Fisc: A Large-scale Cloud-native-oriented File System
summary
A cloud-native file system that offloads heavy client functions into local agents and DPU-based virtio devices and introduces vrpc and a file-based range table to route reads to local chunk servers. Deployed in production for over three years serving 3 million cores, it reports about 60% CPU reduction in clients, sustains ~4 million IOPS, and observes end-to-end read latency under 500 µs for online search while enabling local reads, RDMA transport, and storage-aware retry and recovery.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:37Change in Cloud: Evolving To Cloud-native
- 1:48Bottleneck 1: Heavy Clients
- 2:54Bottleneck 2: Unsuitable Network Load Balancer
- 4:58Design Goals
- 5:21Fisc Architecture
- 8:15Lightweight Clients : VRPC
- 9:11Software and Hardware Co-design
- 10:22Storage-aware Distributed Gateway
- 11:41File-based Route Table & Locality-aware Read
- 13:22Microbenchmark: Lightweight Fisc Clients
- 13:43In Production: Lightweight Fisc Clients
- 14:37E2E Latency and SLA in Production
- 15:26Discussion: what's more
- 16:10Conclusion
related talks
14mCETOFS: A High-Performance File System with Host-Server Collaboration for Remote Storage
11mWhen Cloud Storage Meets RDMA
19mBifrost: Analysis and Optimization of Network I/O Tax in Confidential Virtual...
18mFileMR: Rethinking RDMA Networking for Scalable Persistent Memory
15mScalio: Scaling up DPU-based JBOF Key-value Store with NVMe-oF Target Offload
28mCeph RBD Integration with DPUs
30mOrion: A Distributed File System for Non-Volatile Main Memory and RDMA-Capable Networks
26mApplying New High-Throughput Network Solutions in IEC (Integrated Edge Cloud)
