Debugging Transient Faults in Data Centers using Synchronized Network-wide Packet
summary
Implements SyNDB: programmable-switch packet recording with nanosecond-synced data-plane clocks, ring-buffered compressed “p-records,” and control-plane eviction on a configurable fault trigger. Prototype runs on Barefoot switches with a Rust runtime and P4 programs; simulation and testbed show capture of ~4–11 ms history across the fabric (hundreds of RTTs) and successful root-cause of microbursts and misconfigurations. Typical cost is 5–7 MB SRAM per switch and export only on triggers, avoiding continuous line-rate telemetry.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:35Cloud Reliability is Critical
- 1:05Data Center Network Failures
- 1:34Transient Faults: Microbursts
- 2:13What do we need from the network?
- 4:24How do we solve?: An Outline
- 6:34In-Network Fault Detection
- 7:04Packet record Collection
- 8:56SYNDB Runtime
- 9:27SYNDB Implementation & Evaluation
- 10:04Retrospection & Correlation Simulation
- 10:39SRAM Overhead
related talks
21mDistributed Network Monitoring and Debugging with SwitchPointer
26mF10: A Fault-Tolerant Engineered Network
32mActive Fault Finding in Networks
46mSelf-Programming Networks: Network Clock Synchronization, Event Playback and Real-time Action
16mRuntime Programmable Switches
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
25mStroboscope: Declarative Network Monitoring on a Budget
15mFiDe: Reliable and Fast Crash Failure Detection to Boost Datacenter Coordination