Scalable Tail Latency Estimation for Data Center Networks
summary
Claims a practical estimator that cuts ns-3’s 11+ hour run for a 6,000-host, 5s trace to about a minute by decompose into independent single-link simulations, independently simulate per-link congestion (preserving RTTs and source fan-in), bucket delays by flow size, and convolve per-link delay distributions Monte Carlo-style to get end-to-end FCT tails. Prototype Parsimon prototype runs in ~80–120 seconds, matches ns-3 tail slowdowns within ~9–10% for moderate loads, and degrades mainly under persistent high load.
AI-written summary · about
chapters (5)
- 0:00Motivation for tail latency
- 1:37Limitations of ns3
- 3:44Strategy for approximation
- 9:13Parsimon implementation
- 12:46Evaluation and conclusion
related talks
17mHow to diagnose nanosecond network latencies in rich end-host stacks
22mAdapting TCP for Reconfigurable Datacenter Networks
45mReducing Network Latency: Innovations for a Faster Internet
16mCredence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions
18mMeasuring Congestion in High Performance Datacenter Interconnects
46mSelf-Programming Networks: Network Clock Synchronization, Event Playback and Real-time Action
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
16mRock around the Clock (Synchronization): Improve Performance...