NUMAP: NUMA-aware Multi-core Pinning and Pairing for Network Slicing at the 5G Mobile Edge
Wen-Ping Lai, Kuan-Chun Chiu · 2019
Based on the concept of network functions virtualization (NFV) by adopting the virtualized micro-service-based event-driven model, this paper studies the system resource allocation problems of deploying network/service slices on the mobile edge server, performing as a pivotal control/data/information hub in between radio-access and core networks or even the Internet clouds, in the coming era of 5G digital transformation. A non-uniform memory access (NUMA)-aware multi-core pinning-and-pairing method, called NUMAP, for network/service slicing with respect to different traffic levels is proposed to improve the system performance of a light-weight EPC slice (vEPCLW) on top of an ×86 Dell PowerEdge Server (R740) for the MEC cloudlet platform. This server is equipped with two CPU sockets, each containing 12 physical cores sharing the same local memory bank, denoted as a NUMA node; namely remote memory access time to another NUMA node is longer than the local one. The novelty of the proposed NUMAP method lies in the fact that it is aware of three important pairing schemes based on their pairing distances, namely inter-node-pair, intra-node-pair and hyper-threaded-pair, and NUMAP can thus serve as a New Map for multi-core assignment. Preliminary experimental results show that the NUMAP algorithm outperforms the default one which is based on symmetric multi-processing (SMP).