RSC: Enabling High-Performance RDMA for Serverless Services with Adaptive Network Multiplexing

Abstract

Serverless functions are short-lived, yet RDMA Reliable Connections (RCs) are expensive to create and assume long-lived endpoints. Even when compute sandboxes are warm, an invocation can still suffer a network-cold start if it establishes a new RC or accesses a remote endpoint for the first time. We present RSC (Reliable Serverless Connectivity), a high-performance user-space RDMA stack that eliminates RC setup from the invocation hot path. RSC exposes multiple operating modes that trade off latency, CPU efficiency, isolation, payload registration, and physical QP budget to support diverse serverless workloads. Its density-adaptive architecture scales efficiently to thousands of tenants, providing an in-process path for latency-critical calls, a pre-registered path for large transfers, memory-window isolation, and a diagnostic shared-agent path that measures the cost of physical QP consolidation. Our evaluation on 100 Gbps RoCE shows that native RC setup costs 7.94 ms, while reused-connection 64 B calls reach 59.87 μ s in adaptive sleep mode and 4.70 μ s with aggressive busy polling. For 2 MB payloads, the pre-registered path reaches 236.80 μ s. Workflow, request‑mix, resource, noisy‑neighbor, endpoint‑scale, and QP‑density diagnostics confirm that connection reuse removes millisecond‑scale RC setup from warm service invocation paths. These results demonstrate that RSC can make RDMA economically viable for serverless services.