How Alibaba’s SMC‑R Brings Zero‑Modification RDMA to Cloud Applications

From TCP to RDMA

RDMA is a zero‑copy, protocol‑offload technology widely used in high‑performance computing, high‑frequency trading and other latency‑sensitive scenarios. Alibaba Cloud has deployed RDMA in its data centers, supporting core services such as ESSD and PolarDB.

From verbs to socket

Traditional RDMA uses the IB verbs interface, which is incompatible with the POSIX socket API, creating a high adoption barrier. Projects like rsocket and libvma attempted to bridge this gap, but they suffer from user‑space resource management issues and compatibility limitations.

SMC‑R: Kernel‑level socket‑compatible RDMA

SMC‑R (Shared Memory Communications over RDMA) implements a TCP‑compatible socket family that uses RDMA RC for data transfer. It operates in kernel space between the socket layer and the IB verbs layer, allowing applications to switch from AF_INET to AF_SMC without code changes.

Alibaba Cloud Linux 3 and Anolis OS add sysctl controls and whitelist mechanisms to enable transparent, zero‑modification replacement of TCP with SMC‑R, supporting both RoCE and iWARP networks.

Automatic negotiation and safe fallback

During connection setup, peers exchange a TCP option indicating SMC‑R support. If either side cannot use RDMA, the stack falls back to standard TCP over the established connection, ensuring reliability.

Shared‑memory ring buffer communication

SMC‑R uses a remote memory buffer (RMB) organized as a ring, with producer and consumer cursors to coordinate writes and reads, guaranteeing data integrity and avoiding overrun.

Monitoring and diagnostics

Tools such as sysctl, proc entries, and the smc‑tools suite provide visibility into SMC‑R sockets and performance metrics.

Performance results

Benchmarks on Alibaba Cloud with Elastic RDMA show that SMC‑R improves Redis QPS by ~50% for latency‑sensitive workloads and increases Thrift and Netty RPC throughput by 30% and 12% respectively, while maintaining full socket compatibility.

Limitations

RDMA connection establishment is slower than TCP, and memory consumption per connection is higher, making SMC‑R less suitable for short‑lived connections or public‑internet exposure.

Future work

The SMC‑R code will be open‑sourced in the OpenAnolis high‑performance network SIG, with documentation on Alibaba Cloud Linux 3, aiming to make zero‑modification RDMA adoption easier for more cloud applications.