Improved pre-copy algorithm using statistical prediction and compression model for efficient live memory migration

Abstract

Xen hypervisor is used to execute and migrate the guests on different architectures using a pre-copy algorithm. There are three major categories to improve pre-copy using live migration algorithms: 1) reducing dirty pages; 2) predicating dirty pages; 3) compressing memory pages. The methods based on reducing dirty pages can lead to performance degradation so the new approach called combined approach (including prediction and compression) is proposed in this paper. The prediction of dirty pages during a migration is performed using autoregressive integrated moving average (ARIMA) model. A least recently used (LRU) stack distance-based delta compression algorithm is proposed for compression model to achieve efficient virtual machine migration. The results show that ARIMA-based model is able to predict 93% in the case of high dirty pages environment. The combined approach is able to reduce 19.16% downtime and 10.76% total migration time on an average compared to Xen's pre-copy algorithm.