Implementation of NVMe Over TCP Using SPDK and Its Performance Measurement

In data centers, massive GPGPUs and SSDs are employed to implement massively parallel computing and high-speed huge data storage. To meet the demand for such computing facilities, we believe that disaggregated computer architecture are effective, i.e., provisioning GPGPU clusters and SSD clusters separately and interconnecting these clusters based on NVMe over Fabrics (NVMe-oF) standard. In this paper, we report an implementation of NVMe over TCP, an architecture based on NVMe over Fabrics standard, using the Storage Performance Development Kit (SPDK) and performance measurement.