CN111611073A - A traffic-aware-based container placement method in a containerized data center - Google Patents

Abstract

The invention discloses a container placement method based on traffic awareness in a containerized data center. The invention aims to solve the problem of increased delay when users access services due to inappropriate container placement strategies in the practice of containerized cloud environment, and serious problems Technical issues that reduce the quality of service. This method utilizes the regular characteristics of container traffic in the microservice architecture, and uses the traffic refinement component to place containers that are frequently communicated in the same application into different logical modules according to their traffic correlations, and then place different logical modules belonging to the same application. placed on different physical servers. At the same time, logical modules of different resource-sensitive types are consolidated into the same physical server. Through the above steps, the method reduces the traffic overhead between frequently communicated containers, and also improves the load balancing degree of the data center.

Description

A traffic-aware-based container placement method in a containerized data center

technical field

The invention relates to the technical field of container placement in cloud data centers, in particular to a container placement method based on traffic perception in a containerized data center.

Background technique

Container technology is an operating environment isolation technology similar to the sandbox mechanism. Users can create and run operating systems in containers to achieve operating system-level virtualization. Compared with traditional virtual machines, container technology achieves lightweight application isolation by sharing kernel resources. Due to its various advantages over virtual machines, many cloud service providers have containerized data centers and used microservices architecture to improve the performance of high concurrent requests within the data center.

In a containerized data center, the microservice architecture is often used to decompose an application into several microservices, and each microservice operates according to a certain dependency to improve performance in a concurrent environment. In the microservice architecture, containers belonging to the same application require different levels of network communication to deliver tasks. Real-time applications such as search engines often distribute a user's search request to hundreds or thousands of containers for high concurrent operations.

According to statistics, intra-data center traffic accounts for 70% of the data center's total bandwidth usage, of which cross-rack traffic accounts for about 40%-90%. With the widespread deployment of microservice architectures in containerized data centers, enterprises mainly deploy containers from the perspectives of reducing traffic overhead and load balancing. On the one hand, placing containers with close communication in a nearby location can minimize the traffic overhead between containers, but a single point of failure will make services unavailable; on the other hand, deploying containers to different physical servers can improve load balancing. to a larger optimization effect, but it will increase the communication overhead between containers. Therefore, it is particularly important to balance the reduction of traffic overhead and load balancing in the container placement strategy.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the above-mentioned defects in the prior art, and to provide a container placement method based on traffic awareness in a containerized data center. At the same time, the degree of load balancing is improved.

The purpose of the present invention can be achieved by adopting the following technical solutions:

A container placement method based on traffic awareness in a containerized data center. After receiving a container request, the data center obtains a container placement strategy through an initialization module and a traffic refinement module, and finally the placement module handles the container placement operation. The container placement method Include the following steps:

S1. After the initialization module receives the container request, it creates a traffic matrix in units of a single application according to the traffic exchange between the containers, and further generates a container usage map according to the traffic matrix, which is used to judge the traffic correlation between the containers. and the amount of requests the container makes for computing resources;

S2. The traffic refinement module integrates the containers according to the container usage map, divides the containers belonging to the same application into three levels according to the traffic correlation degree, and stores them in different logic modules. In different logic modules, according to their internal containers Attach corresponding labels to the sensitive types of resources;

S3. Determine the resource-sensitive type of the logical module, mix logical modules of different resource-sensitive types and place them in the virtual machine, and store three logical modules of different levels in the same application separately in different physical servers.

Further, in the step S1, the initialization module receives the container request, and the initialization module collects the information to generate the container usage map as follows:

S11. Parse the container request metadata, and extract the size of the CPU, memory, and network bandwidth requested by the container from the resource request field;

S12. Take a container instance as a node, and its node weight represents the request amount of the container for computing resources; connect the container nodes with communication by edges, and the edge weight between the two nodes represents the network communication volume between the containers.

Further, the described step S2 process is as follows:

For each container instance in the application, the logical modules are divided into three logical modules with gradually increasing correlation according to the traffic correlation between the containers, and the sensitive types of the containers in the logical modules to the requested computing resources are marked.

Further, the traffic refinement module presents a certain Zipf distribution through the container traffic under the micro-service architecture, and divides the divided logical modules into three levels according to the degree of traffic correlation between the containers, which are respectively unrelated and weakly related. and strong associations; among them, strong associations refer to the containers that account for the top 30% of network traffic, weak associations refer to containers that account for the top 30% to 50% of network traffic, and the rest of the containers are defined as unrelated; in addition , the logical modules are further marked as CPU-sensitive and memory-sensitive according to the resource sensitivity of the container in the logical module.

Further, the described step S3 process is as follows:

S31. When the association degree of the placed logic module is marked as weak association, it is temporarily placed in the waiting queue, and the logic modules with no association and strong association are preferentially placed;

S32. In the process of placing the logical modules, mix and place two different sensitive types of logical modules according to their resource sensitive tags, and adopt a resource reservation policy for each virtual machine to reserve certain resources for the weakly associated logical modules ;

S33. When placing a weakly associated logical module, select an optimal virtual machine for the current logical module, in order to improve the load balancing degree of the physical server.

Further, the container usage graph is a graph-like data structure independently established according to each application in the container request, and is used to save the metadata of the container under the application, so as to facilitate the calculation of the degree of traffic correlation between the containers. and the requested resource.

Further, maintain a container tuple for temporary comparison in the logic module, extract the container that communicates with the container in the current temporary comparison container tuple from the list of containers not added to the logic module, and further update the temporary comparison container element. Initially, calculate the total traffic of the containers under the application, extract the two containers whose inbound and outbound traffic accounts for the largest proportion of the total traffic, add them to the logic module, and set the container group as a temporary comparison container tuple.

Further, in the step S32, the mixed placement process of two different sensitive types of logic modules according to their resource sensitive tags is as follows:

For the placement of each logical module, it is regarded as the knapsack problem in the classical algorithm, and the optimization goal is to find a container placement method, so that the CPU and memory utilization of physical servers in the data center can reach more than 65%.

Compared with the prior art, the present invention has the following advantages and effects:

(1) The present invention adopts the means of greedy calculation to accurately calculate the flow correlation degree between the containers, so as to avoid the influence on the placement decision time caused by the high complexity of the algorithm in the real-time calculation process.

(2) The present invention divides the container request into several logical modules with different traffic correlation degrees according to the degree of association between the containers, and ensures that the containers between the logical modules have no network communication as much as possible in the algorithm.

(3) The present invention considers improving the load balancing degree of the data center while reducing the exchange cost of network traffic.

Description of drawings

FIG. 1 is an application architecture diagram of a container placement method based on traffic awareness in a containerized data center disclosed in the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Example

As shown in Figure 1, it is an application architecture diagram of a container placement method based on traffic awareness in a containerized data center. It is applied in a containerized data center under the microservice architecture. To solve the problem of high latency caused by large size, the container is integrated by using the traffic rules presented by the container exchange under the microservice architecture to reduce the traffic overhead. At the same time, it optimizes the single point of failure problem caused by the low degree of load balancing of the data center in the process of reducing traffic overhead. The present invention conducts a comparative test on this greedy algorithm, and evaluates the performance impact of the present invention on the services provided by the data center from multiple angles.

In order to clarify the application scenario of the present invention more clearly, the following detailed analysis is made in conjunction with the application architecture diagram, which is mainly divided into three steps:

1) Build a container usage graph. According to the Zipf distribution of container traffic under the microservice architecture, a corresponding container usage graph is created for each application, which is used to represent the request for computing resources of all containers under the application and the network communication traffic between containers. For i application, its container usage graph can be expressed as G _i (V, E), V is the container node in the graph, and its node weight can be expressed as C _A =<CPU,MEM,Disk,Bandwidth>, indicating that the container The number of resources requested; E represents the edge connecting the container nodes in the graph, and its edge weight represents the network communication traffic between the two containers, for example, E<C _A ,C _B >=6, E<C _A ,C _D > =56. In order to facilitate the description of the network communication flow characteristics between containers, it can be expressed in the form of (C _i , C _j , D _ij ) or (C _j , C _i , D _ij ), that is, the network communication flow of container C _i and container C _j is D _ij . When the containers are placed, the physical link distance between the containers will be added to the edge weight, which is convenient for calculating the traffic overhead generated by the data center after the placement is completed, and is used to compare the impact of different placement strategies on the traffic overhead.

对于C₃来说，其占比最大的对外通讯流量发生在与C₄的网络交流中，其值等于

此时不满足“占比最大”条件。其余容器添加进DWS的过程为：对于当前最新的LCPDWS，按照一对一的原则在其余容器中找到与处于LCPDWS中的两个容器网络通讯流量最大的两个容器，将这两个容器添加进DWS中。至此，精炼算法完毕。对于分割算法，在精炼完毕后的每一个DWS中，应用分割算法将其分割为三个不同关联等级的逻辑模块。将前x％频繁交流的容器放置于强关联逻辑模块中，这些容器为上文所述的“交流较大”容器，将接下来的y％容器放置于弱关联逻辑模块中，这些容器为上文所述的“交流较少”容器，将剩余容器放置无关联逻辑模块中，这些容器为上文所属的“无交流”容器。具体x,y的值取决于其依赖的DWS，其中DWS_i表示在DWS中的第i个容器所对应与其他容器的网络通讯总量。根据Zipf分布，设定x的取值在(0,0.3)中，即在前30％容器中找到一个下标i满足DWS_i-DWS_i+1最大。对于y的取值，则在(x,0.5)范围中取得。对于无关联逻辑模块，则在(y,1)范围中取得。至此，分割算法完毕。在完成分割算法的同时，根据逻辑模块中容器对资源请求的敏感程度，将逻辑模块进一步标签化为CPU敏感型和内存敏感型。2) Traffic refinement. The refinement of the communication traffic of the container network in the present invention is mainly accomplished by two greedy algorithms. As shown in the traffic refinement module in FIG. 1 , the two algorithms can be respectively called a refinement algorithm and a segmentation algorithm. After the traffic refinement module receives the container usage map, it first reorders the container sequence according to the traffic correlation degree of the container to generate a DWS (Descending Weight Sequence), and then divides the sorted container sequence into three different traffic correlation degrees. logic module. For the refining algorithm, such as the current four containers C ₁ , C ₂ , C ₃ , C ₄ , there are (1, 2, 1), (1, 3, 4), (1, 4, 6), (2, 3, 7), (2, 4, 8), (3, 4, 5), after adding the edge weights connecting each node, for C ₁ , C ₂ , C ₃ , C ₄ get 11, 16 respectively , 16, 19 (eg, 11=E<C ₁ ,C ₂ >+E<C ₁ ,C ₃ >+E<C ₁ ,C ₄ >). First, for the largest two values of 19 and 16, find the two containers C ₂ and C ₄ with the largest proportion of external communication traffic, add them to DWS, and set them as the initial LCPDWS (Last container pair in DWS). For C2 _{, C4} , respectively

For C ₃ , the largest proportion of external communication traffic occurs in the network communication with C ₄ , and its value is equal to

At this time, the "largest proportion" condition is not met. The process of adding other containers to DWS is as follows: For the current latest LCPDWS, find the two containers with the largest network communication traffic with the two containers in the LCPDWS according to the one-to-one principle, and add these two containers to the DWS. in DWS. At this point, the refining algorithm is complete. For the segmentation algorithm, in each DWS after refining, the segmentation algorithm is applied to segment it into three logical modules with different association levels. Place the first x% of frequently communicated containers in the strong association logic module, these containers are the "big communication" containers described above, and place the next y% containers in the weak association logic module, these containers are the upper For the "less communication" container described in the article, the remaining containers are placed in unrelated logic modules, and these containers are the "no communication" containers to which the above belongs. The specific value of x, y depends on the DWS it depends on, where DWS _i represents the total amount of network communication with other containers corresponding to the ith container in the DWS. According to the Zipf distribution, the value of x is set in (0, 0.3), that is, a subscript i is found in the top 30% of the containers to satisfy the maximum of DWS _i -DWS _i+1 . For the value of y, it is obtained in the range of (x, 0.5). For unrelated logic modules, it is obtained in the (y,1) range. At this point, the segmentation algorithm is completed. While completing the segmentation algorithm, according to the sensitivity of the container in the logic module to resource requests, the logic module is further labeled as CPU-sensitive and memory-sensitive.

3) Place the logic module. For the placement of logical modules, we mainly refer to the solution in the multi-knapsack problem. Logical modules with different degrees of association under the same application are placed in virtual machines in different physical servers, and logical modules of different resource-sensitive types are mixed at the same time. For example, CPU-sensitive logic modules and memory-sensitive logic modules are mixed and placed under the condition that the above conditions are satisfied. The placement process of logical modules is divided into two rounds. In the first round, logical modules with strong association and no association are placed preferentially. For placement of weakly associated logic modules. In the second round, weakly related logic modules are placed.

To sum up, this embodiment proposes a traffic-aware-based container placement strategy in a containerized data center. This method utilizes traffic exchange between containers to present a special Zipf distribution. Containers with the same degree of association are placed together, and then different The logical modules of the related degree are placed separately on different physical servers, and finally mixed according to the resource-sensitive type of the logical modules. While reducing the data center traffic overhead, the load balancing degree is improved.

The above-mentioned embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above-mentioned embodiments, and any other changes, modifications, substitutions, combinations, The simplification should be equivalent replacement manners, which are all included in the protection scope of the present invention.

Claims (8)

1. A container placement method based on flow perception in a containerized data center is characterized in that the container placement method comprises the following steps of:

s1, after the container request is received by the initialization module, a flow matrix is created by taking a single application as a unit according to the flow communication condition among the containers, and a container quantity map is further generated according to the flow matrix and is used for judging the flow association degree among the containers and the request quantity of the containers for computing resources;

s2, the flow rate refining module integrates containers according to the container quantity diagram, divides the containers belonging to the same application into three grades according to the flow rate association degree, stores the three grades into different logic modules, and attaches corresponding labels to the different logic modules according to the sensitive types of the containers in the different logic modules to the resources;

s3, judging the resource sensitivity types of the logic modules, mixing the logic modules with different resource sensitivity types, placing the logic modules into the virtual machine, and separately storing the three logic modules with different levels in the same application into different physical servers.

2. The method for placing the container based on the flow sensing in the containerized data center according to claim 1, wherein in step S1, the initialization module receives the container request, and the process of the initialization module collecting the information to generate the container volume map is as follows:

s11, analyzing the metadata of the container request, and extracting the sizes of the CPU, the memory and the network bandwidth of the container request from the resource request field;

s12, taking a container instance as a node, wherein the node weight of the container represents the request amount of the container to the computing resource; the container nodes with communication are connected by edges, and the edge weight between the two nodes represents the network communication quantity between the containers.

3. The method for placing the container based on the flow sensing in the containerized data center according to claim 1, wherein the step S2 comprises the following steps:

for each container instance in each application, dividing the logic module into three logic modules with gradually-enhanced association degrees according to the traffic association degrees among the containers, and marking the sensitive types of the containers in the logic modules to the requested computing resources.

4. The method for placing the container based on the flow perception in the containerized data center according to claim 1, wherein the flow refining module presents certain Zipf distribution through the container flow under the micro-service architecture, and the divided logic module level is divided into three levels according to the flow association degree among the containers, wherein the three levels are respectively non-association, weak association and strong association; wherein, the strong correlation refers to a container with the network communication amount accounting for 30% of the first time, the weak correlation refers to a container with the network communication amount accounting for 30% to 50% of the first time, and the rest containers are defined as no correlation; in addition, the logic modules are further marked as CPU sensitive type and memory sensitive type according to the sensitivity of the containers in the logic modules to the resources.

5. The method for placing the container based on the flow sensing in the containerized data center according to claim 1, wherein the step S3 comprises the following steps:

s31, when the relevance degree of the placed logic module is marked as weak relevance, temporarily placing the logic module into a waiting queue, and preferentially placing the logic module without relevance and with strong relevance;

s32, in the process of placing the logic modules, the two logic modules with different sensitive types are placed in a mixed mode according to the resource sensitive marks of the logic modules, and a resource reservation strategy is adopted for each virtual machine to reserve certain resources for the logic modules with weak association;

and S33, when the weakly associated logic module is placed, selecting the optimal virtual machine for the current logic module, so as to improve the load balancing degree of the physical server.

6. The method as claimed in claim 1, wherein the container volume map is a map-like data structure separately established according to each application in the container request, and is used to store metadata of the container under the application, so as to calculate the traffic association degree between the containers and the requested resources.

7. The container placement method based on flow perception in the containerized data center according to claim 1, wherein a container tuple for temporary comparison is maintained in the logic module, the container with the largest communication with the container in the container tuple currently and temporarily compared is extracted from a container list not added to the logic module, the temporarily compared container tuple is further updated, in an initial situation, the total flow of the containers in the application is calculated, two containers with the largest total flow of the incoming degree and the outgoing degree in comparison to the total flow are extracted and added to the logic module, and the container group is set as the temporarily compared container tuple.

8. The method for placing the container based on the traffic awareness in the containerized data center according to claim 5, wherein in step S32, the process of placing the logic modules of two different sensitive types in a mixed manner according to the resource sensitivity labels thereof is as follows:

regarding the placement of each logic module as a knapsack problem in a classical algorithm, the optimization aims at finding a container placement mode so that the utilization rate of a CPU and a memory of a physical server in a data center reaches more than 65%.

Images