Reducing Data Center Resource Over-Provisioning Through Dynamic Load Management for Virtualized Network Functions.

Network Function Virtualization aims at replacing specialized hardware network appliances by commodity servers. In this paper, we address sub-second variations in data center network workloads, which place highly volatile processing demands on the servers. This makes an efficient dimensioning of the hardware resources dedicated to network function execution challenging. Based on the observation that short-term peak workloads typically do not hit all machines at exactly the same time, we propose to enable the servers to reuse under-utilized resources of their peers by selectively redirecting packets when local resources are exhausted. To satisfy line rate performance demands, we present a hardware load management layer, which is located in the ingress path of each server. Our simulative evaluation shows that the load management layer can reduce the hardware resources required for network function execution by up to 24% while maintaining network throughput and latency performance. Especially in large data centers, these resource savings can significantly reduce network expenses.