FarReach: Write-back Caching in Programmable Switches
summary
FarReach implements an in-switch write-back cache on a Tofino switch with P4 and a RoxDB backend, claiming up to 6.6× throughput versus a NetCache baseline under skewed write workloads with 128 simulated servers. It uses non-blocking cache emission to avoid data-plane stalls, visible cache eviction metadata so evicted entries are safely loaded to servers, and crash-consistent snapshot generation plus client-side preservation for zero-loss recovery; control-plane bandwidth stays ~1.4 MB/s at 2.5s snapshots.
AI-written summary · about
chapters (13)
- 0:00FarReach introduction
- 1:25Switch architecture overview
- 2:28In-switch caching challenges
- 4:14FarReach design solutions
- 5:43Non-blocking cache admission
- 8:37Visible cache eviction
- 11:29Crash consistency snapshots
- 14:17Zero data loss preservation
- 15:21Evaluation methodology
- 16:46Throughput results
- 17:08Scalability with 128 servers
- 17:51Snapshot performance impact
- 18:55Conclusion
related talks
17mPushing the Limits of In-Network Caching for Key-Value Stores
12mDebugging Transient Faults in Data Centers using Synchronized Network-wide Packet
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
20mFLAIR: Accelerating Reads with Consistency Aware Network Routing
12mConcordia: Distributed Shared Memory with In-Network Cache Coherence
23mFast Software Cache Design for Network Appliances
17mSwiSh: Distributed Shared State Abstractions for Programmable Switches
16mQueue-Mem: Energy-Efficient Hardware Storage for Advanced Network Function Acceleration