Research Of Virtual Machine Live Migration Based On Xen

Live migration of virtual machine(VM) is one key technology of virtualization, live migration can make VM continued to provide non-stop services, while being migrated from one physical host to another by a complete and seamless way. Memory migration becomes most complex object to be send due to the continuous modification by applications during live migration. At present, pre-copy is a mainstream method in memory live migration. In pre-copy, memory pages are frequently modified by the source VM, so these dirty pages should be resend to the target host. That causes a long total migration time and a decline efficiency of migration. So to predict writable working set(WWS) of memory pages is a key issue to avoid to resend dirty pages and accelerate the migration process.To analyze the framework of live migration and the source code of pre-copy in Xen, we propose a weighted forecast algorithm based on time series(WFTS) combining Exponentially Weighted Moving Average(EWMA) to decide dirty pages to be resend. Based on the principle of locality of the operating system and by adopting the idea of EWMA, WFTS use memory page's history change information to forecast its dirty probability in the next iteration and avoid to resend intermediate state of high dirty rate pages. By studying the source code of compression algorithm in Remus of Xen4.2.1 and combining Run Length Encoding(RLE), we design Zero Word Run Length Encoding(ZWRLE). ZWRLE algorithm compress dirty pages in the first round of the iteration. ZWRLE based on 32 bit compression unit only encode word 0 in RLE, and do not compress none-zero words. ZWRLE algorithm in the first iteration can greatly reduce redundant data of dirty pages to send.Experimental results show that WFTS can forecast effectively the trend of memory pages to be dirtied in high dirty rate environment, reduce the intermediate state of frequently dirtied memory pages to resend, and shorten the total migration time. ZWRLE can greatly reduce the transmission of redundant memory data by compressing pages. However, our test environment is in LAN without considering the influence of network, the focus of future research is how to improve our framework with combining WFTS and ZWRLE in complicated and unstable network environment.