RDC: Energy-Efficient Data Center Network Congestion Relief with Topological
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
RDC proposes inserting circuit switches at the edge between ToR switches and servers to regroup servers across racks by reconfiguring circuits based on measured flows; a controller estimates demand (proactive or reactive), runs a graph‑partitioning optimizer, and pushes circuit + ToR routing changes. A 16‑server optical testbed with four logical racks and a 512‑server HTC packet‑level simulation show concrete gains: localizing inter‑rack traffic (Facebook trace drops ~96%→38%), speeding a 4:1 oversubscribed fabric by 1.78–3.9× (over 10× at scale), and improving performance per watt up to 2.4×.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:41Top-of Rack Design: Almost Every Data Center Uses
- 2:11Key Observation: Traffic are Skewed at Server-level
- 3:11Power of Regrouping Servers to Break the Rack Boundary
- 5:45Circuit Switching: Key Building Block of RDC
- 6:26RDC Controller
- 6:41Collect Flow Pattern
- 7:13Topology Optimization (Traffic Localization)
- 7:59Topology Reconfiguration
- 8:12Testbed Setup
- 8:28HDFS Write Need to Place the 3rd Copy onto a Different Rack
- 9:43Simulation Setup
- 12:30Energy Consumption Analysis
- 12:56Summary
related talks
21mLarry: Practical Network Reconfigurability in the Data Center
22mAdapting TCP for Reconfigurable Datacenter Networks
20mRDMA at Cloud Scale: The OCI Experience
26mApplying New High-Throughput Network Solutions in IEC (Integrated Edge Cloud)
17mDRack: A CXL-Disaggregated Rack Architecture to Boost Inter-Rack Communication
19mTackling DC Congestion and Bursts
23mMinimal Rewiring: Efficient Live Expansion for Clos Data Center Networks
25mRDMA @Scale