Performance investigation on disaggregated artificial intelligence data centers beyond rack scale with optical switching

Accompanying the ever-increasing scale of big data applications, artificial intelligence data centers are facing the issue of resource fragments, resulting in a low network resource utilization ratio. Disaggregating the network resources is an efficient solution to improve the network resource utilization ratio by allocating the required amount of resources. In this paper, we focus on the problem of CPU and GPU resource disaggregation for an artificial intelligence data center. We carry out investigations for data center disaggregation exploiting optical switching. The results show that PCIe over optical (PO) guarantees 3 µs latency with 62 m of fiber. Compared with the PCIe with Ethernet switch (PE) solution, the PO scheme saves 48.3% completion time for the backprop application. Moreover, we compare the cost and power consumption of a data center architecture that scales out as the square of the port count of an optical packet switch (OPSquare) employing a PO scheme with respect to variants of OPSquare and Leaf-Spine under different network scales and interface bandwidths. Results show that the optical network architecture with the PO scheme saves 31.6% in cost and 12% in power consumption, respectively, compared with the Leaf-Spine with Ethernet solution.