Rock around the clock (synchronization
summary
Claim: deploying high precision clocks (Clockwork's software implementation of Stanford HPTS ideas) and hardware like PTP-capable NICs lets systems measure true oneway delay and control congestion. Demo 1: EKS e‑commerce under a Spark shuffle at 150 Gbps sees retransmits spike and 99th/Max latencies hit 8–10s then recover to milliseconds when receiver‑driven pause signals use synchronized clocks. Demo 2: 20‑node AWS cluster (10 Gbps per node, ns offsets) drops retransmits to near zero and enforces 6G/3G/0.8G bandwidth slicing.
AI-written summary · about
chapters (11)
- 0:00Introduction to time sync
- 1:26History of time measurement
- 3:54Quartz and atomic clocks
- 5:21Distributing time in networks
- 7:27NTP protocol explained
- 11:26High frequency trading needs
- 12:53Precision Time Protocol
- 13:54Congestion control with sync
- 16:06Demo: E-commerce latency
- 20:59Demo: Bandwidth allocation
- 24:35Key takeaways and Q&A
related talks
14mLightning Talk: Beyond NTP: High-accuracy clocks for distributed observability
46mSelf-Programming Networks: Network Clock Synchronization, Event Playback and Real-time Action
13mLive in the Express Lane
14mSyncWise: Error-Aware Time Synchronization for Reconfigurable Data Center Networks
47mAdventures in Spacetime
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
30mOk, it's time to synchronize our clocks...
38mTechniques and Tools for a Coherent Discussion about Performance in Complex...