TWI434185B - Method and manager physical machine for virtual machine consolidation - Google Patents

Description

Virtual host consolidation method and management entity host

The disclosure relates to a virtual machine merging method and a management physical machine.

Today, due to the trend of cloud computing, the data center is more important than ever. A typical data center contains hundreds or thousands of physical machines, such as servers. Each physical host can host multiple virtual hosts running on it. Users in the data center can rent virtual hosts to execute their software applications. For data center operators, this virtual host-based architecture is easier to manage and more efficient than a direct host-based architecture. An entity host with multiple virtual hosts can serve multiple users simultaneously. It is easier to implement a resource limit available to the user via the virtual host than via the physical host. In addition, virtual hosts are more robust than physical hosts because virtual hosts can migrate to another physical host in the event of a physical host failure without disrupting the services provided to the user.

For operators running large-scale data centers, electricity bills are a heavy burden. The cost of electricity consumption accounts for up to 20% of the total operating cost of the entire hardware architecture. The reduction in clock frequency and the power management mode of the processor are not very helpful, because the power consumption of some components of the physical host (such as memory and hard disk) is not affected by the clock frequency and the power management mode of the processor. The best way to save power is to completely cut off the power to the redundant physical host.

Therefore, the present disclosure provides a virtual host merging method. The method attempts to minimize the number of required physical hosts without degrading the performance of individual virtual hosts. In other words, the method attempts to maximize the number of redundant physical hosts to be shut down by merging virtual hosts onto a minimum number of surviving physical hosts.

The disclosure also provides a management entity host for virtual host consolidation. The management entity host performs the virtual host merging method mentioned above. The managed entity host attempts to minimize the number of virtual host migrations required for virtual host consolidation. In addition, the management entity host migrates virtual hosts in parallel, thus minimizing the time required to complete all virtual host migrations.

According to an embodiment of the disclosure, the disclosure provides a virtual host merging method. The method is performed by a management entity host. A network connects the management entity host with a plurality of server physical machines. A plurality of virtual hosts are running on the server entity host. The Servo Entity Host provides N kinds of resources and the virtual host consumes resources. N is a preset positive integer. The method comprises the following steps. The management entity host classifies the server entity host as a redundant entity host and a surviving entity host, wherein the virtual host running on the redundant entity host may perform an interleaved N-dimensional matching decompression filling algorithm (interleaved N-dimensional) Best-fit decreasing packing algorithm) successfully migrated to the surviving entity host. The management entity host determines a migration path of the virtual host running on the redundant entity host to the surviving entity host according to an interleaved N-dimensional matching deprecated padding algorithm and a network topology. The management entity host determines a parallel migration sequence of the virtual host running on the redundant entity host based on the migration path. The management entity host migrates the virtual host running on the redundant entity host to the surviving entity host in parallel according to the parallel migration sequence.

According to another embodiment of the disclosure, the disclosure provides a management entity host. The management entity host includes a memory and a processor coupled to the memory. The processor executes the virtual host merging method mentioned above and stores the data generated by the steps of the method and used by the steps, such as migration paths and parallel migration sequences, in memory.

FIG. 1 is a schematic diagram showing a hardware architecture of a data center according to an embodiment of the present disclosure. In the architecture, network 110 connects management entity host 120 with a plurality of server entity hosts 130. Each servo entity host 130 can host one or more virtual hosts running thereon. Each servo entity host provides N different resources, such as processor cycles, memory, network bandwidth, and the like. N is a preset positive integer. The virtual host utilizes and consumes resources provided by the physical host. One of the goals of this embodiment is to consolidate all virtual hosts to the least number of physical hosts as much as possible and to minimize the time required to migrate virtual hosts. Turning off the power of the redundant physical host after the virtual host is merged in principle helps to reduce the power cost of the data center.

The virtual host merging of this embodiment is somewhat similar to the conventional vector bin packing problem. Each servo entity host 130 can be represented by an N-dimensional vector, each element of which corresponds to the remaining capacity of a resource provided by the servo entity host. Similarly, each virtual host can be represented by an N-dimensional vector, each element of which corresponds to the consumption of a resource of the virtual host. The input of the above virtual host merge problem is a vector indicating the host of the servo entity and the virtual host, and a detailed record of the accommodation mapping between the host entity and the virtual host. In order to solve the virtual host consolidation problem, all servo entity hosts must be classified into redundant entity hosts and surviving entity hosts, and then the virtual hosts running on the redundant entity hosts must be migrated to the surviving entity hosts. For maximum efficiency, the number of redundant physical hosts to be shut down should be maximized, and the number of virtual hosts being migrated and the time required to migrate virtual hosts should be minimized.

There are many differences between the conventional vector packing problem and the virtual host merging problem of this embodiment. First, the boxes in the conventional vector boxing problem are initially empty, and the server entity in the virtual host merge problem of this embodiment has initially been partially populated by the virtual host running on it. The initial remaining capacity of each resource of each surviving entity host may be different. Second, the virtual host consolidation of this embodiment allows for load balancing. The final remaining capacity of the surviving physical host should be as equal as possible. Third, some virtual hosts may not migrate due to their heavy load and the performance degradation caused by migration. The virtual host consolidation of this embodiment takes into account such virtual hosts. Fourth, virtual host merging of this embodiment minimizes the time required to migrate virtual hosts by parallelizing the migration sequence.

FIG. 2 is a schematic diagram of the management entity host 120 of FIG. 1. The management entity host 120 includes a memory 124 and a processor 122 coupled to the memory 124. Processor 122 performs each step and all calculations of the virtual host merging method illustrated in Figures 3-10. The processor 122 stores the data generated by the steps of the virtual host merging method mentioned above and generated in the steps in the memory 124. For example, the memory 124 can store a detailed record of the above-mentioned N-dimensional vector representing the servo entity host and the virtual host, the accommodation relationship between the servo entity host and the virtual host, and the virtual host merge mentioned above. The queues, migration paths, graphics, and parallel migration sequences produced by the method.

FIG. 3 is a flow chart showing a virtual host merging method in this embodiment of the disclosure. First, the management entity host checks if it is appropriate to trigger a virtual host merge (step 310). If it is not appropriate to trigger a virtual host merge at this time, the process ends. If it is suitable to trigger virtual host merging at this time, the management entity host classifies the servant entity host as a redundant entity host and a surviving entity host, so that the surviving entity host can accommodate all virtual hosts, and the remaining redundant entity hosts are powered off. Power is saved (step 320). The management entity host determines a migration path from the virtual host running on the redundant entity host to the surviving entity host according to an interleaved N-dimensional superior depreciation padding algorithm and network configuration (step 330). The managed entity host determines a parallel migration sequence of virtual hosts running on the redundant physical host based on the migration path (step 340). The managed entity host migrates the virtual hosts running on the redundant physical host to the surviving physical host in parallel according to the parallel migration sequence (step 350). Next, the management entity host shuts off power to the redundant physical host to save power costs (step 360). Details of the steps mentioned above will be explained later.

In step 310, a series of numbers ( M ₁ , M ₂ , ..., M _N ) are calculated for all i , where 1 <= i <= N . The number M _i is an estimated number of servo entity hosts required to accommodate the consumption of the i-th resource of all virtual hosts, calculated from a one-dimensional best-fit decreasing packing algorithm. For example, if the third of the resources is memory, the number M ₃ is the number of servo entity hosts required to accommodate the total consumption of the memory of all virtual hosts according to the one-dimensional superior depreciation filling algorithm. . Next, the number M is calculated, where M = max{ M ₁ , M ₂ , ..., M _N }. The managed entity host triggers virtual host consolidation when M * K < P , where K is a preset constant and P is the number of servo entity hosts. Since the one-dimensional superior declining padding algorithm only considers the resource capacity of a single resource of the servo entity host and the resource consumption of a single resource of the virtual host, the estimated quantity M _i is slightly inaccurate. One purpose of the constant K is to compensate for the estimation error. Another purpose of the constant K is to set a threshold for triggering virtual host merging.

4 is a flow chart of one of the above-mentioned dimensions of the preferred decremental padding algorithm corresponding to M _i . First, assume that all of the servo entity hosts are empty, with no virtual hosts running (step 410). All virtual hosts are placed in a queue and the virtual hosts in the queue are sorted in descending order of consumption of the i-th resources (step 420). Next, the first virtual host is taken from the queue (step 430). The fetched virtual host is assigned to a Serving Entity Host with a minimum i-th resource remaining capacity that can accommodate the consumption of the i-th resource of the fetched virtual host (step 440). In step 440, only the capacity of the i-th resource of the host entity and the consumption of the i-th resource of the fetched virtual host are considered, and the capacity of other resources and the consumption of other resources are completely ignored.

The initial remaining capacity of a resource of a servant entity host is equal to the full capacity of this resource of the servant entity host. In this embodiment of the disclosure, the full capacity of each resource of the servo entity host is defined as a predetermined percentage of the total amount of the same resource of the host of the servo entity. This preset percentage is not necessarily 100%. Experiments have shown that when the utilization of the resources of the servo entity host approaches 100%, the power consumption of the servo entity host will increase exponentially. For each resource of each servo entity host, there is an ideal threshold, such as 70%, 80% or 90%. As long as the resource utilization does not exceed the corresponding ideal threshold, the power consumption will not be exponential. increase. The ideal threshold for the servo entity host can be obtained experimentally. The preset percentage of each resource of each of the servo entity hosts mentioned above is equal to the ideal threshold of the same resource of the host of this servo entity.

Next, since the host of the assigned physical servo remaining capacity of the resource i of the i-th subtracting the consumption of resources of the virtual hosts taken to update the remaining capacity of the i-th servo physical machine resources of the assigned (step 450). Next, it is checked if the queue is empty (step 460). If the queue is not empty, then return to step 430. The estimated number M _i is the number of servo entity hosts for accommodating all virtual hosts in the one-dimensional matching depreciation filling algorithm corresponding to the i-th resource illustrated in FIG.

FIG. 5 is a flow chart showing the details of the classification of the servo entity host in step 320 of FIG. First, all servo entity hosts are sorted in order of resource utilization (step 510). The resource utilization of the Servo Entity Host can be defined in several ways. For example, resource utilization can be defined as the maximum of the utilization percentage of each resource of the server entity host. Alternatively, the resource utilization may be defined as the sum or average of the utilization percentages of each of the resources of the servo entity host. When using one of the resource utilization definitions mentioned above, the Serving Entity Hosts must be ordered in ascending order of resource utilization.

Alternatively, the resource utilization of the Serving Entity Host may be defined as the sum of the difference between the above-mentioned ideal threshold for each of the resources of the Servo Entity and the utilization percentage of the same resource of the Servo Entity Host. When using this resource utilization definition, the Serving Entity Hosts must be ordered in descending order of resource utilization.

There may be some virtual hosts that meet the preset "non-removable" conditions, such as very busy virtual hosts or heavily loaded virtual hosts. In this case, such virtual hosts should not be migrated, and the Serving Entity Hosts with such virtual hosts should be classified as surviving entity hosts. When there are one or more virtual hosts that meet the preset conditions, the ordering in step 510 may be slightly different, so that the sort order of the server host having at least one virtual host that satisfies the preset condition does not satisfy the preset condition. After the virtual host's server entity host. Servo entity hosts of virtual hosts that do not meet the pre-conditions are ordered in the order of resource utilization as discussed above.

Next, the variables L and R are initialized (step 520). The binary search method is used to find the division points in the sorted servo entity host, and the servo entity host is classified into a redundant entity host and a surviving entity host. The variable L represents the starting point of the binary search method search range, and the variable R represents the end point of the binary search method search range. The initial value of the variable L is the number M used in step 310 of Figure 3, and the initial value of the variable R is the number P of all servo entity hosts.

Next, it is checked if the variable R is greater than the variable L (step 530). If the variable R is less than or equal to the variable L , the flow ends. If the variable R is greater than the variable L , the average value i of the variables L and R is calculated (step 540). The fractional part of the average value i is removed such that i is an integer and 1 <= i <= P . Next, the first i of the sorted servo entity hosts are classified as redundant entity hosts, and the sorted servo entity host back faces ( P - i ) are classified as surviving entity hosts (step 550). Performing an interlaced N-dimensional superior declining padding algorithm to see if the virtual host running on the redundant entity host can be successfully migrated to the surviving entity host (step 560), and then checking the interleaved N-dimensional superior declining padding algorithm. The result (step 570). If the returned result is successful, the variable L is set to i +1 (step 580) and the process returns to step 530. If the result of the return is a failure, the variable R is set to i (step 590) and the process returns to step 530.

The binary search method in the loop of steps 530-590 can find the maximum i that can successfully migrate the virtual host running on the redundant entity host to the surviving entity host according to the interlaced N-dimensional superior decrement filling algorithm. The final partitioning of the redundant physical host and the surviving physical host is determined by the last execution of step 550. The migration path of step 330 of Figure 3 is determined by the last successful execution of the interlaced N-dimensional superior decremental padding algorithm in the loop of steps 530-590.

6 is a flow chart of the interlaced N-dimensional superior decrement filling algorithm performed by step 560 of FIG. First, a normalization factor corresponding to each resource is calculated based on the consumption of resources of all virtual hosts (step 605). Subtracting the total consumption of the i-th resources of all virtual hosts from the sum of the capacity of the i-th resources of all surviving entity hosts, and then dividing the above-mentioned subtraction result by the number of surviving entity hosts, thereby calculating a normalization factor corresponding to the i-th resource Where i is an integer and 1<= i <= N .

Next, the standardized consumption of resources for each virtual host running on the redundant physical host and the representative consumption of each virtual host running on the redundant physical host are calculated based on the normalization factor (step 607). For any virtual host running on a redundant physical host, the consumption of the i-th resource of the virtual host can be divided by the normalization factor corresponding to the i-th resource to calculate the standardized consumption of the i-th resource of the virtual host. For example, if the N-dimensional vector of resource consumption of a virtual host is ( v ₁ , v ₂ , . . . , v _N ), the standardized consumption of resources of the virtual host is ( v ₁ / n ₁ , v _{2 )} / n ₂ , ..., v _N / n _N ), where n _i is a normalization factor corresponding to the ith resource, 1 <= i < =N . The representative consumption of each virtual host running on the redundant physical host consumes the largest standardized consumption of the standardized consumption of all resources of the virtual host. For example, if the standardized consumption v ₃ / n ₃ is the normalized consumption of a virtual host, the N-dimensional vector ( v ₁ / n ₁ , v ₂ / n ₂ , ..., v _N / n _N ) , then standardize consumption v ₃ / n ₃ for this representative consumption of this virtual host. The representative consumption represents the overall resource consumption of the virtual host.

Next, each virtual host running on the redundant physical host is placed in one or more queues according to the standardized consumption of resources for each virtual host (step 610). There are a total of N in the above-mentioned queues, and there is a one-to-one mapping relationship between the N queues and the N resources. Each virtual host running on the redundant physical host is placed in a queue corresponding to the resources corresponding to the representative consumption of this virtual host. For example, if the representative consumption of a virtual host consumes the standard consumption of the second resource of the virtual host, the virtual host is placed in the second queue corresponding to the second resource of the N resources. . Each virtual host is placed in a queue.

Next, the virtual hosts in each queue are sorted in descending order of representative consumption (step 615). In other words, i-th queue by the virtual host resources corresponding to the i-th normalized consumption of v _i / n _i in descending order of the sort.

Next, check all queues to see if all queues are empty (step 620). If all queues are empty, it means that all virtual hosts running on the redundant entity host have been successfully populated into the surviving entity host, returning a successful result (step 625). Otherwise, a queue is selected from the queues according to a normalized sum corresponding to the remaining capacity of the resources (step 630).

In order to select a queue from the queue, the processor of the management entity host divides the sum of the remaining capacity of all the surviving entity hosts of each resource by the normalization factor corresponding to the resource, thereby calculating corresponding to each resource. The standardized sum of remaining capacity. For example, corresponding to the i normalized sum of the remaining capacity of the resource is to survive the remaining capacity of the i-th physical machine resources divided by the normalization factor corresponding to the i n _i is calculated to produce the resources. The processor managing the entity host may define a set comprising N - 1 queues, the set including the remaining N - 1 queues except for the queue corresponding to the resource corresponding to the smallest of the remaining capacity normalized sums . The processor managing the entity host selects the queue of the first virtual host with the largest representative consumption among the first virtual hosts in each of the sets.

Next, the first virtual host is taken from the selected queue (step 635). The selection criteria mentioned above ensure that the above selected queues are not empty. The staggered selection of the queues mentioned above ensures that the virtual host is balancedly populated into the surviving entity host. Next, a check is made to see if there are any surviving entity hosts that can accommodate the consumption of all resources of the fetched virtual host (step 640). If there is no surviving entity host that can accommodate the consumption of all of the resources of the fetched virtual host, then the failure result is returned (step 650). Otherwise, the fetched virtual host is assigned to a surviving physical host with the smallest representative remaining capacity consuming the consumption of all resources of the fetched virtual host (step 645) and returns to step 620.

In order to assign the fetched virtual host, the processor managing the physical host must calculate the representative remaining capacity of each surviving physical host. The representative remaining capacity of each surviving entity host is the minimum of the standardized remaining capacity of all resources of this entity host. To calculate the standardized remaining capacity of a resource of an entity host, the remaining capacity of this resource of the entity host can be divided by the normalization factor corresponding to the resource. For example, if the N-dimensional vector of the remaining capacity of a resource of a surviving entity host is ( p ₁ , p ₂ , . . . , p _N ), the normalized remaining capacity of the resource of the entity host is ( p ₁ / n ₁ , p ₂ / n ₂ , ..., p _N / n _N ), where n _i is a normalization factor corresponding to the ith resource, 1 <= i <= N . The representative remaining capacity of this entity host is the minimum of the normalized residual capacity ( p ₁ / n ₁ , p ₂ / n ₂ , ..., p _N / n _N ). The virtual host fetched in step 635 is assigned to a surviving physical host with the smallest representative remaining capacity that can accommodate the consumption of all resources of the fetched virtual host. This optimization strategy helps balance the load distribution of virtual hosts on surviving physical hosts. The processor managing the entity host subtracts the resource consumption of the fetched virtual host from the remaining capacity of the resource of the assigned entity host to update the remaining capacity of the resource of the assigned entity host.

The interleaved N-dimensional superior declining padding algorithm of Figure 6 has two alternative implementations. The first alternative has been discussed above. The second alternative is described below.

The second alternative is different from the first alternative in steps 610, 630, and 635. In step 610, there are a total of N *( N -1) queues. Each queue corresponds to a different arrangement of two of the N resources. For example, when N = 3 and the resource is processor cycle, memory, and network bandwidth, any two of the three resources have six different permutations, namely, {processor, memory}, {processor, Network}, {memory, processor}, {memory, network}, {network, processor} and {network, memory}. There are six queues in the example mentioned above, and there is a one-to-one mapping relationship between the six queues and the six arrays. For each virtual host running on the redundant entity host, as long as the standardized consumption of the i-th resource of the virtual host is greater than the standardized consumption of the j-th resource, it is placed corresponding to the first i-th resource and then the j-th resource. Arranged in the queue. Where i and j are integers, 1<= i <= N , 1<= j <= N and i is not equal to j . For example, if the standardized consumption of the network bandwidth of a virtual host is greater than the standardized consumption of the memory of the virtual host, the virtual host is placed in a queue corresponding to the {network, memory} arrangement. In a second alternative, one virtual host can be placed in multiple queues because one virtual host may satisfy multiple queue criteria at the same time.

In step 630, the processor of the management entity host selects a queue corresponding to the resource corresponding to the largest remaining capacity normalized sum and then the resource corresponding to the smallest remaining capacity normalized sum. For example, in the sum of the remaining capacity of each resource, if the sum of the remaining capacity of the processor cycle is the largest and the sum of the remaining capacity of the network bandwidth is the smallest, the processor selection of the managed entity host corresponds to { Processor, network} array of queues. The interleaving selection of the queue in the second alternative also ensures that the selected queue is not empty.

In step 635, in addition to taking the first virtual host from the selected queue, the processor of the management entity host also removes the extracted virtual host from all queues, because the extracted virtual host may also exist in other ports. In the column.

The other steps in Figure 6 are the same in both the first alternative and the second alternative.

The migration path from the virtual host to the surviving entity host running on the redundant physical host is determined by the last successful execution of the interlaced N-dimensional superior depreciation padding algorithm performed in step 560 of FIG. Each migration path corresponds to one of the virtual hosts running on the redundant physical host. The starting point of each migration path is a redundant entity host on which the corresponding virtual host runs, and the destination of the migration path is the corresponding successful execution of the interleaved N-dimensional matching deprecated padding algorithm corresponding to the virtual host. Survival entity host. The rest of each migration path can be determined based on the network form. For example, the remainder of each migration path may pass through routing elements of the network, such as switches and routers, or even through other physical hosts connected by the network.

FIG. 7 is a flow diagram of determining a parallel migration sequence in step 340 of FIG. 3 in accordance with an embodiment of the present disclosure. First, a conflict graph is defined (step 710). Each node of the conflict graph represents one of the migration paths. Each node of the conflict graph has a weight that represents the estimated time required to migrate a virtual host corresponding to a particular node to the corresponding surviving entity host. Since the time required to migrate the virtual host is mainly spent on the memory content of the transfer virtual host, the time required to migrate the virtual host can be estimated based on the memory consumption of the virtual host and the network bandwidth of the migration path. If there is any partial overlap of the two migration paths corresponding to the two nodes in the conflict graph, there is a connection between the two nodes. The connection in the conflict graph means that the two virtual hosts corresponding to the two nodes at the end of the connection cannot migrate in parallel due to the overlapping portion of their migration paths.

Next, use the maximum vertex-coloring algorithm on the conflict graph to find a coloring method that includes all nodes, so that the sum of the weights of the colors of this coloring method is among all possible coloring methods. Minimal (step 720). The weight of each color is defined as the maximum of the weights of all nodes of this color. For example, if some nodes in the conflict graph are painted yellow by the maximum node shading algorithm, the weight of yellow is the maximum weight of all yellow nodes. Since the two nodes at the end of the same line are always painted in different colors by the maximum node shading algorithm, the virtual hosts corresponding to the nodes of the same color can be migrated in parallel because there is no conflict in their migration path.

Next, the migration path corresponding to the node of each color is added to the parallel migration sequence as a batch migration path (step 730). In other words, the migration paths corresponding to the nodes of the same color are added to the parallel migration sequence as a batch migration path. The migration path corresponding to the node of another color is added to the parallel migration sequence as another batch migration path, and so on.

The flow in Figure 7 can be used to determine parallel migration sequences with minimal migration times. However, the computational complexity of step 720 is NP-hard. A data center may have thousands of virtual hosts waiting to be migrated, and step 720 may take too much time in this case. In order to speed up the decision of the parallel migration sequence, this problem must be simplified.

FIG. 8 is a flow diagram of determining a parallel migration sequence in step 340 of FIG. 3 in accordance with another embodiment of the present disclosure. In this embodiment, the problem is simplified by assuming that the time required to migrate each virtual host is the same. First, a directed graph is defined based on the migration path and network morphology (step 810). The nodes of the directed graph contain redundant physical hosts and surviving physical hosts, and, if necessary, routing elements for networks such as switches and routers. The connection in the above directed graph is defined by the above migration path.

Next, the maximum set of disjoint paths are found in the directed graph (step 820). This is a well-known problem in algorithmic graph theory. There are many conventional algorithms that can be used to find the largest set of disjoint paths in a directed graph. The largest set of each disjoint path corresponds to a set of virtual hosts that can be migrated in parallel. Next, the migration path corresponding to the largest set of disjoint paths is added to the parallel migration sequence as a batch migration path (step 830). The own line map removes the line corresponding to the largest set of disjoint paths (step 840). Check to see if there is still at least one connection in the directed graph (step 850). If there is still at least one connection in the directed graph, then return to step 820.

FIG. 9 is a flow chart of determining a parallel migration sequence in step 340 of FIG. 3 in accordance with another embodiment of the present disclosure. In this embodiment, the problem is further simplified by assuming that the time required to migrate each virtual host is the same, and the two migration paths do not overlap as long as their origin entity hosts are different and their destination entity hosts are different. The above simplification applies to some simple networks, such as multiple servo entity hosts connected by a single switch. First, a bipartite graph is defined based on the migration path (step 910). The nodes of this bipartite graph are redundant entity hosts and surviving entity hosts. If there is a migration path from a redundant physical host to a surviving physical host, there is a connection between the redundant physical host and the surviving physical host.

Next, a maximum matching is found in the bipartite graph (step 920). This is also a well-known problem in the calculation of graph theory. There are many conventional algorithms that can be used in bipartite graphs to find the largest match. Each maximum match corresponds to a set of virtual hosts that can be migrated in parallel. Next, the migration path corresponding to this maximum match is added to the parallel migration sequence as a batch migration path (step 930). The line corresponding to this maximum match is removed from the bipartite graph (step 940). Check to see if there is still at least one connection in the bipartite graph (step 950). If at least one connection still exists in the bipartite graph, then return to step 920.

FIG. 10 is a flow chart showing parallel migration in step 350 of FIG. 3 in accordance with an embodiment of the present disclosure. The parallel migration may be based on the parallel migration sequence determined by the flow of FIG. 7, FIG. 8, or FIG. As discussed above, a parallel migration sequence contains one or more batch migration paths. First, the first batch of migration paths in the parallel migration sequence is selected (step 1010). The virtual hosts corresponding to the selected batch are migrated in parallel to the surviving entity host according to the migration path of the selected batch (step 1020). Next, a check is made to see if there is at least one migration path after the current selected batch (step 1030). If there is at least one migration path after the current selected batch, the next migration path in the parallel migration sequence is selected and returned to step 1020 (step 1040).

11A through 11H illustrate an example of determining a parallel migration sequence according to the flow of FIG. As shown in FIG. 11A, this example includes four redundant physical hosts, three surviving physical hosts, and ten virtual hosts to be migrated. Each of the wires of the bipartite graphs of FIGS. 11A to 11H corresponds to one migration path. First, the migration path corresponding to the three connections shown in FIG. 11B is added as a first batch migration path to the parallel migration sequence. Next, three connections are removed as shown in FIG. 11C. Next, the migration path corresponding to the three connections shown in FIG. 11D is added as a second batch migration path to the parallel migration sequence. The three connections are removed as shown in Figure 11E. Next, the migration path corresponding to the two lines shown in FIG. 11F is added as a third batch migration path to the parallel migration sequence. Remove the two wires as shown in Figure 11G. Finally, the migration path corresponding to the last two connections shown in FIG. 11H is added as a fourth batch migration path to the parallel migration sequence.

In summary, the method for managing virtual host merging provided by the above embodiments and the management entity host attempt to maximize the number of redundant physical hosts that can be shut down without degrading the performance of the individual virtual hosts. The method and management entity host attempt to minimize the time required to complete all virtual host migrations by parallelizing the migration sequence. The method and management entity host uses a binary search method to find the optimal number of surviving entity hosts. The method and the management entity host classify the server entity with light load as a redundant entity host to minimize the number of virtual hosts to be migrated. In addition, the method and the management entity host populate the virtual host into the surviving entity host in an interleaved match decrement manner to achieve better merge results, such as load balancing.

The present disclosure has been disclosed in the above embodiments, but it is not intended to limit the disclosure, and any one of ordinary skill in the art can make some changes and refinements without departing from the spirit and scope of the disclosure. The scope of protection of this disclosure is subject to the definition of the scope of the patent application.

110. . . network

120. . . Management entity host

122. . . processor

124. . . Memory

130. . . Servo entity host

310-360. . . Process step

410-460. . . Process step

510-590. . . Process step

605-650. . . Process step

710-730. . . Process step

810-850. . . Process step

910-950. . . Process step

1010-1040. . . Process step

FIG. 1 is a schematic diagram showing a hardware architecture of a data center according to an embodiment of the present disclosure.

2 is a schematic diagram of a management entity host for virtual host merging according to an embodiment of the disclosure.

FIG. 3 is a flow chart showing a virtual host merging method according to an embodiment of the present disclosure.

4 to 10 are more detailed flowcharts of a virtual host merging method according to an embodiment of the present disclosure.

11A-11H are schematic diagrams showing an example of generation of a parallel migration sequence according to an embodiment of the present disclosure.

310-360. . . Process step

Claims (30)

A virtual host merging method is performed by a management entity host, wherein a network connects the management entity host and a plurality of server entity hosts, and multiple virtual hosts run on the server entity host, and the server entity provides N And the virtual host consumes the resource, where N is a preset positive integer, the method includes: the management entity host classifies the server entity host as a redundant entity host and a surviving entity host, where The virtual host running on the redundant entity host may be successfully migrated to the surviving entity host according to an interlaced N-dimensional optimal depreciation filling algorithm; the management entity host is dequantized and filled according to the interlaced N-dimensional matching And a form of the network determining a migration path of the virtual host running on the redundant entity host to the surviving entity host; the management entity host determining the redundant entity based on the migration path a parallel migration sequence of the virtual host running on the host; and the management entity host will be in the I run it on the physical host virtual hosts in parallel migrate to the surviving entity host. The method of claim 1, further comprising: estimating, according to the one-dimensional matching deprecated padding algorithm, the servant host required to consume the ith resource of all the virtual hosts for each i quantity Mi, where 1 <= i <= N; and when M * K <P triggered when the combined virtual host, where M is the maximum of all of the M _i, K is a constant and P is the preset servo The number of physical hosts. The method of claim 2, wherein the one-dimensional matching depreciation filling algorithm corresponding to the M _i comprises: assuming that all of the servo entity hosts are empty, wherein there is no virtual host; The virtual host is placed in a queue, and the virtual hosts in the queue are sorted in descending order of consumption of the i-th resource; the first virtual host is taken out from the queue; Assigning the fetched virtual host to the server entity host having the minimum remaining capacity of the i-th resource that can accommodate the consumption of the i-th resource of the fetched virtual host; the servo assigned the i th entity of host resources of the remaining capacity by subtracting the sum of the i-th virtual host of the resource consumption removed, thereby updating the first entity as host of the assigned servo The remaining capacity of the i resource; and returning the first virtual host from the queue if the queue is not empty, wherein M _i is used in the one-dimensional optimal depreciation filling algorithm To accommodate all of the virtual hosts The number of servo entity hosts. The method of claim 1, wherein the step of classifying the servo entity host comprises: sorting all the server entities in order of resource utilization; and prior to the sorted server entity host i Classified as the redundant entity host, and classify the sorted server entity host ( P - i ) into the surviving entity host, where i is an integer and 1<= i <= P , P is a number of all of the servo entity hosts; and finding a maximum of the virtual host that can be successfully migrated to the surviving entity host on the redundant entity host according to the interleaved N-dimensional matching deprecated padding algorithm i . The method of claim 1, wherein the step of classifying the servo entity host comprises: sorting all of the servo entity hosts, wherein according to the sort order, having at least one virtual meeting a predetermined condition The servo entity host of the host is located after the server entity of the virtual host that does not meet the preset condition, and the server entity of the virtual host that does not meet the preset condition is sorted according to resource utilization order; after the previously ordered the servo host entity i category entity to the redundant host, and the host entity behind the ranking after the servo (P - i) th entity is classified as the host survival, where i is An integer and 1 <= i <= P , P is the number of all of the servo entity hosts; and finding the operation that can be run on the redundant entity host according to the interlaced N-dimensional superior decrement filling algorithm The maximum i of the virtual host successfully migrated to the surviving entity host. The method of claim 1, wherein the interlaced N-dimensional superior declining padding algorithm comprises: calculating a standardization corresponding to each of the resources based on consumption of the resources of all the virtual hosts Generating a standardized consumption of the resources of each of the virtual hosts running on the redundant entity host and each of the virtual hosts running on the redundant entity host based on the normalization factor Representative consumption; placing each of the virtual hosts running on the redundant entity host in one or more queues according to the standardized consumption of the resources of each of the virtual hosts; Determining the descending order of representative consumptions to sort the virtual hosts in each of the queues; checking all of the queues, if all of the queues are empty, returning a successful result; according to the corresponding a standardized sum of remaining capacity of resources, selecting a queue from the queue; taking the first virtual host from the selected queue; assigning the extracted virtual host to accommodate the The surviving entity host having the smallest representative remaining capacity consumed by all of said resources of the virtual host; if there is no such consumption of the consumption of all of the resources of the fetched virtual host The entity host returns a failed result; and returns to the step of checking all of the queues. The method of claim 6, wherein the sum of the consumptions of the i-th resources of all the virtual hosts is subtracted from the sum of the capacities of the i-th resources of all the surviving entity hosts And dividing the result of the subtraction by the number of the surviving entity hosts, thereby calculating the normalization factor corresponding to the ith resource, where i is an integer and 1<= i <= N ; The consumption of the i-th resource of the virtual host is divided by the normalization factor corresponding to the i-th resource, thereby calculating each virtual host running on the redundant entity host Said standardized consumption of the i-th resource; said representative consumption of each of said virtual hosts running on said redundant entity host being the maximum standardized consumption of said standardized consumption of said resources of said virtual host ; i-th resource of the survival of the remaining capacity of the host entity divided by the normalization factor corresponding to the resource of the i, thereby calculating a remaining capacity corresponding to the resource of the i-th The standardization sum. The method of claim 7, wherein there are a total of N queues, the N queues and the N resources have a one-to-one mapping relationship, and the redundancy will be in the redundancy The step of placing the virtual host running on the physical host in the queue includes: placing each of the virtual hosts running on the redundant entity host in a representation corresponding to the virtual host Consuming the queue of the corresponding resource; and selecting a queue from the queue includes: excluding the queue corresponding to the resource corresponding to the smallest of the remaining capacity normalized sum The definition includes a set of the remaining N - 1 of the queues; and among the first virtual hosts of each of the sets, the first virtual host having the largest representative consumption is selected Queue. The method of claim 7, wherein there are a total of N *( N -1) queues, each of the queues corresponding to a different arrangement of two of the N resources, and the The step of placing the virtual host running on the redundant physical host in the queue includes: for each virtual host running on the redundant physical host, when the virtual host When the standardized consumption of the i-th resource is greater than the standardized consumption of the j-th resource of the virtual host, placing the virtual host in a queue corresponding to the arrangement of the i-th resource and then the j-th resource Where j is an integer, 1<= j <= N and j is not equal to i ; and the step of selecting a queue from the queue comprises: selecting corresponding to the sum of the remaining normalized capacity of the remaining capacity The resource is then a queue of resources corresponding to the smallest standardized sum of the remaining capacity; and the step of fetching the first virtual host from the selected queue comprises: removing the location from all of the queues Take out the virtual host. The method of claim 6, wherein the step of assigning the fetched virtual host comprises: calculating the representative remaining capacity of each of the surviving entity hosts, wherein each of the surviving entity hosts The representative remaining capacity is a minimum value of the standardized remaining capacity of the resource of the surviving entity host, and the standardized remaining capacity of each of the resources of the surviving entity host is a resource of the resource of the surviving entity host Calculating the remaining capacity divided by the normalization factor corresponding to the resource; assigning the fetched virtual host to the consumption of all of the resources that can accommodate the fetched virtual host has a minimum representation The surviving entity host of the remaining capacity; and subtracting the consumption of the resource of the fetched virtual host from the remaining capacity of the resource of the assigned entity host, thereby updating the The remaining capacity of the resource of the assigned entity host. The method of claim 1, wherein each of the migration paths corresponds to one of virtual hosts running on the redundant entity host, and the starting point of the migration path is that the virtual host runs on the The redundant entity host, and the destination of the migration path is the surviving entity host corresponding to the virtual host assigned by the interlaced N-dimensional superior depreciation filling algorithm, and the rest of the migration path is based on the Determined by the form of the network described. The method of claim 11, wherein the determining the parallel migration sequence comprises: defining a conflict graph, wherein each node of the conflict graph represents one of the migration paths, each of the The node has a weight indicating an estimated time required to migrate the virtual host corresponding to the node to the corresponding surviving entity host, if any of the two migration paths corresponding to the two nodes of the conflict graph Partially overlapping, there is a connection between the two nodes; a maximum node coloring algorithm is used on the conflict graph to find out A coloring manner of the node such that a sum of weights of the colors of the coloring mode is minimal among all coloring modes, wherein the weight of each of the colors is the weight of the node of the color a maximum value; and the migration path corresponding to the node of each of the colors is added to the parallel migration sequence as a batch migration path. The method of claim 11, wherein the determining the parallel migration sequence comprises: defining a directed graph according to the migration path and the form of the network; in the directed graph Finding a largest set of disjoint paths, and adding the migration path corresponding to the largest set to the parallel migration sequence as a batch migration path; removing from the directed graph corresponding to the maximum a line connecting the sets; and if there is still at least one connection in the directed graph, returning to the step of finding the largest set. The method of claim 11, wherein the determining the parallel migration sequence comprises: defining a bipartite graph according to the migration path; finding a maximum match in the bipartite graph, and corresponding to The migration path of the maximum match is added to the parallel migration sequence as a batch migration path; the connection corresponding to the maximum match is removed from the bipartite graph; and if at least the bipartite graph still exists A connection returns to the step of finding the maximum match. The method of claim 1, wherein the step of migrating the virtual host running on the redundant entity host comprises: selecting a first batch of migration paths in the parallel migration sequence; The migration path of the batch moves the virtual host corresponding to the selected batch to the surviving entity host in parallel; and if there is at least one migration path after the selected batch, the parallel is selected Migrating the next batch of migration paths in the sequence and returning to the step of migrating the virtual hosts corresponding to the selected batch. A management entity host for virtual host consolidation, wherein a network connects the management entity host and a plurality of server entity hosts, and multiple virtual hosts run on the server entity host, and the server entity provides N resources. And the virtual host consumes the resource, where N is a preset positive integer, the management entity host includes a memory and a processor coupled to the memory, wherein the processor is the server of the servo entity Classified as a redundant entity host and a surviving entity host, wherein the virtual host running on the redundant entity host can be successfully migrated to the surviving entity host according to an interlaced N-dimensional superior declining padding algorithm; Determining, by the processor, a migration path of the virtual host running on the redundant entity host to the surviving entity host according to the interleaved N-dimensional matching depreciation filling algorithm and a form of the network; The storage path is stored in the memory; the processor determines the virtual host running on the redundant entity host based on the migration path a parallel migration sequence, and storing the parallel migration sequence in the memory; the processor migrating the virtual host running on the redundant entity host in parallel according to the parallel migration sequence to The surviving entity host. The management entity host of claim 16, wherein the processor estimates, according to the one-dimensional matching deprecated padding algorithm, for each i to consume the consumption of the i th resource of all the virtual hosts. The number of servo entity hosts M _i ,1<= i <= N ; when M * K < P , the processor triggers the virtual host to merge, M is the maximum value of all M _i , and K is a preset constant And P is the number of the servo entity hosts. The management entity host according to claim 17, wherein the one-dimensional matching depreciation filling algorithm corresponding to the M _i comprises: assuming that all the servo entity hosts are empty, wherein no virtual host; All the virtual hosts are placed in a queue, and the virtual hosts in the queue are sorted in descending order of consumption of the i-th resource; the first virtual is taken from the queue Hosting the virtual host to the server entity having the minimum remaining capacity of the i-th resource that can accommodate the consumption of the i-th resource of the fetched virtual host; the i-th virtual host of the resource consumption of the physical machine resources of the i-th servo remaining capacity by subtracting the taken out, thereby updating the assigned entity as host of the assigned servo The remaining capacity of the i-th resource; and the step of retrieving the first virtual host from the queue if the queue is not empty, wherein M _i is a depreciation filling algorithm in the one-dimensional matching Used to accommodate all The number of physical machines of the virtual host servo. The management entity host of claim 16, wherein in order to classify the server entity hosts, the processor sorts all of the server entity hosts in order of resource utilization; the processor sorts the after the physical machine prior to the servo classified into the i-th redundant entity behind the physical machine and the host after the ranking of the servo (P - i) th to the classified physical host survival, i is an integer and 1 <= i <= P , P is the number of all the servo entity hosts; the processor finds that the virtual host running on the redundant entity host can be successful according to the interlaced N-dimensional matching depreciation filling algorithm The maximum i that is migrated to the surviving entity host. The management entity host of claim 16, wherein in order to classify the server entity host, the processor sorts all the server entity hosts, and according to the sort order, at least one meets a preset. The server entity host of the conditional virtual host is located after the server entity of the virtual host that does not meet the preset condition, and the server entity of the virtual host that does not meet the preset condition is resource-utilized. sort order; physical machine prior to the servo processor after sorting of the classified into the i-th redundant host and the entity behind the physical machine after the ranking servo (Pi) Category host entity to the survival , i is an integer and 1 <=i<= P , P is the number of all the servo entity hosts; the processor finds that the redundant entity can be found according to the interlaced N-dimensional matching depreciation filling algorithm The virtual host running on the host successfully migrates to the maximum i of the surviving entity host. The management entity host of claim 16, wherein the interlaced N-dimensional superior decrement filling algorithm comprises: calculating, corresponding to each of the resources, based on consumption of the resources of all the virtual hosts a normalization factor; calculating, based on the normalization factor, a standardized consumption of the resources of each of the virtual hosts running on the redundant entity host and each of the virtualities running on the redundant entity host Representative consumption of the host; placing each of the virtual hosts running on the redundant entity host in one or more queues according to the standardized consumption of the resources of each of the virtual hosts; Sorting the virtual hosts in each of the queues in descending order of the representative consumption; checking all of the queues, if all of the queues are empty, returning a successful result; a normalized sum of the remaining capacity of the resource, selecting a queue from the queue; taking the first virtual host from the selected queue; assigning the extracted virtual host to a capacity The surviving entity host having the smallest representative remaining capacity consumed by the all of the resources of the fetched virtual host; if there is no such all of the resources that can accommodate the fetched virtual host The surviving entity host consumes a failure result; and returns to the step of checking all of the queues. If the application of the host management entity patentable scope of item 21, wherein said processor subtracts the sum of all the virtual host resources i from the i-th resource capacity of all of the survival of the host entity and The sum of the consumptions, and then the processor divides the result of the subtraction by the number of the surviving entity hosts, thereby calculating the normalization factor corresponding to the ith resource, where i is an integer and 1< = i <= N ; the processor divides the consumption of the ith resource of each of the virtual hosts by the normalization factor corresponding to the ith resource, thereby calculating the redundancy The standardized consumption of the i-th resource of each of the virtual hosts running on the physical host; the representative consumption of each of the virtual hosts running on the redundant physical host is the virtual host a maximum normalized consumption of said standardized consumption of said resource; said processor dividing said sum of said remaining capacity of said i-th resource of said surviving entity host by said normalization corresponding to said i-th resource Factor, whereby the calculation corresponds to The normalized sum of said first remaining capacity of the resource i. The management entity host of claim 22, wherein there are a total of N queues, and there is a one-to-one mapping relationship between the N queues and the N resources; When the virtual host running on the host host is placed in the queue, the processor places each of the virtual hosts running on the redundant entity host corresponding to the virtual host The representative consumption is in the queue of the resource corresponding to the resource; when the queue is selected from the plurality of queues, the processor defines a set, the set includes the N -1 a queue, the queue corresponding to the resource corresponding to the smallest of the remaining capacity normalized sums; and the processor among the first virtual hosts of each of the arrays, Select the queue in which the first virtual host with the largest representative consumption is located. The management entity host as described in claim 22, wherein there are a total of N *( N -1) queues, and each of the queues corresponds to a different arrangement of two of the N resources; When the virtual host running on the redundant physical host is placed in the queue, for each virtual host running on the redundant physical host, as described in the virtual host When the standardized consumption of the i-th resource is greater than the standardized consumption of the j-th resource of the virtual host, the processor places the virtual host in an arrangement corresponding to the first i-th resource and then the j-th resource In the queue, where j is an integer, 1<= j <= N and j is not equal to i ; when the queue is selected from the plurality of queues, the processor selection corresponds to the first and the largest The resource corresponding to the remaining capacity normalized sum is then a queue of the resources corresponding to the smallest of the remaining capacity normalized sums; when the first virtual host is retrieved from the selected queue The processor removes the removed from all of the queues The virtual host. The management entity host of claim 21, wherein the processor calculates the representative remaining capacity of each of the surviving entity hosts, in order to assign the fetched virtual host, wherein each of the The representative remaining capacity of the surviving entity host is a minimum of the standardized remaining capacity of the resource of the surviving entity host; the processor dividing the remaining capacity of each of the surviving entity hosts by the corresponding The normalization factor of the resource, thereby calculating the standardized remaining capacity of each of the resources of the surviving entity host; the processor assigning the fetched virtual host to accommodate the fetched The surviving entity host having the smallest representative remaining capacity consumed by the all of the resources of the virtual host; the processor subtracting the remaining capacity of the resource from the assigned entity host The consumption of the resources of the virtual host is fetched, thereby updating the remaining capacity of the resource of the assigned entity host. The management entity host of claim 16, wherein each of the migration paths corresponds to one of virtual hosts running on the redundant entity host, and the starting point of the migration path is the virtual host Running the redundant physical host, and the destination of the migration path is the surviving physical host corresponding to the virtual host assigned by the interlaced N-dimensional superior depreciation filling algorithm, the processor according to the The form of the network determines the rest of the migration path. The management entity host of claim 26, wherein in order to determine the parallel migration sequence, the processor defines a conflict map and stores the conflict graph in the memory, wherein the conflict graph Each of the nodes represents one of the migration paths, and each of the nodes has a weight indicating an estimated time required to migrate the virtual host corresponding to the node to the corresponding surviving entity host, if corresponding to The collision graph has any partial overlap of the two migration paths of the two nodes, and there is a connection between the two nodes; the processor uses a maximum node coloring algorithm on the conflict graph. To find a coloring manner of the node, such that the sum of the weights of the colors of the coloring mode is the smallest among all the coloring modes, wherein the weight of each of the colors is the node of the color Describe the maximum value of the weight; the processor adds the migration path corresponding to the node of each of the colors to the parallel migration sequence as a batch migration path. The management entity host of claim 26, wherein in order to determine the parallel migration sequence, the processor defines a directed graph according to the migration path and the form of the network, and The directed graph is stored in the memory; the processor finds a maximum set of one of the disjoint paths in the directed graph, and uses the migration path corresponding to the largest set as a batch migration a path is added to the parallel migration sequence; the processor removes a line corresponding to the largest set from the directed graph; if at least one connection still exists in the directed graph, the processing The step returns to find the maximum set. The management entity host of claim 26, wherein in order to determine the parallel migration sequence, the processor defines a bipartite graph according to the migration path and stores the bipartite graph in the memory. The processor finds a maximum match in the bipartite graph, and adds the migration path corresponding to the maximum match to the parallel migration sequence as a batch migration path; The bipartite graph removes the connection corresponding to the maximum match; if there is still at least one connection in the bipartite graph, the processor returns to the step of finding the maximum match. The management entity host of claim 16, wherein the processor selects the first batch of migration paths in the parallel migration sequence in order to migrate the virtual host running on the redundant entity host; The processor migrating the virtual host corresponding to the selected batch to the surviving entity host in parallel according to the migration path of the selected batch; if there is at least one migration after the selected batch Path, the processor selecting the next batch of migration paths in the parallel migration sequence and returning to the step of migrating the virtual host corresponding to the selected batch.