write latency is ~4× a Linux NFS server on NVMe.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:40What is File Storage Service
- 1:50Architecture of File Storage Service
- 4:15Problems with Standard File Servers
- 4:47Scale Up: FSS Distributes Data Structures Across Servers
- 7:01Represent Inodes as Tabular Data instead of implementing the inodes directly
- 8:08Store the Tabular Data in a Distributed B-tree
- 9:31Updating the B-tree Requires Atomic Operations Across Servers
- 11:25The Problem with 2PC
- 12:04Paxos Doesn't Get Stuck
- 12:48FSS Uses Paxos to Implement two-phase commit
- 13:18How Big is an Extent?
- 14:26Multipage Store Conditional (MPSC) On top of 2PC, We program FSS using an
- 16:46A Simple Throughput Model
- 17:57Looks Like a Speedup Curve
related talks
17mSingularFS: A Billion-Scale Distributed File System Using a Single Metadata Server
16mHadaFS: A File System Bridging the Local and Shared Burst Buffer for Exascale
15mInfiniFS: An Efficient Metadata Service for Large-Scale Distributed Filesystems
12mFacebook's Tectonic Filesystem: Efficiency from Exascale
20mscoutfs: Large Scale POSIX Archiving
21mNew Techniques to Improve Small I/O Workloads in Distributed File Systems
45mA Distributed File System for Secure P2P Applications
19mConsolidating Storage Backends at Scale with Tectonic Filesystem
