A NUMA-aware Scheduler For Xen VMs Based On Load Balancing

It’s challenging to schedule resource like CPU and memory in modern multi-core computer system. In NUMA (Non-Uniform Memory Access), frequently remote memory access will reduce system performance. Content of node’s resource among virtual machines will influence performance too. This project implements a scheduler based on Xen virtual machine. This project focuses on scheduling virtual machines to balance the work load of each NUMA node and gain higher memory access speed. This project optimizes the performance of the entire system.The main work of this dissertation are as follows:1) Proposing an adaptive K-means algorithm to aggressive the VMs:We propose an adaptive K-means algorithm to classify VMs by multi-dimensions information of VMs’ load.2) Assigning VMs to node based on shared-cache and load balancing:We Schedule entire CPUs of a VM to the same NUMA node to lower LLC miss rate. We combine heavy-load VMs with light-load VMs to make balance of each NUMA node.3) Proposing a minimum cost VM migration algorithm:We calculate the minimum number of VMs need to be migration to reduce the time of migrations. We find a best node for each VMs to get least remote memory access probability.4) Implementing a scheduler:We implement a scheduler for Xen VMs based on the aggressive algorithm and load balancing and minimum cost migration algorithm.5) Test of scheduler:We run a set of test cases to exam whether the scheduler can improve the system performance or not.