CN104102528A - Quick performance restoration method of virtual machine monitor - Google Patents

Abstract

The invention provides a fast virtual machine monitor performance recovery method, the steps are as follows: step 1, memory-based VM state passivation; step 2, VMM micro-restart; step 3, VM activation. The method provided by the invention can quickly clear the internal state of the VMM, reduce the impact on other VMs due to frequent disk access and a large amount of bandwidth occupation, and prevent the restarted system from entering the crash state again due to inconsistent data.

Description

Fast way to recover virtual machine monitor performance

technical field

The invention belongs to the technical field of performance recovery, in particular to a fast virtual machine monitor performance recovery method.

Background technique

The phenomenon of software decay refers to the state degradation or performance degradation of a software system that continues to run for a long time, eventually leading to system crashes. Software degradation often manifests itself in the form of slow leakage of system resources, unreleased file locks, and data corruption. Over time, system error states gradually accumulate, resulting in system performance degradation or instantaneous failure.

In recent years, in order to effectively meet the needs of complex Internet-oriented applications for large-scale computing power, massive data processing and information services, the international academic and industrial circles have begun to use virtualization technology to realize the sharing and availability of resources in heterogeneous environments. Management and collaboration, and supports large-scale deployment, migration, and operation and maintenance of applications. Virtualization technology shields the dynamics, distribution and heterogeneity of the hardware platform by introducing a virtual layer between software and hardware, supports the sharing and multiplexing of hardware resources, and provides independent and isolated computing for each common user. environment while providing administrators with centralized management of hardware and software resources. Referring to Figure 1, the virtual machine monitor VMM (Virtual Machine Monitor) is the middle layer software that implements virtualization technology. Commonly used VMMs include Xen and VMware. With the help of virtualization technology, a virtual layer is introduced between software and hardware. This layer shields the dynamics, distribution and heterogeneity of the hardware platform, supports the sharing and multiplexing of hardware resources, and provides independent, An isolated computing environment while providing administrators with centralized management of hardware and software resources. Usually, the independent operating environment provided by VMM for applications is called a virtual machine VM (Virtual Machine). For example, the Linux system and applications running in VMware constitute a VM, and the operating system in the VM is called a guest operation. System GOS (Guest Operating System), VMM is called the host of VM. For each VMM, it can host multiple VMs, and each VM is isolated from each other. However, for any software, it is inevitable to experience software decay when it is running, and VMM is no exception. Because the VMM is the host of the VMs, when the VMM performs performance recovery, the state of the VMs will be lost, resulting in the failure of the services provided by the VM and increasing the failure time.

In order to cope with system performance degradation or even failure due to software degradation, researchers have proposed many methods for restoring system performance, such as Restart, Microreboot, Checkpointing, Rio, RootHammer, Otherworld, Recovery Box, etc., to improve system performance , to increase system reliability and availability.

Restarting Restart (B. Randell. System Structure for Software Fault Tolerance. IEEE Transactions on Software Engineering, SE-1(2): 220-232, 1975.) is the simplest and most effective means of performance recovery, often used in operations System and process restarts. Among them, operating system restart is the most common method for solving operating system runtime problems. However, operating system restart involves shutting down the runtime system and reloading the operating system kernel, which requires a lot of time overhead and will cause running The application process terminates unexpectedly causing data loss. The status of the VMM in the present invention is equivalent to that of the operating system in the traditional computing system. If the restart technology is directly applied to the VMM, the state of the VM will be lost.

Microreboot (G. Candea, S. Kawamoto, Y. Fujiki, G. Friedman, and A. Fox, Microreboot-A Technique for Cheap Recovery, 6th Symposium on Operating Systems Design and Implementation, San Francisco, CA, pp.31-44, 2004) is a recursive restart technology. The core idea of this technology is to restart the fine-grained software components first. If the software performance fails to recover after restarting, restart the coarser-grained software components, and so on until To achieve the purpose of restoring performance. The present invention only considers that the VMM has recovery requirements, and considers that the VM carried by the VMM is robust, so there is no need to apply the recursive restart technology, but it is necessary to ensure that the VM state is not lost when the VMM is restarted.

By using restart technology and Microreboot technology, the system returns to a healthy state from an invalid state after restarting. However, due to the loss of a large amount of data and state, many applications and services are interrupted, and the performance is greatly reduced.

In order to solve the problem of losing the state of the components carried by the software components to be restored during the recovery process, Checkpointing and Rio file caching methods are proposed. Checkpointing (S.Feldman and C.Brown.IGOR: A System for Program Debugging via Reversible Execution.InProceedings of the Workshop on Parallel and Distributed Debugging, pages112–123, 1989) is applied to the recovery process, the basic idea is to treat recovery software Before the component is restored, the Checkpointing technology is used to save the state of the carried component, but after the restoration, the state of the carried component is reloaded to avoid the loss of the state of the component. Rio File Cache Method (P.Chen, W.Ng, S.Chandra, C.Aycock, G.Rajamani, and D.Lowell, The Rio File Cache: Surviving Operating System Crashes, in Proc.Int'l Conf.ASPLOS, 1996 , pp.74–83.) can save old file caches. When the system crashes, Rio can save the dirty file cache to the hard disk, and can prevent any modification or corruption of the file due to restart. Rio depends on operating system and hardware, and some operations cannot be supported by hardware, while the method of the present invention depends on VMM and is independent of hardware. In addition, Rio focuses on the reliability of the system, which has little effect on improving system performance after restarting.

Both Checkpointing and Rio file caching methods are based on external memory to save component state, which requires more time overhead, so researchers consider memory-based state preservation methods, such as Otherworld (M.Le, A.Gallagher, and Y.T amir, Challenges and Opportunities with Fault Injection in Virtualized Systems, First International Workshop on Virtualization Performance: Analysis, Characterization and Tools, Austin, TX, April 2008). Otherworld allows the Linux kernel to recover from a failed state while saving the state of running processes. The new kernel starts in a reserved memory space, so the content of the failed system's memory space is preserved. In Otherworld, a thread creates a new process descriptor and copies the original value from the old memory area. However, the repaired kernel components need to pass many complex data structures, which is very likely to damage the kernel. Although the present invention also adopts the idea of memory-based VM state preservation, it makes full use of the memory space allocated by the original VM, and can conveniently identify the modified memory pages through the P2M mapping table to narrow the range of memory pages that need to be reserved, and the required additional memory overhead very small.

The Recovery Box (M.Baker and M.Sullivan.The Recovery Box:Using FastRecovery to Provide High Availability in the UNIX Environment.In Proceedings of the Summer USENIX Conference,pages31–44,1992) only saves the state of the operating system and applications, and Save the state in the unused memory space, and reload these states after recovery, but it also requires hardware support to achieve the purpose of not losing the memory state after restarting. The method of the present invention relies on VMM, is independent of hardware, and can Save the complete VM state.

Contents of the invention

The purpose of the present invention is to provide a virtual machine monitor performance recovery method that integrates VM passivation and activation and VMM micro-restart to implement fast.

The technical solution that realizes the object of the present invention is: a kind of fast virtual machine monitor performance restoration method, and the steps are as follows:

Step 1, memory-based VM status passivation, specifically:

Save all state of each VM in the VMM to a reserved memory block before implementing VMM performance recovery;

Step 2, VMM micro-restart, specifically:

Save the heap information, static data segment and memory page information of VMM before restart, load the original kernel image to realize system restart, and solve the problem of consistency of state before and after restart;

Step 3, VM activation, specifically:

Create a new VM according to the configuration information of each VM, and reload all the states of the VM saved in the reserved memory block into the new VM.

Compared with the prior art, the present invention has significant advantages: (1) before the VMM implements performance recovery, the state information of the passivation VM, especially the used memory pages, can shorten the process of rebuilding the VM and read files from the file system The time to fill the memory page, thereby avoiding the impact on other VMs due to frequent disk access and a large amount of bandwidth; (2) The micro-restart method can not only quickly clear the internal state of the VMM, but also integrate the old VMM and The VMs state method quickly restores system performance, which greatly reduces space and time costs; (3) After the system restarts, by judging the inconsistency inducing factors, it uses different mechanisms to solve the consistency problem in a targeted manner, so that The restarted system will not enter the crash state again due to inconsistent data, thus ensuring the recovery success rate.

The present invention will be described in further detail below in conjunction with the accompanying drawings.

Description of drawings

Figure 1 is a scene diagram of a single server virtualization system;

Fig. 2 is a flow chart of the fast virtual machine performance recovery method of the present invention;

Fig. 3 is a process diagram of VM state passivation;

Fig. 4 is a flow chart of VMM micro-restart;

Figure 5 is a memory structure diagram of the VMM.

Detailed ways

In conjunction with Fig. 2, the present invention provides a kind of quick virtual machine performance restoration method, and the steps are as follows:

Step 1, memory-based VM status passivation, specifically:

Save all state of each VM in the VMM to a reserved memory block before implementing VMM performance recovery;

Step 2, VMM micro-restart, specifically:

Save the heap information, static data segment and memory page information of VMM before restart, load the original kernel image to realize system restart, and solve the problem of consistency of state before and after restart;

Step 3, VM activation, specifically:

Create a new VM according to the configuration information of each VM, and reload all the states of the VM saved in the reserved memory heap into the new VM.

Combined with Figure 3, the specific process of step 1 is as follows:

Step 1.1, obtain the number n of VMs carried by the VMM, the configuration information of each VM and the state of each VM through the VM state management interface provided by the VMM, set VM_LIST to represent the list of VMs carried by the VMM, and record it as VM_LIST={VM ₁ ,VM ₂ ,...,VM _n }, for any VM _i (1≤i≤n), its state space is S={s ₁ ,s ₂ ,s ₃ }, where s ₁ means the VM is in the running state , s ₂ means it is in sleep state and s ₃ means it is in shutdown state; the configuration information of VM includes the number of virtual CPUs, the number of virtual network adapters, the size of virtual memory and the size of virtual disk space, etc.; if VM _i is in s ₁ or s ₂ , Then the VMM reserves a corresponding memory heap for VM _i , and allocates additional memory units to store the configuration information and execution status of VM _i ; if VM _i is in s ₃ , it is denoted as r∈[1,n];

Step 1.2, select the running state and dormant Any VM _j in , where p,q∈[1,n], ${\mathrm{VM}}_j\∈\{{\mathrm{VM}}_p^{{\mathrm{the\; s}}_1},{\mathrm{VM}}_q^{{\mathrm{the\; s}}_2}\},$ initialize j = 1;

In step 1.3, the VMM sends a suspend command to VM _j , that is, the suspend command. Once VM _j receives the suspend command, the suspend handler of VM _j calls the detach interface of the front-end driver to unload all virtual devices that have been loaded on VM _j ; Each VM has a front-end driver, which is responsible for managing all virtual devices on the VM, such as virtual disks, virtual network adapters, etc.; once all virtual devices are unloaded, the suspend handler of VM _j initiates a suspend hypercall to VMM;

In step 1.4, the VMM reserves a corresponding memory block for the VM _j that initiates the suspend hypercall to save the P2M mapping table and cache memory pages, wherein the P2M mapping table maintains the mapping relationship between the physical address managed by the VM memory and the machine address managed by the VMM, The VMM is responsible for managing the P2M mapping table; each memory page managed by the VM is called a physical memory page, and each page has a unique number called PFN (physical frame number), and each memory page managed by the VMM is called The machine memory page also has a unique number, which is called MFN (machine frame number), so each entry in the P2M mapping table is a PFN-to-MFN mapping. Once VM _j allocates a memory page, the PFN-to-MFN mapping relationship corresponding to the memory page is stored in the P2M mapping table; cached memory pages are those memory pages used by the VM, which can be calculated by querying the P2M mapping table. The specific method is Traversing the P2M mapping table, the PFN in each mapping table entry is the PFN of the cache memory page, and these cache memory pages need to be preserved during the recovery operation of the VMM; the machine memory page itself corresponding to the P2M mapping table and the cache memory page The VMM is responsible for the management, so the preservation of these two parts of the VM state does not require additional memory space, and the VM _j is directly closed;

Step 1.5, obtain the state of VM _j again through the VM state management interface provided by VMM, if VM _j is not in the closed state, then re-execute steps 1.3-1.4; if the state is closed, set j←j+1, repeat step 1.3 -1.4, until the VM status management interface provided by the VMM obtains the status of each VM _j is in the closed state, and step 1 ends.

Combined with Figure 4, the specific process of step 2 is as follows:

Step 2.1, keep the state of VMM; the memory structure of VMM is shown in Figure 5, the data preserved in the VMM micro-restart process includes the VMM static data segment, VMM heap information and VMM machine memory page information, where the static data segment contains Key information required in the new VMM startup process, such as interrupt descriptors, pointers to VMM heap memory structures, pointers to VM memory structures and VM lists, the specific method is:

Step 2.1.1, modify the VMM link program linker, allocate the reserved memory space to the bss segment to realize the expansion of the bss segment, so that the increased memory space is enough to accommodate the data of the bss segment and static data segment of the VMM before restarting, and set Copy the data of the bss segment and the static data segment of the VMM before restarting to the expanded bss segment of the new VMM;

Step 2.1.2, modify the VMM startup program bootup, before the memory heap of the new VMM is allocated, identify unused memory pages from the memory heap of the VMM before restarting by traversing, and add the unused memory pages To the end of the memory heap, which keeps the previously used memory pages unchanged before and after restart; in this way, the state information of the VM stored in the VMM memory heap will not be lost before and after restart;

Step 2.2, in order to ensure that the state of VMM and VM _j is preserved, a fast VMM loading method is designed using the kexec mechanism provided by the Linux kernel. This method can avoid the loss of memory state due to hardware reset during system restart. The specific process is as follows: The fast VMM loading method xexec that implements the kexec mechanism in the VMM is responsible for loading the original kernel image of the VMM and the kernel of the privileged virtual machine VM ^* , and allocating initial memory and disk space for the privileged virtual machine VM ^* ; the privileged virtual machine VM ^* Cooperate with VMM to manage other virtual machines VM ₁ , VM ₂ ,...,VM _n of VMM;

Step 2.3, restart the VMM and complete the initialization, copy the data in the static data segment in the extended bss segment to the new static data segment after the restart, these data include the privileged virtual machine pointer, the storage free allocation table pointer, and the P2M mapping table pointer , time event object pointer, domain list pointer, hash table pointer, system time variable, IRQ description;

Step 2.4, eliminate the state inconsistencies between the old and new VMMs, between the VMM and the VM, and between the VMM and the hardware, that is, eliminate the state inconsistencies between the old and new VMMs, between the VMM and the VM, and between the VMM and the hardware:

The state inconsistency between the old and new VMMs is caused by partial updates of key data structures, unreleased locks, and memory leaks caused by VMM failures; in view of the above three causes of inconsistencies, different relative measures are taken to eliminate state inconsistencies. The specific methods are as follows : For partially updated data structures, use the log information recorded before the VMM failure to rebuild the data structure of the new VMM; for unreleased locks, track all locks and signals and re-initialize; for memory leaks, after restoring the VMM, the system recycles leaked storage space;

The state inconsistency between VMM and VM is caused by non-monotonically increasing system time, partially executed hypercalls, and undelivered virtual interrupts; for the above three causes of inconsistency, different relative measures are taken to eliminate state inconsistency; specifically The method is as follows: for the problem of non-monotonically increasing system time, save the VMM time structure before the VMM crashes, and restore the time structure before the VMs run after the VMM restarts to ensure that the system time is consistent with the time of the VMs; for partially executed hypercalls, Obtain the hypercall execution information through the log, and re-execute the incomplete hypercall; for the undelivered virtual interrupt, the timeout mechanism is used to resend the virtual interrupt request by the timeout handler.

Resolve inconsistencies between the VMM and hardware by resetting the I/O controller or responding to all pending interrupts.

For specific methods to eliminate state inconsistencies between old and new VMMs, between VMMs and VMs, and between VMMs and hardware, see the following articles:

Candea G, Kawamoto S, Fujiki Y, et al.Microreboot-A Technique for CheapRecovery[C]//OSDI.2004,4:31-44.

Chen P M, Ng W T, Chandra S, et al.The Rio file cache: Surviving operating system crashes[M].ACM,1996.

Bailey K, Ceze L, Gribble S D, et al. Operating system implications of fast, cheap, non-volatile memory[C]//Proceedings of the13th USENIX conference on Hottopics in operating systems. USENIX Association, 2011:2-2.

Kourai K.Cachemind: Fast performance recovery using a virtual machine monitor[C]//Dependable Systems and Networks Workshops(DSN-W),2010International Conference on.IEEE,2010:86-92.

Park E, Egger B, Lee J. Fast and space-efficient virtual machine checkpointing[C]//ACM SIGPLAN Notices.ACM,2011,46(7):75-86.

The specific process of step 3 is as follows:

Step 3.1, traverse VM_LIST, select any VM _l , initialize variable l=1, 1≤l≤n;

Step 3.2, if VM ₁ is in running state or dormant state, then according to the configuration information of VM ₁ and the P2M mapping table, allocate the memory space specified by the configuration information for VM ₁ , and go to step 3.3; if VM ₁ is in a closed state, then go to step 3.5 ;

Step 3.3, load the virtual device according to the configuration information, and establish communication with the front-end driver;

Step 3.4, start VM ₁ , and update the P2M mapping table by VMM, set the execution state of VM ₁ ;

Step 3.5, l←l+1; if l≤n-1, go to step 3.2; otherwise, end step 3.

Claims (4)

1. a virtual machine monitor method for restoring performance fast, is characterized in that, step is as follows:

Step 1, the vm health passivation based on internal memory, is specially,

Before implementing VMM performance recovery, all states of each VM in VMM are saved in reserved memory block;

Step 2, VMM is micro-restarts, be specially,

Heapinfo, static data section and the page information of front VMM is restarted in preservation, is written into original kernel reflection and realizes system and restart, and solves the consistency problem of state before and after restarting;

Step 3, VM activates, is specially,

Create new VM according to the configuration information of each VM, all states that are kept at the VM in the reserved heap of VMM are re-loaded in new VM.

2. the method for restoring performance of virtual machine monitor fast according to claim 1, is characterized in that the detailed process of step 1 is as follows:

Step 1.1, the vm health management interface providing by VMM obtains the configuration information of number n, each VM and the state of each VM of the VM of VMM carrying, is designated as VM_LIST={VM ₁, VM ₂..., VM _n, VM_LIST is the VM list of VMM carrying, and is each in run mode with dormant state reserved corresponding memory block is stored its corresponding configuration information and executing state, wherein p, q ∈ [1, n];

Step 1.2, chooses run mode with dormant state in any one VM _j, ${\mathrm{VM}}_j\∈\left\{{\mathrm{VM}}_p^{s_1}{,\mathrm{VM}}_q^{s_2}\right\},$ Initializing variable j=1;

Step 1.3, VMM is to VM _jsend suspend instruction, each VM _jdetach interface unload all virtual units that loaded, unloaded after each VM _jsuspend handling procedure initiate suspend hypercalls to VMM;

Step 1.4, VMM is the VM that initiates suspend hypercalls _jpreserve P2M mapping table and buffer memory page, close VM _j;

Step 1.5, the vm health management interface again providing by VMM obtains VM _jstate, if VM _jnot in closed condition, re-execute step 1.3-1.4; If state, for closing, makes j ← j+1, repeating step 1.3-1.4, until the vm health management interface that VMM provides obtains each VM _jstate all in closed condition.

3. the method for restoring performance of virtual machine monitor fast according to claim 1, the detailed process of step 2 is as follows:

Step 2.1, preserves heapinfo, static data section and the page information of restarting front VMM, and concrete grammar is:

Step 2.1.1, the chain program linker of amendment VMM, the bss section of expansion VMM, will restart the bss section of front VMM and the data Replica of static data section to bss section after the expansion of new VMM;

Step 2.1.2, the start-up routine bootup of amendment VMM, identification is restarted in the heap of front VMM without the page using and is added to heap afterbody;

Step 2.2, based on the quick VMM loading method of kexec mechanism, loads original kernel reflection and the franchise virtual machine VM of VMM ^*kernel, and be franchise virtual machine VM ^*distribute initial internal memory and disk space;

Step 2.3, restarts VMM and completes initialization, and data in the static data section in the bss section of expansion are copied to and are restarted in rear new static data section;

Step 2.4, eliminates between new and old VMM, state between VMM and VM and between VMM and hardware is inconsistent.

4. the method for restoring performance of virtual machine monitor fast according to claim 1, is characterized in that the detailed process of step 3 is as follows:

Step 3.1, traversal VM_LIST, chooses any VM _l, initializing variable l=1,1≤l≤n;

Step 3.2, if VM _lin run mode or dormant state, according to VM _lconfiguration information and P2M mapping table be VM _lthe memory headroom that assignment configuration information is specified, goes to step 3.3; If VM _lin closed condition, go to step 3.5;

Step 3.3, loads virtual unit according to configuration information, and communicating by letter between foundation and front-end driven program;

Step 3.4, starts VM _l, and upgrade P2M mapping table by VMM, VM is set _lexecuting state;

Step 3.5, l ← l+1; If l≤n-1, forwards step 3.2 to; Otherwise, end step 3.