TW201622381A - Date center network provisioning method and system thereof - Google Patents

Abstract

Data center network provisioning method includes partitioning a plurality of virtual machines in the data center network into a plurality of groups according to topology information of the data center network. A plurality of virtual machines in each group is assigned to a plurality of servers in the date center network. A pair-wised transmission path of the plurality of servers is assigned according to a traffic load of each switch in the data center network.

Description

Data center network configuration method and system thereof

The invention discloses a data center network configuration method and a system thereof, in particular, a network configuration method and a system thereof under a Software Defined Network.

In recent years, with the flourishing development of cloud computing, virtualization technology has become a hot research topic. Host virtualization can convert a single entity host into multiple virtual machines (VMs) that work together. The host performs parallelization to provide reliable service quality. But virtualization technology is applied to cloud-level networks, requiring huge amounts of computing power, memory, and data storage. To this end, Stanford University has developed a software defined network (SDN) system, and developed the OpenFlow architecture. The original goal was to extend the programmable programming features of the campus network's switching lines and provide a corresponding virtual platform. In general, a software-defined network consists of a centralized controller (Controller) and tens of thousands of switches (Switch) that are connected to each other and provide a transmission path to all physical machines. The connection relationship is a topology structure, which also constitutes a data center (Date Center) system under the software definition network.

The current cloud network uses software to define the data center under the network, and will face several key issues. First, the physical machine under the data center establishes a connection with another physical machine through several switches. However, if the link distance of the link is not optimized, the core switch of the uppermost layer of the data center ( Core Switch) will be vulnerable to traffic congestion and increase network latency. Second, if the number of switches used on a physical machine is not optimized, it can cause power loss problems. Third, in two entities When the data is transmitted to each other on the machine, the connection is established through several switches. However, when a switch generates a traffic congestion (Traffic Congestion), the transmission speed through the switch is greatly reduced. At this time, if there is no optimized reroute mechanism, the throughput of the data transmission (Throughput) will also decrease, resulting in deterioration of the connection quality.

Therefore, it is very important to develop a data center network configuration method and its system to find a balanced optimal solution among the above three transmission problems.

An embodiment of the present invention provides a data center network configuration method, which includes dividing a plurality of virtual machines in a data center network into a plurality of groups according to topology information of a data center network, and each group A plurality of virtual machines are allocated to a plurality of servers in the data center network, and the transmission paths of the servers are allocated according to the transmission traffic of each switch in the data center network.

Another embodiment of the present invention provides a data center network configuration system, including a plurality of switches, a plurality of servers, a plurality of virtual machines, and a control machine. Each of the switches is connected to each other according to topology information. The servers are connected to the switches, and the virtual machines are connected to the servers, and the controller is connected to the switches, and the servos are connected to the switches. And the virtual machines, wherein the control machine divides the plurality of virtual machines in the data center network into a plurality of groups according to the topology information of the data center network, and the control machine divides the plurality of virtual machines in each group Allocating to a plurality of servers in the data center network, and allocating the transmission paths of the servers according to the transmission traffic of each switch in the data center network.

100‧‧‧Data Center

1 to 12‧‧‧ switches

13‧‧‧ Servo

14, 16, 17‧‧ virtual machines

15‧‧‧Control machine

S201 to S203‧‧‧ steps

G1‧‧‧First Group

G2‧‧‧ second group

1 is an architectural diagram of an embodiment of a data center of a software-defined network of the present invention.

Figure 2 is a flow chart of the network configuration method of the embodiment of Figure 1.

Fig. 3 is a schematic view showing a method of reducing the winding distance in the embodiment of Fig. 1.

Fig. 4 is a schematic view showing a method of rewinding in the embodiment of Fig. 1.

1 is an architectural diagram of an embodiment of a data center of a software-defined network of the present invention. As shown in FIG. 1, the data center 100 includes a plurality of core layer switches 10, a plurality of aggregation layer switches 11, a plurality of edge layer switches 12, a plurality of server machines 13, a plurality of virtual machines 14, and a control unit 15. In this embodiment, the data center 100 is built with a wide tree-like information (Fat-Tree Information), has a three-layer switch structure, and a plurality of core layer switches 10 are top-level switches of the topological structure, and a plurality of switches. The aggregation layer switch 11 is a switch of the intermediate layer of the topology structure, and the plurality of edge layer switches 12 are the switches of the lowest layer of the topology structure. The three-layer switches are connected to each other in the manner shown in FIG. The plurality of servos 13 are physical machines in this embodiment, and the paired servos 13 can transmit data after establishing a transmission path through the switch. The virtual machine 14 is a non-physical machine used by each user, and these virtual machines 14 are attached under the servo 13, using the hardware resources and bandwidth of the server 13. For example, in FIG. 2, three virtual machines 14 are executed under the leftmost server 13, and the three virtual machines 14 can transmit data to other virtual machines through the leftmost server 13. The control unit 15 is coupled to all of the switches 10 to 12, the servos 13, and the virtual machines 14 for controlling the distribution and transmission of traffic throughout the data center 100. However, the conventional data center does not optimize the Hop Distance of the paired servo 13 link paths, and does not optimize the number of virtual machines 14 configured on each of the servos 13, nor does it Consider an optimized reroute mechanism that avoids a significant reduction in transmission speed when a switch 10, 11 or 12 generates a traffic congestion (Traffic Congestion). Therefore, the reliability of data transmission in the entire traditional data center is poor and energy consumption is high. In order to simultaneously improve the transmission bandwidth, reliability, and power consumption, the present invention proposes a method for configuring the data center 100 network, and the steps thereof will be described in detail below.

2 is a flow chart of a method for configuring a data center 100 network of the present invention. In FIG. 2, the data center 100 network configuration method includes three steps S201 to S203, as shown below: S201: according to the topology information of the data center network 100, the data center network The plurality of virtual machines 14 are divided into a plurality of groups; S202: assigning a plurality of virtual machines 14 of each group to a plurality of server machines 13 in the data center network; and S203: according to the data center 100 The transmission traffic of each of the switches 10 to 12 distributes the transmission paths of the servos 13 to each other.

In the flowchart of the configuration method of the data center 100 network, step S201 is a step of performing re-grouping of a plurality of virtual machines 14 in the data center network, the purpose of which is to reduce the Hop Distance Reduction. Step S202 is to reallocate the hardware resources of the server 13 to the plurality of virtual machines 14, the purpose of which is to save energy saving. Step S203 is a plan for performing re-routing of the transmission paths of the servos 13 to each other in order to balance the load distribution between the servos 13. Therefore, the configuration method of the data center 100 network of the present invention can be regarded as an algorithm that simultaneously considers the optimization of the path distance, energy consumption, and traffic distribution balance. The process and principle of steps S201 to S203 will be described in detail below.

Please refer to FIG. 3, which is a schematic diagram of the data center 100 of FIG. 2 performing step 201 to reduce the winding distance. In this step, the virtual machine 14 of the data center 100 is divided into two groups (the present invention is not limited thereto, and other embodiments may be divided into N groups, and N is a positive integer greater than 2). The first group G1 includes a plurality of virtual machines 16, and the second group G2 includes a plurality of virtual machines 17. The data transmission correlation in each group is very high, that is, the virtual machines in the same group will transmit data to another virtual machine in the same group at a higher probability. The data transmission correlation is low in two different groups, that is, the probability that the virtual machine 16 in the first group G1 transmits data to the virtual machine 17 in the second group G2 is lower. How these two groups G1 and G2 are distinguished by the original data center 100 in Fig. 1 will be described below. The data center 100 of Fig. 3 has a plurality of servos 13, assuming that the index values (Index) of these servos 13 are 1 to M, and M is a positive integer. At this time, since the paired servos 13 establish a connection path through a plurality of switches in the data center 100, the data can be transmitted. Therefore, the connection path corresponding to the paired servos 13 corresponds to the value of the number of switches used. This value is called the Hop Number. Thus, the i-th servo 13 via the plurality of switches of the j-th servo 13 for transmission, the number of jumpers 15 by the control unit that, here assumed to be h _ij. Therefore, in the data center 100 of FIG. 1, the number of jumpers transmitted by all the servos 13 can be expressed as a matrix of the number of relay points, as follows:

Where H is the number of relay point matrices, and the number of relay point matrices H is a symmetric matrix ( H=H ^T ) and each of the elements is a real number. Furthermore, the diagonal value of the relay point number matrix H is 0, which is because the server does not need to go through the switch when it transmits the data to itself. When the number of the relay build a dot matrix H is completed, the relay will be based on the number of dots matrix H, corresponding to the establishment of load transfer matrix, as defined below. When the i-th server 13 transmits the j-th server 13 via a plurality of switches, its traffic load can be known by the controller 15, which is assumed to be l _ij . Therefore, in the data center 100 of FIG. 2, the transmission load amount of all the servos 13 can be expressed as a transmission load amount matrix as follows:

Where L is the transmission load matrix, and the transmission load matrix L is a symmetric matrix ( L=L ^T ) and each of the elements is a real number. Furthermore, the Diagonal value of the transmission load matrix L is partially zero. This is because the servo does not need to transmit through the switch if it transmits the data to itself. According to the relay point quantity matrix H and the transmission load amount matrix L , the server and the corresponding used virtual machine can be grouped via some transmission indicators. For example, calculate the sum of the values on each column (Row) of the relay point number matrix H , . The numerical meaning is the sum of the number of switches used for all paths outside the i-th server connection; calculate the sum of the values on each column of the transmission load matrix L (Row), . The numerical meaning is the sum of the transmission loads of all paths outside the i-th server. Therefore, according to the characteristics of the relay point number matrix H and the transmission load amount matrix L , the switch can be grouped as shown in FIG. In Figure 3, because virtual machines in the same group will transfer data to another virtual machine in the same group at a higher probability. Data transfer correlation is low in two different groups. Therefore, in the data center 100 of FIG. 3, the usage rate of the core layer switch 10 through which the cross-group transmission must pass is greatly reduced, so that the probability of network congestion can be reduced. Moreover, since the virtual machine transmits data in the same group at a high probability, the virtual machine 16 in the first group G1 or the virtual machine 16 in the second group G2 can have a shorter path. Transfer data. Therefore, step S201 can achieve the effect of reducing the Hop Distance Reduction.

After performing step S201, the virtual machine and the server are divided into two groups G1 and G2, and the controller 15 also obtains the number of virtual machines in each group. Next, in order to enable the servos of all entities in the data center 100 to operate in a minimum energy consumption mode, the controller 15 performs step S202 to allocate a plurality of virtual machines of each group to the data center 100 network. There are a plurality of servos in the machine to minimize the number of servos with large loads, and the method is described in detail below. First, the controller 15 in the data center 100 obtains system resource requirements for all virtual machines in each group. For example, the controller 15 will obtain the system resource requirements of all the virtual machines 16 in the first group G1 and the system resource requirements of all the virtual machines 17 in the second group G2. The system resource requirements described herein may be processor load requirements, memory capacity requirements, and/or bandwidth usage requirements. However, the present invention is not limited thereto, and system resource requirements may be other resource indicators. The control unit 15 also obtains the upper limit of the system resources supported by each of the servos 13. The system resource upper limit described herein may be a processor load upper limit, a memory capacity upper limit, and/or a bandwidth usage upper limit. However, the present invention is not limited thereto, and the system resource upper limit may be another resource upper limit. Next, the control unit 15 adds the system resource requirements of each virtual machine separately, and sorts them from large to small. For example, the first group G1 i-th virtual machine 16, which includes a processor resource requirements CR _i load demand, demand for memory capacity and bandwidth usage, the MR _i demand BR _i. Therefore, the total resource requirement of the i-th virtual machine 16 is TR _i = CR _i + MR _i + BR _i . Assuming that there are MG virtual machines in the first group G1, the control machine 15 sorts the total resource requirements {TR ₁ , ..., TR _MG } of the virtual machine from large to small, and the total resource demand is high. The virtual machine starts to be distributed to the server 13, and the allocation method uses the flow of the Knapsack Problem Algorithm to allocate each group of virtual machines to the server. For example, the first open servo 13 has a processor load upper limit of CU ₁ , a memory capacity upper limit of MU _{1 ,} and a bandwidth usage upper limit of BU ₁ . Therefore, when the virtual machines are assigned to the server one by one, it is necessary to monitor whether the servo 13 has exceeded the load amount at any time. For example, the first open server 13 has already allocated two virtual machines, thus complying with CU ₁ >CR ₁ +CR ₂ , MU ₁ > MR ₁ + MR ₂ , BU ₁ > BR ₁ + BR ₂ relationship, each resource load has not been overloaded. If the third virtual machine is redistributed to the servo 13, and CU ₁ <CR ₁ +CR ₂ +CR ₃ occurs, MU ₁ <MR ₁ +MR ₂ +MR ₃ , BU ₁ <BR ₁ +BR ₂ + In the case of more than one of BR ₃ , it indicates that one or more of the resource loads are overloaded. At this time, the controller 15 turns on a server 13 to allocate the third virtual machine. In the above calculations for system resource requirements and system resource caps, these parameters are quantized and normalized in order to calculate metrics that are consistent. In step 202, the controller 15 tries to maximize the number of virtual machines to which a single server 13 is allocated. In other words, when the system load is not full, the number of idle states or unused servers is also At most, equivalent to step 202, the minimum number of servers 13 can be used to meet the power requirements of all virtual machine resources of the data center 100, thereby saving energy savings.

After performing steps S201 and S202, the topology of the material center 100 is divided into two groups, and the controller 15 opens a minimum number of servers 13 for each virtual group to use for the corresponding virtual machine. However, when the number of the server 13 and the virtual machine of the entity is determined in step S202 After that, each server 13 is planned to be completed through a path connecting several switches to another server 13. At this time, in step S203, the controller 15 considers the problem of re-routing. First, the control unit 15 detects or monitors the transmission traffic of each switch in the data center 100 network, and calculates the sum of the transmission traffic of the switches through which each pair of servos 13 communicate with each other. If the sum of the transmission traffic of the switch through which the transmission path passes is greater than the threshold (Threshold), the control unit 15 searches for a non-conflicting path (Disjointed Path) and reroutes the transmission path, wherein the new path The sum of the transmission traffic of the switch through which it passes is not greater than the threshold.

Figure 4 is a schematic illustration of the rewinding step of the embodiment of the present invention. To simplify the description, Figure 4 uses only 10 switches to describe the routing process. In Figure 4, the 10 switches are switch 0 to switch 9, respectively, and each switch has its transmission traffic. In Fig. 4, the transmission path of the switch 6 to the switch 8 is considered. First, the predetermined path planned by the control unit 15 is the path P1, which is expressed as follows: P1: switch 6 → switch 2 → switch 0 → switch 4 → switch 8 However, the control machine 15 adds up all the transmission traffic in the path P1 and finds It exceeds the threshold value, so it is judged that the path P1 is an over-utilized path, and there may be a high probability of transmission congestion. Therefore, the control unit 15 performs the operation of re-routing the P1 path, and corrects the transmission path of the switch 6 to the switch 8 as the P2 path, which is expressed as follows: P2: switch 6 → switch 3 → switch 1 → switch 5 → switch 8 After re-bypassing, the probability of transmission congestion occurring when the switch 6 to the switch 8 perform data transmission can be effectively reduced. Therefore, in step 203, since the control unit 15 detours the path where the transmission congestion may occur, a further improvement in the quality of the overall data transmission can be obtained.

In summary, the present invention describes a method for configuring a data center network, which uses a relay point quantity matrix and a transmission load matrix to group data center topology structures to reduce the number of switches used between the two servers. , the effect of reducing the distance of the transmission path. And use each The system resource requirements of a virtual machine between groups and the system resource limit of the server allocate virtual machines to appropriate servers to enable virtual machines with a minimum number of servers to achieve energy saving. Furthermore, the control unit will check whether the path used by the pair of servos to transmit data may cause a transmission congestion, and if necessary, a path bypass of the transmission congestion may occur, so that the quality of the overall data transmission can be further improved. Therefore, in the network configuration method of the present invention, the data center can simultaneously achieve low transmission distance, high energy saving, and high transmission quality.

The above are only the preferred embodiments of the present invention, and all changes and modifications made to the scope of the present invention should be within the scope of the present invention.

100‧‧‧Data Center

10 to 12‧‧‧ switches

13‧‧‧ Servo

14‧‧‧Virtual Machine

15‧‧‧Control machine

Claims (10)

A data center network configuration method includes: dividing, according to one topology information of the data center network, a plurality of virtual machines in the data center network into a plurality of groups; each of the plurality of groups The virtual machine is allocated to a plurality of servers in the data center network; and the transmission paths of the servers are allocated according to the transmission traffic of each switch in the data center network. The method of claim 1, wherein the topology information of the data center network is a wide tree-like information (Fat-Tree Information), and the switches in the data center network include a plurality of switches Core Level Switch, a number of Aggregate Level Switches, and a number of Edge Level Switches. The method of claim 1, wherein the virtual machines in the data center network are divided into the groups according to the topology information of the data center network, including: according to the data center network Establishing a relay point quantity matrix according to the topology information; establishing a transmission load quantity matrix according to the topology information of the data center network; and according to the relay point quantity matrix and the transmission load quantity matrix, the data The virtual machines in the central network are divided into the groups. The method of claim 3, wherein the relay point quantity matrix and the transmission load quantity matrix are Symmetric Matrix, and all elements in the two matrix are Real Number. The method of claim 3, wherein the virtual machines in the data center network are divided into the groups according to the relay point quantity matrix and the transmission load quantity matrix, according to the relay point Quantity moment And the transmission load matrix, the virtual machines in the data center network are divided into a plurality of groups corresponding to different transmission traffic intervals. The method of claim 1, wherein the assigning the virtual machines of each group to the servers in the data center network comprises: obtaining support by each of the servers a system resource cap; obtaining a system resource requirement of each virtual machine in each of the virtual machines of each group; and determining a system resource limit supported by each server and a system resource requirement of each virtual machine Allocating the virtual machines of each group to the servers through a Knapsack Problem Algorithm; wherein the system resource upper limit includes a processor load upper limit, a memory capacity upper limit, and/or Or a bandwidth usage limit, and the system resource requirement includes a processor load requirement, a memory capacity requirement, and/or a bandwidth usage requirement. The method of claim 1, wherein the transmission paths of the servers are allocated according to the transmission traffic of each switch in the data center network, including: detecting each switch in the data center network. Transmitting traffic; calculating the sum of the transmission traffic of the switches through which each pair of servos pass each other; and rerouting the transmission path if the sum of the transmission traffic of the switches through which the transmission path passes is greater than a threshold. The method of claim 7, wherein if the sum of the transmission traffic of the switch through which the transmission path passes is greater than the threshold, the transmission path is re-routed if the sum of the transmission traffic of the switch through which the transmission path passes is greater than the Threshold value, re-bypassing the transmission path to a path where the sum of the transmitted traffic of the passing switch is not greater than the threshold value. A data center network configuration system includes: a plurality of switches, each switch is connected to each other according to a topology information; a plurality of servers are connected to the switches; and a plurality of virtual machines are connected to the servers; and a control machine, connected to the switches, the servers, and the virtual machines; wherein the controller divides the plurality of virtual machines in the data center network according to one of the data center network information a plurality of groups, the controller allocates a plurality of virtual machines of each group to a plurality of servers in the data center network, and allocates the network according to the transmission traffic of each switch in the data center network. The transmission paths of these servers to each other. The configuration system of claim 8, wherein the plurality of switches comprises: a plurality of core layer switches; a plurality of aggregation layer switches connected to the core layer switches; and a plurality of edge layer switches connected to the aggregation layers The switch; wherein the core layer switches are the uppermost switches in the topology of the data center network, and the edge layer switches are the lowest switches in the topology of the data center network.