MoTor: Resource-efficient cloud-native network acceleration with programmable switches

Abstract

With the rapid increase in the number of services and the frequency of service invocations in distributed cloud-native applications, the volume of inter-service communication traffic in these applications has grown significantly over the years. As a result, cloud-native network, especially, inter-service communication, is crucial for maintaining the overall performance of cloud-native applications. However, our study indicates that extra intricate network processing induced by packet transmission imposes substantial overhead on inter-service communication. Therefore, we propose MoTor, a programmable switch-assisted network solution to accelerate cloud-native network in a resource-efficient way. The core idea of MoTor is to minimize network processing overhead by offloading essential network functions from the CPU to programmable switches. MoTor utilizes eBPF to enable non-intrusive network bypassing and forward inter-service traffic to the switch directly. The switch is then leveraged to perform critical network functions required for inter-service communication. To address the memory constraints of programmable switches, MoTor introduces a table entry prediction algorithm and a no memory-cost per connection consistency (PCC) preservation strategy. The evaluation results show that MoTor achieves salient performance and resource gains while maintaining PCC. Specifically, MoTor improves application’s throughput by up to 1.66$\times$ and reduces latency by up to 62.3% without any code modifications. Although other programmable switch-assisted solutions offer performance improvement comparable to MoTor, they require additional host or switch resources. In contrast, the PCC preservation strategy and prediction algorithm in MoTor enable simultaneous reductions in both host and switch resource consumption. Specifically,it reduces CPU and memory usage by up to 55% and up to 70%, respectively, compared to existing network solutions, while lowering the switch’s memory consumption by 60%.