Söze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
summary
Söze uses a single network telemetry signal—the maximum queuing delay from the path (via CIC format)—to drive decentralized per-flow rate updates and realize weighted bandwidth shares without per-flow or topology knowledge. It’s implemented as a Linux kernel module and an eRPC kernel-bypass dataplane and converges in constant time across topologies, matching a centralized water‑filling baseline. Simulations and testbed runs show TPC‑H job completion improvements of up to 0.59× and 0.79× on average, demonstrating practical accuracy and low overhead.
AI-written summary · about
chapters (12)
- 0:00Weighted bandwidth allocation intro
- 1:24Data center bandwidth sharing
- 2:48Minimal info for rate allocation
- 3:33Weighted rate allocation equations
- 5:03Queue stabilization indicator
- 6:50Queue delay to rate-weight mapping
- 8:38Convergence to weighted allocation
- 9:45Flows disguising weights
- 10:48Bottleneck hops in networks
- 12:15INT for bottleneck signaling
- 13:41Implementation and evaluation
- 15:50Conclusion and Q&A
related talks
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
20mETC: An Elastic Transmission Control Using End-to-End Available Bandwidth...
17mPrecise Data Center Traffic Engineering with Constrained Hardware Resources
25mVirtualizing Traffic Shapers for Practical Resource Allocation
15mJustitia: Software Multi-Tenancy in Hardware Kernel-Bypass Networks
36mNetwork Bandwidth-Aware Kubernetes Cluster
19mNet-P4ct: Enhanced WAN Bandwidth Fair Sharing Using P4 Programmable Switches