NUMASFP: NUMA-aware Dynamic Service Function Chain Placement in Multi-core Servers

Many prior works investigated service function chain (SFC) placement in multi-core NFV servers with an assumption of treating all CPU cores equal. However, sophisticated servers follow non-uniform memory access (NUMA) architecture in which CPU cores are distributed across different NUMA nodes to enhance scalability. Our preliminary experiments show that placing VNFs in arbitrarily selected cores can degrade SFC performance, owing to contention for system resources at intra-node and inter-node levels. To handle the aforementioned problem, we study the impact of placing SFCs in a server by jointly considering the discrepancy of cores in different NUMA nodes and carefully allocating shared resources such as last level cache and memory bandwidth. We first propose a mathematical formulation, to optimally choose a node to place each SFC and allocate required shared system resources, in the form of a non-linear integer programming (NLIP) problem. Then we develop NUMASFP, a NUMA-aware dynamic SFC placement mechanism that places SFCs inside a server and migrates SFCs among NUMA nodes based on their current traffic rates to maximize the aggregate throughput of all SFCs. Furthermore, it ensures performance isolation for SFCs that are running inside the server. We evaluate the performance of NUMASFP by implementing it on a prototype system and also conducting large-scale simulation studies. The performance results demonstrate that the proposed mechanism can effectively handle SFCs over state-of-the-art and baseline approaches.