Impact of RDMA Technology on High‑Performance Data Centers and Its Adoption at JD.com

JD.com IT Resource Service Department recently organized a Future Data Center Core Technology Seminar, inviting directors and technical experts from the company’s AI, Big Data, and Cloud Computing teams.

During the seminar, participants discussed the hot topic of RDMA (Remote Direct Memory Access) and high‑performance networking, emphasizing the growing impact of deep learning and cloud computing on application performance.

Impact of RDMA on High‑Performance Data Centers

RDMA enables direct memory‑to‑memory data transfer without involving the operating system kernel, achieving kernel bypass.

Compared with traditional TCP, TCP requires extensive CPU participation for packet creation, CRC calculation, flow control, and memory address lookup, consuming significant CPU resources and being vulnerable to CPU noise, which degrades network performance.

With RDMA, all these operations are handled by the network‑card hardware, resulting in low latency, high bandwidth, and minimal CPU utilization.

JD.com’s Use of RDMA in Data Centers

RDMA can be used directly via the verbs API, allowing applications to fully exploit its performance.

Popular deep‑learning frameworks such as TensorFlow and big‑data frameworks like Spark can transmit data over RDMA.

JD’s AI research team employs RDMA for high‑performance model file transfer in distributed training scenarios.

Major database systems now support RDMA interfaces, and parallel file systems such as IBM GPFS, Lustre, and GLustre provide native RDMA support.

Middleware can encapsulate RDMA for common compute and storage use cases.

All major MPI implementations (OpenMPI, Intel MPI, MVAPICH, etc.) support RDMA, and it is widely used in the world’s fastest supercomputers.

Storage protocols also leverage RDMA: iSCSI over RDMA (iSER), Ceph over RDMA, and the newer NVMe‑over‑Fabric protocol all require RDMA for data transfer.

Special‑purpose applications in finance, securities, and multimedia use VMA to gain RDMA’s low‑latency benefits without changing socket APIs.

JD.com IT Resource Service Department leader Lv Ke noted that JD, Microsoft, and Facebook are strengthening RDMA usage to reduce data‑center costs, improve application performance, and address TCP scalability issues.

Testing by JD’s hardware systems team showed that, in virtual‑machine environments on the same physical server, RDMA can free nearly 100% of CPU resources for computation, allowing more VMs to run on the same hardware.

JD Technology public account focuses on sharing technical insights.