HASO: A hot-page aware scheduling optimization method in virtualized NUMA systems

In the situation of CPU and memory overcommit, it is inevitable that some NUMA nodes will be overloaded or hotspotted and become hot nodes, leading to the VM application performance degradation in virtualized NUMA (vNUMA) systems. However, the virtual machine monitor (VMM) can not be aware of the NUMA feature and the distribution array of hot memory pages amongst NUMA nodes effectively. Aiming at eliminating the hot nodes and load imbalance in a vNUMA system, this paper proposes a hot-page aware scheduling optimization method (HASO) and implements a HASO scheduling system. At first we monitor the NUMA node load state, find the hot nodes in which we choose the hot VMs that cause the node hotspots. Then, we predict the distribution of future hot memory pages of the hot VM, and evaluate the cost of migrating the hot pages between NUMA nodes. At last, the hot pages of hot VM with minimized migration cost are migrated to idling nodes, so as to eliminate the hot node and improve the VM application performance. In contrast to the default scheduling mechanism of VMM, our HASO scheduling method can not only improve the memory intensive benchmark cg by up to 27.06% and the benchmark stream by up to 15.63%, but also balance the load of NUMA nodes.