Cloud video conferencing and cloud video command play an important role in various fields. The balanced utilization of the underlying resources of cloud video conferencing systems is an important research direction for cloud video. At present, most of the cloud video server cluster resources are native scheduling algorithms, which are simple and weak in scene specificity. The native algorithm can meet the basic needs,but cannot well refine and balance scheduling resources. The previous research mostly carried out scheduling by formulating rescheduling strategies. The configuration is complex, and the amount of engineering is large. In this paper, the Refined Balanced Resource Allocation algorithm is proposed by introducing the coefficient of variation. Compared with the default algorithm, it is easy to implement in engineering, and can better achieve balanced scheduling of resources, avoid resource fragmentation, and distribute pods on the cluster more dispersed. After simulation experiments, the algorithm proposed in this paper improves Spread Score, and the balance of large cluster resources has increased by 14.28%. The algorithm proposed in this paper is quite effective and feasible.
1. Introduction 1
Cloud video services play an important role in communication in our contemporary life. Since the
COVID-19 pandemic, people's offline travel and communication have been greatly restricted. Due to a
large number of users, the resource consumption of the server cluster is relatively large. In certain
scenarios, it needs to face the problems of a sudden increase in access services and excessive server
pressure. The emergence of cloud computing and container[
Since 2013, with the rise of container technology [
Based on the above research, the survey found that the current cloud video has three challenges. First, the default scheduling strategy is too simple to meet the scheduling requirements of specific scenarios. Second, the default algorithm avoids fragmentation of resources on nodes during balanced scheduling, but high CPU and memory loads on specific nodes will occur, resulting in unbalanced use of cluster resources. Third, the unbalanced distribution of pods increases the node failure rate.
the coefficient of variation. While the resources in the nodes are used in a balanced manner, the pods on the cluster is more dispersed.
with the default algorithm, the RBRA algorithm is easier to implement in engineering, and can better achieve balanced scheduling of resources, avoid resource fragmentation.
A Kubernetes cluster usually consists of a master and multiple nodes[
Predicates: It filters out nodes that do not meet the conditions from all nodes in the current cluster according to the scheduling policy.
Priorities: Score these nodes; finally select the node with the highest priority and bind it to the Pod.
The default balanced resource allocation algorithm, which belongs to priorities, selects nodes with the most balanced resource usage. Usually, when the number of resources is two, the algorithm first calculates the CPU and memory usage on the node. The difference is calculated between CPU and memory usage. 10 minus 10 times the difference between CPU and memory usage. The final score is obtained.
= =
ScoreDefualtBalance = 10 − 10| the final Score − = =
, = + + )2 + ( − )2 + ( −
1 )2]2/3, ScoreDefualtBalance = (1 −
) ∗ 100
Scholars have also made improvements to the default algorithm. The existing research mostly adopts
the combined scheduling strategy to schedule resources. Du Jun[
Ghofrane [
(a) (b) Figure 1: Result of Default algorithm and one of the possible desired results. When a list of pod requirements is submitted, according to the default algorithm, pods will be preferentially scheduled to the nodes with the largest remaining node resources, as shown in (a); (b) represents the scheduling result after introducing the coefficient of variation to eliminate the influence of the order of magnitude of nodes in the cluster.
Based on the possible business requirements on the server, a Refined Balanced Resource Allocation
(RBRA)algorithm is proposed. After deconstructing the resource scheduling algorithm[
To eliminate the influence of the level of data values and the different measurement units on the
measurement value of dispersion degree, it is necessary to calculate the coefficient of variation[
the degree of dispersion of the set of data.
For resource utilization data, set as 1, 2, … , , n represents the number of resources, represents the ratio of the nth resource, and ̅ represents the average number of 1, 2, … , . Then we calculate
2 = ∑ =1( − ̅)2
∑ = √ =1 ( − ̅)2 ,
, ̅ =
∑ =1 , =
√∑ =1 ( − ̅)2 ∑ =1 = ̅
1) Read Pods nodes information 2) The number of resources ≤2, =
CPUtotal CPUuse , Score is the same as the default algorithm. Score= 10 − 10| respectively. respectively.
N=3, the CPU usage request, memory usage request, and storage usage request are counted as,
N resources of each node, and obtain the available resources of the node, for example, N=3,
= [( cpuFraction - mean )2 + ( memoryFraction − mean )2 + ( storageFraction − mean )2]1/2/3
=
Select the node with the highest score, when there are multiple nodes with the same highest score, randomly select the node as the deployment node.
This section introduces the experimental results of the proposed RBRA algorithm based on simulation. First, the basic configuration of the experimental environment is introduced, and then the parameter settings of the experiment are described. In Section 4.1, we introduce the parameter configuration and scheduling results of the RBRA algorithm when the node resource difference is an order of magnitude 10. In Section 4.2, we show the scores of the RBRA algorithm under different cluster sizes and the imbalance of the cluster.
For a cluster, the cluster standard deviation reflects the balance of cluster resources. It can be
obtained by calculating the number of Pods on the cluster which is the same as the way used by Lin [
the number of pods on nodej ( ∈ ).
, J represents the number of nodes, ( )means Then the average of each node is √∑ ∈ ( ( ) − dispersed the resource distribution of the cluster is. 4.1.
Through the Kube-scheduler-simulator and local Kubernetes cluster environment, the algorithm proposed in this paper is well verified. The experimental machine is Lenovo Blade 7000, processor Intel Core I7-9700K, memory 32G, hard disk capacity 1T, operating system Windows10, CentOS7.9, software VmWare.
When nodes with orders of magnitude difference appear in the cluster, they can be randomly scheduled to achieve balanced scheduling of resources by RBRA. First, we set up a master with two nodes. The allocable resource of the Node0 is CPU 4000m (m is one-thousandth of a core), Memory 32GB, Storage 400GB, and the allocable resource of the Node1is CPU 40000m, Memory 320GB, Storage 4000 GB. Then the resource requested by Pod is CPU 300 m Memory 1GB Storage 10GB. The default algorithm continues to deploy Pods to Node1, and our algorithm can implement scheduling to Node0 or Node1 after introducing the coefficient of variation, which can better disperse Pods and make cluster resources more balanced.
We set clusters with different numbers of nodes and conduct scheduling experiments. The results are displayed in 4.2 4.2.
Pods distribution on nodes -Default algorithm 0 20 40 60 80 100 120 140 node 8 node 7 node 6 node 5 node 4 node 3 node 2 node 1 node 8 node 7 node 6 node 5 node 4 node 3 node 2 node 1
Pods distribution on nodes -Our algorithm
As shown in Figure 3 and Figure 4, the range of the default algorithm scheduling is larger, and its degree of dispersion is larger than that ours. Our algorithm makes the distribution of Pods on the cluster more balanced. We will next set and calculate the score of the cluster.
Experiments, statistics and analysis of the above two algorithms are carried out respectively. Spread Score=P1* P2…*Pn n represents the number of nodes, and Pn represents the number of Pods on the nth node. Table 1 shows the Spread Score. 3 134 30996 64260 5 206 26911170 51705024 8 382 618271898400 2653862215680 Compared with the default algorithm, our algorithm avoids fragmentation of cluster resources, and the Spread Score is higher according to table 1. Number of Nodes
Default
Ours 8
In this paper, by introducing the coefficient of variation, we propose the RBRA algorithm. When the order of magnitude difference of cluster resources is large, the resources can still be allocated to a node in a balanced manner, while ensuring that the resource utilization within the nodes is also balanced. It solves the problem of single policy mismatch caused by the default algorithm, resulting in a high load on the CPU, memory, etc. on the node, and uneven use of cluster resources, reducing the node failure rate. We conduct simulation experiments based on the local environment. The RBRA algorithm can better achieve balanced scheduling for clusters with large differences in the order of magnitude of resources, avoiding the devastating impact of a single node failure. According to the experimental data, it has a better application effect on the balanced scheduling of pods of different cluster sizes.