How Baidu Cloud’s Sonata Achieves Predictable Performance Across Multi‑Vendor, Multi‑Generation NICs (NSDI’27)

Ba​idu Intelligent Cloud’s new data‑plane implementation, Sonata , was accepted to NSDI’27. The paper asks how a cloud network data‑plane can stay high‑performance and stable when NICs come from multiple vendors and span several generations.

Design Choice: Software‑Centric Data Plane

Sonata deliberately avoids deep binding to any NIC class. It uses only the standard interfaces that all NICs share to offload stable, long‑lived traffic, while leaving complex, short‑lived flows and consistency logic to high‑performance software. This "Software‑Centric" approach lets the same data‑plane run on heterogeneous hardware without per‑vendor customisation.

Three Core Challenges

Performance : Massive short‑connection workloads (live streaming, P2P, serverless) cannot be offloaded to NICs and would overwhelm the software session table. Sonata aggregates similar sessions into "session groups" using an Aggregated Match Cache (AMC), reducing the number of entries the fast path must handle to about one‑fifth of the original count.

Robustness : Frequent network updates (millions per day per region) cause churn. Instead of a flush‑all that clears the hardware flow table, Sonata’s Proactive State Orchestration (PSO) extracts only the affected rules, generates probe packets to verify actions, and updates mismatched entries, keeping unrelated flows untouched.

Isolation : Policies must be expressed uniformly in software but enforced efficiently in hardware. Sonata builds on the standard srTCM/trTCM token‑bucket meters (HQoS) and abstracts them so that rate‑limiting can be applied across all NICs, even though each NIC provides only a limited number of meters.

Key Mechanisms

Session Aggregation : By grouping flows with identical actions, Sonata cuts lookup work and preserves priority semantics while keeping the fast path lightweight.

Proactive State Orchestration (PSO) : Updates only the subset of hardware rules that actually change, avoiding upcall storms and minimizing disruption.

Thread‑Level Hot Upgrade : New dynamic libraries are loaded in‑process, global offset tables are rewritten, and hardware queues, huge‑page memory, and offload contexts remain untouched, limiting upgrade‑induced latency to a few hundred microseconds regardless of VM count or NIC type.

HQoS Isolation : Uses shared token‑bucket meters as building blocks; a 200 Gbps NIC can host roughly ten 20 Gbps “elephant” flows, which are identified and constrained, while the remaining millions of “mouse” flows are handled in software.

Production Evaluation

In real‑world deployments covering tens of thousands of servers, Sonata increased the BvS software CPS (new connections per second) by 2.13‑2.51×, surpassing the OVS data‑plane by 2.37‑2.75×. Enabling hardware offload added a further 4.3‑11.5× boost. Rule updates and hot upgrades introduced only microsecond‑level jitter; hot upgrades stayed under 300 µs, while a traditional active‑standby process caused tens of milliseconds of pause.

Across four different NIC platforms, incremental optimisations showed consistent CPS gains: session aggregation alone yielded ~2.2× improvement, and adding hardware offload approached an order of magnitude on some platforms. Policy‑change latency no longer degraded with session size, and HQoS accurately throttled elephant flows without harming other traffic. In a short‑video service, PSO and hot upgrades reduced configuration‑change jitter and tail latency, and session structures survived six version upgrades without rebuilding active sessions.

Conclusion

By keeping the data‑plane software‑centric, Sonata solves the three challenges—performance, robustness, and isolation—across any vendor or generation of NICs, delivering predictable cloud‑network behavior. The authors argue that this approach generalises to any infrastructure that must evolve on heterogeneous hardware, because software‑level capabilities can be reused while hardware advances are absorbed through standard interfaces.