Hashing Design in Modern Networks: Challenges and Mitigation Techniques
summary
Demonstrates that limited uncorrelated hardware hash functions cause hash polarization that can lose about 33% of usable capacity; builds two deployable fixes: a topology-aware color recombining scheme that reuses hash functions safely in multi-stage DCNs, and a controller-driven co-prime resizing method that resizes ECMP groups to break correlation in mesh/direct-connect topologies. Evaluation uses Broadcom RTX7 hashes on real traces and topologies and reports COV improvements from ~0.6 to ~0.05 and production COV as low as 0.03.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:28Multipath forwarding in modern networks
- 1:21Hash polarization
- 4:16Hashing requirements of Google's datacenter networks
- 8:00Hash function reuse - color recombination
- 9:49Hash function reuse - Copriming
- 10:54Copriming illustration
- 12:47Copriming applied
- 16:13Simulation: direct connect topology
- 16:52Trade-off between weight skew and switch table usage
- 17:17Production Deployment
related talks
13mHashing Linearity Enables Relative Path Control in Data Centers
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
17mTopoOpt: Co-optimizing Network Topology and Parallelization Strategy for Distributed...
22mAdapting TCP for Reconfigurable Datacenter Networks
14mRDC: Energy-Efficient Data Center Network Congestion Relief with Topological
16mGRO Library Enhancements
25mVirtualizing Traffic Shapers for Practical Resource Allocation
17mMeta’s Network Journey to Enable AI