DPU Performance Benchmark Methodology and Implementation (2022)

The article introduces the DPU performance benchmark testing system framework and process, covering the test system, test requirements, and test activities.

Three types of DPU test systems are defined: Single-End, End-to-End, and Multi-End.

Single-End systems consist of a host and a DPU directly connected via a bus (typically PCIe) forming a closed environment; they measure the acceleration of specific compute tasks performed by the DPU, using metrics such as data‑processing time and scale.

End-to-End systems link two Single-End systems with a direct network cable, designating one as initiator/client and the other as target/server; they evaluate the DPU’s network‑acceleration capabilities, testing operations like RDMA Send/Receive and NVMe‑oF, and measuring latency, throughput, and peak network performance.

Multi-End systems interconnect multiple Single-End units through a complex network topology, simulating real‑world data‑center environments; they assess DPU handling of network, storage, and security acceleration, measuring latency, throughput, and connection capacity under varied traffic and attack scenarios.

The testing workflow executes each test case in two rounds, each round comprising a warm‑up run, a measured run, a result‑check phase, and a clear‑up step; the system configuration must remain unchanged throughout, and any unrecoverable errors invalidate the round.

Test reports must present performance metrics for each case and include complete configuration details—hardware resources, OS and software versions, compilation options, and any anomalies observed during execution.

Required components include the DPU Benchmark program, open‑source tools, benchmark driver, DPU driver, operating system, computing device (CPU), connected devices (PCIe, CXL), I/O hardware, the DPU product under test, and other hardware resources such as memory and storage.