How DPU‑Based Architectures Revolutionize High‑Performance Storage Networks

Background

Increasing network bandwidth and storage performance have shifted a large portion of server resources—approximately 30%—to handling networking and storage protocols on the host CPU. At the same time, CPU performance improvements are slowing, reducing the energy‑efficiency of data‑center services.

Data Processing Unit (DPU) Overview

A DPU is a specialized processor built for data‑centric workloads. It uses a software‑defined architecture to virtualize infrastructure‑layer resources and offload networking and storage protocols, alleviating CPU bottlenecks.

Core DPU Technologies

IO hardware device virtualization

VPC overlay network acceleration

EBS distributed storage acceleration

Local storage virtualization acceleration

RDMA‑based high‑speed data transfer

Security hardware acceleration

Elastic bare‑metal support

Resource pooling capabilities

NVMe‑over‑RDMA Offload Modes

Non‑offload : All data passes through the embedded CPU, requiring multiple DMA copies between host memory, embedded CPU memory, and the NIC.

Zero‑copy : Data moves directly from host memory to remote storage without traversing the embedded CPU cache, reducing DMA hops. This mode relies on SPDK bdev and requires RDMA support on the storage side.

Full‑offload : Both control and data planes are handled entirely by hardware, eliminating embedded CPU involvement. It provides the highest throughput but limits software control over the storage backend.

High‑Performance Storage Architecture with DPU

By separating compute from storage, DPUs move network and storage protocol processing from the host CPU to the DPU, decreasing CPU utilization and increasing throughput. Integrated accelerators (encryption, compression, etc.) further speed up data handling. Hardware‑level storage virtualization allows a single physical device to appear as multiple virtual devices, reducing data copies and latency in virtualized environments.

Application Scenarios

(1) Bare‑Metal

In bare‑metal deployments, users have exclusive access to physical servers, running operating systems and applications directly on the hardware. This eliminates virtualization overhead, delivering higher performance, lower latency, and better isolation—suitable for large databases and high‑performance computing workloads.

(2) Virtualized

In virtualized environments, cloud providers split physical machines into multiple VMs to improve hardware utilization and reduce data‑center costs. However, VM access to network storage suffers from memory copies, virtualization overhead, and network device limits. DPUs can virtualize remote storage as local NVMe devices using SR‑IOV, dramatically reducing latency and achieving near‑native performance.

Performance Considerations

Traditional data paths involve multiple context switches and data copies between user space, kernel space, and network stacks, consuming CPU cycles and increasing latency. Offload modes reduce these overheads: Zero‑copy eliminates the embedded CPU cache hop, while Full‑offload removes software control entirely, offering the greatest throughput at the cost of flexibility.

Conclusion

DPUs provide a powerful mechanism to offload networking and storage workloads, enabling higher performance, lower CPU utilization, and more efficient data‑center resource usage. Selecting the appropriate offload mode requires balancing performance, control, and compatibility with existing storage solutions.

Related Reading

DPU Hardware Standardization Exploration – https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650749086&idx=1&sn=45210c6e19dae3ed31e2534890f99ee4

Advances and Future Innovations in DPU Technology – https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650749080&idx=1&sn=6a0b1f2bbbdbc470b4155abb253f1feb

Practical DPU Deployments and Use Cases – https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650746845&idx=1&sn=0813ac94a58503d068bc270a9b12d753