Resource Affinity Model Optimization And Research Driven By SR-IOV NIC

Non-uniform memory access (NUMA) takes the advantage of Sym-metric Multi-Processing (SMP)’s ease of programming and Massively Parallel-Processing (MPP)’s scalability, So in large servers, NUMA is the main-frame. But it also brings in a classic problem, when a processor is process-ing memory data, it has a better efficiency to run close with the data. And in high performance virtualization network, Single Root I/O Virtualization (SR-IOV) interface achieves good performance. However, the scheduler brings in impact to the network performance, as the vCPU’s scheduling doesn’t consider the Virtual Machine (VM)’s buffer’s allocation, which causes the classical NUMA problem.Our paper first consider of the location of VM’s network buffer, which have different affinity with different NUMA node. Then we also consider of CPU’s workload balance, which will avoids the situation that we choose a NUMA node, which has a good affinity but also a high workload pres-sure to run the vCPU. At last, we try to dynamically allocate the vCPU’s to the best position of pCPU and achieves a significant performance im-provement.