Taurus: Towards A High-Performance and Generic Congestion Control Framework for Datacenter Networks#

Luyang Li et al. (ICT CAS / Chinese Academy of Sciences / Douyin Vision Co., Ltd. / HKUST)

• Why: Congestion-control frameworks (CCP, DOCA-PCC) are how production DCs keep RDMA/lossless-ish fabrics usable across diverse workloads, and this is one of the few papers with a real 400Gbps production deployment behind it — with authors from ByteDance's Douyin group, it carries deployment weight.
• What: A hardware–software co-designed CC framework that lets operators deploy varied CC algorithms at near-native performance. Reported results vs. CCP: +32.3% throughput, −96.4% latency, −158.7% CPU overhead; vs. DOCA-PCC: +9.3% throughput, −28.8% latency (from abstract only).
• How: Partitions CC functionality into distinct tasks and maps them onto appropriate hardware vs. software components: a collaborative signal-collection mechanism supporting diverse congestion feedback, a type-aware message report engine with two-tier aggregation to cut host↔NIC communication overhead, and software built-in handlers so new algorithms can be deployed without changing the datapath.
• Where applicable: Datacenter fabrics at 100–400Gbps; requires FPGA-based NICs (implementation is on commodity servers with FPGA NICs), so it presumes you control the host NIC hardware. The Douyin production deployment makes the results credible but the environment is a specific hyperscale DC — expect the tradeoffs to shift with different NIC generations.