Research On Pre-record Virtual Machine Live Migration And Related Technologies

In recent years,with the rapid development of cloud computing technology,the importance of virtualization technology has become increasingly prominent.Virtual machine(VM)live migration can transparently transfer a running VM from one physical host to another.It is the core technology for cloud computing platform to adapt to the quickly changing business needs,realize dynamic deployment and reconfiguration of resources,and solve the key issues such as load balancing and system fault tolerance.How to implement VM migration quickly and accurately,and complete the dynamic synchronization of VM memory while maintaining VM working efficiency is the key and difficult point in the research of VM live migration.Current mainstream migration methods include pre-copy,post-copy,and the migration based on checkpoint and log playback.Post-copy has the advantages of short downtime,low network transmission,and can quickly release the occupied resources on source host,which has made it a research hotspot all the time.The disadvantage is that VM needs to be resumed when memory is incomplete,which may cause too many page faults and results in VM performance degradation.At present,the most commonly used optimization way is to adjust the sending sequence of memory pages according to the location of page faults.However,this way can not guarantee the accuracy of prediction,and needs the VM resumes for a period of time before it starts functioning.Aiming at the above problems,this paper proposes a live VM migration method based on pre-recording,and deeply studies the key technologies related to it,including the migration of VM core states,remote page faults handling,acquisition of VM memory access trajectory,dynamic backup of dirty pages,optimization and adjustment of memory sending order.The main work of this paper is as follows:(1)Post-copy has excessive page faults and related optimization methods have insufficient in accuracy and timeliness.A live VM migration method named as prerecord was proposed.By pre-executing the VM on source host and recording the accessed memory pages,the memory pages that will be accessed after the VM is restored can be obtained.The adjustment of memory page migration order is completed before migration starts,thus solving the problem of timeliness.The recorded pages were preferentially migrated for use by destination host to reduce page faults.The method is implemented in Xen and the experimental results show that pre-record migration can effectively reduce page faults and improve the stability of VM operation during migration while running various types of load applications.It also has the advantages of less total network transmission,total migration time and downtime.(2)Mixed migration of page table pages and ordinary memory pages will increase the complexity of memory reconstruction.It is easy to cause the loss of association between multi-level page tables,and single page addressing may cause multiple page faults.This paper proposes a VM core states migration strategy for adding page tables to core states.While restoring VM core states,destination host restores the multi-level structure of page table,determines the storage location of memory pages.This paper also solves the problem that accessing to the unmigrated pages after the reconstruction of page table can not trigger page fault and the inability to obtain the page frame number of missing pages.Reconstructing p2m table while refactoring the page table.Analyze the structure and interrelationship between core states,and develop the migration order and recovery process of each core state when the page table is added to the core states.(3)Pre-record migration needs to record the memory pages accessed by VM during pre-execution.How to get these accessed memory pages is a problem that needs to be solved.In addition,pre-execution will modify memory pages,but the results of pre-execution will not be inherited by the destination host.It needs the pages' original content.Therefore,memory pages should be backed up before modification,and the method of backing up all memory pages is inefficient.This paper proposes a method,which using shadow page table(SPT)to get access trajectory of VM in memory,and to dynamically on demand backup the original content of dirty memory pages.The address translation records in SPT can reflect the memory pages accessed by VM.The access trajectory of the VM in the memory is obtained by cyclically saving and emptying the SPT.The function of SPT is extended to map the original content of the page and save it in the pre-allocated storage space while SPT translating the address,thus realizing the dynamic on demand backup of dirty pages.This method is implemented in Xen.The backup pages are saved in an iteratively increasing memory space to avoid reducing VM performance,and the backup pages are recorded by shadow record table to avoid duplicate backup of the same memory page.This method has the advantages of simple and efficient process,reduced data volume,easy implementation,and takes less memory space and higher execution efficiency than the full memory backup.(4)Normal page faults have differently fix procedures from remote page faults.VM monitor can not judge the migration status of the VM with page faults.Type judgment of all page faults will slow down the speed of normal-running VM fixing page faults.To solve this problem,this paper designs and implements RemotePF page table assistant to identify VM's migration status.Enable RemotePF at the beginning of migration.Page faults will be transfer to migration module for fixing.Disable RemotePF at the end of migration,and all page faults are directly fixed by OS.RemotePF is implemented in Xen and controlled by supercall.The experimental results show that RemotePF can effectively shorten the fixing time of page faults when the VM is running normally,and it will not interrupt the fixing process of page faults when the VM is being migrated.(5)The independent remote page fault request processing module is cumbersome in obtaining the VM control interface,maintaining p2m table and the sent page record table.To solve the above problems,this paper optimizes the fixing process of remote page faults at both source and destination hosts.Analyze the work that needs to be done in the process of fixing remote page faults requests.Use multi-threading to analyze,fixing and respond to page faults.Maintain control interface,p2m table,and sent page record table in a multi-threaded manner.This approach reduces the time for coordination and synchronization between modules and improves the efficiency of handling remote page faults.The paper describes the optimized processing flow of remote page faults at both ends of the migration.(6)The pre-recorded pages can be used to find out the pages with high probability of being accessed in the remaining memory pages,and migrate them preferentially to avoid the phenomenon that the number of page faults increases significantly after the pre-recorded pages are sent.This paper proposes a prediction algorithm based on linear regression to optimize memory sending sequence in pre-record migration.Using the locality principle to analyze the memory access trajectory and find out the memory region with continuity.The concepts of incremental path and sequence set are proposed to analyze the continuity of access behavior in each region.Determine whether the access records can constitute a sequence set,and then determine the extension direction and scope of the prediction window.Memory pages in the prediction window are migrated immediately following the pre-recorded pages,so the most likely accessed pages are first sent to destination host,which further reduces the number of page faults.The algorithm is implemented in Xen and experiments are carried out while VM running load application with different access rules.The experimental results show that the algorithm can further reduce page faults based on pre-record migration,improving the stability of VM operations and the performance of load applications.(7)The performance for linear regression prediction will be poor when the number of load applications is large and each application occupies a small amount of memory.This paper proposes a method of using cluster analysis to find out the memory area with high VM access probability,and gives priority to the pages in clustering results to determine the migration order.Cluster analysis algorithm is used to analyze the prerecorded pages to find out the hotspots of VM accessing.Based on the fact that memory access records can form irregular sequence series(ISS),an ISS-DBSCAN clustering algorithm is designed to reduce the time complexity.The priority of pre-recorded pages in hotspots is determined by the access time and number.According to the priority,the hotspots are divided into several neighborhoods,and pages in each neighborhood are migrated in the order of priority.ISS-DBSCAN is implemented in Xen,and the experimental results show that the algorithm can further reduce page faults based on pre-record migration,thus improving the stability of VM operation,and still has a good optimization effect in the case of a large number of parallel load processes and each application occupies a small amount of memory,which makes up for the deficiency of linear regression prediction.