Computing tasks assignment optimization among edge computing servers via SDN
SDN-based optimization of computing task allocation among edge computing servers
As an extension of cloud computing, edge computing has become an important mode for handling new service scenarios of the Internet of Things (IoE) under 5G, especially delay-sensitive computing tasks generated by edge devices . Edge computing provides critical support that meets the delay-sensitive characteristics by deploying servers at the edge of the network. However, a large number of uneven distributed computing tasks at different network edges usually lead to a bottleneck in the task processing delay of a single-edge computing server . Task allocation is mainly based on the local ECS status without considering the global network view, which can easily lead to unbalanced task load among multiple ECSs. This article introduces the novel network idea of software-defined network to the edge computing model. The logically highly centralized control plane is composed of multiple physically distributed control centers in order to coordinate computing tasks in a global view. In order to optimize task allocation and minimize task processing delay , this paper proposes three schemes . First proposedScheme for evaluating the computing characteristics of ECS tasks, And then proposedA plan for predicting the processing time of a unit task in the future of ECS. Therefore, different types of computing tasks can be assigned to appropriate electronic control units that can better handle these tasks while minimizing processing delays. In addition, designedOptimizing the plan for evaluating the time cost, In order to further improve the efficiency of task allocation. Experimental results show that compared with the prior art, this mechanism can more effectively optimize task allocation and minimize task processing delay. Specifically, compared with the corresponding work, this mechanism can increase the average unit task processing delay and ECS load balance by approximately 14% and 23%, respectively.
本文关键词：Edge Computing和SDN
边缘计算主要用来实现边缘设备产生的延迟敏感计算任务
SDN主要用来解决ECSs负载均衡问题
延迟敏感计算任务是什么类型的任务？
个人理解 这类延迟敏感计算任务意味：对低延迟要求较高的任务
三种方案来优化任务分配和最小化任务处理延迟：
评估ECS任务计算特征的方案
预测ECS未来单位任务处理时间的方案
优化评估时间开销的方案
衡量系数：
平均单位任务处理延迟
ECS负载均衡度
负载均衡：
负载均衡建立在现有网络结构之上，它提供了一种廉价有效透明的方法
扩展网络设备和服务器的带宽、增加吞吐量、加强网络数据处理能力、
提高网络的灵活性和可用性。
负载均衡（Load Balance）其意思就是分摊到多个操作单元上进行执行
例如Web服务器、FTP服务器、企业关键应用服务器和其它关键任务服务
器等，从而共同完成工作任务。
With the rapid development of network communication technology, people's demand for high-performance services is becoming more and more urgent. The fifth-generation mobile system (5G) is gradually integrating into people's daily life. It is committed to meeting the characteristics of ultra-high flow density, ultra-high connection density and ultra-high mobility [1]. Therefore, the requirements of new service scenarios have undergone significant changes, and these changes sensitively represent delay characteristics, including reliable control, fast movement, real-time analysis, and location awareness [2, 3] . Some research has been done on optimizing the scheduling of computing tasks among multiple servers in the cloud center (Section 2), which significantly improves the task processing time. However, cloud computing servers are usually located far away from the edge of the network . Considering that with the development of service scenarios under 5G, the number of computing requests from mobile devices is increasing rapidly, and traditional cloud computing may cause higher transmission delays and severe congestion. Obviously, the traditional cloud computing model that focuses on large-scale computing tasks is no longer suitable for this situation. As an extension of cloud computing, edge computing deploys functions such as computing, network, and storage near data resources at the edge of the network [4]. Its distributed edge computing servers are closer to users in order to provide real-time computing services nearby .
作为云计算的衍生，边缘计算服务器处于网络的边缘部分，
接近用户和客户端，更能实时提供服务。
However, according to the edge computing model, distributed ECSs are located in different edge network domains. Each ECS mainly performs calculation tasks based on its local network status. When a single ECS receives a large number of uneven distributed computing requests from its terminal, this often occurs in current network service scenarios, which can easily lead to a delay bottleneck for a single ECS to process computing tasks [6]. The ECS on the other end may be idle at this time. In this paper, we propose that in a global view, multiple computing centers are coordinated to distribute computing tasks that are non-uniformly distributed at different endpoints. In this method, the task is assigned to the appropriate ECS, and the working status of the ECS and the network status are considered at the same time, which not only further optimizes the task processing efficiency, but also balances the workload of the ECS.
全局视图，在多个计算中心之间协作分配非均匀分布在不同端点的计算任务,
要考虑此时的ECS工作状态和网络状态。
防止单个ECS计算任务遇到瓶颈，而此时还有其他的ECS却idle。
Software-defined network (SDN) [7] is a novel network concept, which has the characteristics of simplicity, flexibility and openness, which is naturally suitable for the needs of 5G service scenarios [8, 9]. By introducing the decoupling control and forwarding of SDN into edge computing modeling, the logically highly centralized control plane can be composed of multiple physically distributed ECS [10]. Therefore, the task processing mode of edge computing can be well improved, which is based on the local information of each ECS to manage services and resources [11]. This article introduces the idea of sustainable development network into global task assignment. By taking advantage of the sustainable development network, some research has been done to optimize the distribution of edge computing tasks among multiple computing centers on a global scale (Section 2). Therefore, while improving task processing efficiency, the workload of the electronic control system is well balanced. But in fact, the working scenarios and calculation modes of ECS with different ENDs are usually regional . In other words, the electronic control system of a certain terminal may often handle certain types of computing tasks , which enables ECS to process these corresponding tasks better than other ECSs in less time. Task processing efficiency can be further optimized by assigning tasks to the appropriate ECS. In this article, we have designed a scheme,Through mining and analyzing the historical records of the tasks processed in the pastTo evaluate the task processing characteristics of each ECS (that is, the frequency of processing specific types of computing tasks). In this method, candidate ECSs suitable for processing a class of computing tasks can be initially obtained, which enables tasks to be assigned to more suitable ECSs while further reducing their processing delay.
通过分析以往的处理过的任务的历史记录来评估ECS适合处理该类型（或者说特征）
的任务，以便下次再处理这类任务将更快的将具有该种特征类型的任务分配给ECS。
Among the many candidate ECSs, how to choose the most suitable one to handle the tasks waiting to be assigned remains a challenge. Although ECS is evaluated as suitable for one type of task, it can still cause serious delays to newly assigned tasks. For example, the current available computing power of ECS cannot afford new tasks , so the task will not be processed immediately after allocation until enough computing power is released. In addition, the newly assigned tasks may have a negative impact on the current computing efficiency of ECS. For example, the increased processing time brought to the ECS by the newly assigned tasks may reduce the future unit task processing efficiency of the ECS. In order to solve the above problems, we propose a scheme to evaluate the task processing time of ECS before assigning tasks. In this plan, the future unit task processing time of ECS is defined. It is used as an evaluation criterion for selecting ECS. By considering the past processing time of a class of computing tasks, a method for predicting the future processing time of candidate computing tasks is proposed. In this method, in the current network state, the ECS with the smallest FUT is selected as the ECS most suitable for the corresponding task.
FUT Future Unit task processing Time 未来单元任务处理time
接着刚刚上一步评估了某种ECS适合某种任务类型
接着我们将要提出如果一个新任务分配给ECS，该怎么办呢？
我们就要预计评估新任务的FUT，选择最小的FUT的ECS
However, a task may last for several rounds of evaluation without being assigned. Before each round of new allocation, it is necessary to estimate its future processing time, which obviously will bring a lot of additional re-estimation time overhead. In this article, we further design a method to solve the problem of re-estimating the time cost of tasks participating in multiple rounds of allocation. Therefore, before each round of task allocation, re-estimation operations are greatly reduced, which further improves the overall task allocation efficiency of multiple ECSs.
这里还有一个问题就是任务被连续好几轮估计FUT，就带来额外的开销时间
In this paper, based on the above design scheme, based on the advantages of SDN, a mechanism ACTE is proposed for the collaborative allocation of computing tasks among multiple ECSs. The main contributions are as follows:
The mechanism of collaboratively Assigning Computing Tasks
among multiple ECSs (ACTE)
1.给出了ACTE机制的系统框架，给出了整个系统的详细工作原理和流程。
2.设计了一个评估ECS的计算任务处理特征的方案，以便为不同类型的待
分配任务获得候选环境控制系统。
3.提出了预测ECS未来工作时间的方法，从而为相应的任务选择最合适的、
具有最小未来工作时间的ECS。
4.设计了优化评估时间开销的方法，以进一步提高任务分配效率。
system framework
Tasks collaboratively assignment
ECS'S COMPUTEING FEATURES ASSESSMENT (ECS computing task feature evaluation)
对于不同类型的计算任务，我们首先确定能够更好地处理哪种类型任务的初步合适的ECS。
在这种方法中，可以评估第(t+ 1)个时间段内每个ECS的计算特征(即处理特定类型任务的频率)。
In short: use the frequency of processing certain types of tasks to describe the types of tasks that ECS is suitable for
i represents the ECS are i calculated tasks
subscript indicates the type of task is
the APF i K (t) represents the ECS i K kinds of physical frequency of the type of service time period t
above is the molecule in the ECS i within a time period t The number of k types of services processed
The following denominator is the number of all types of services processed during the period t on ECS i
Here is the model. The
left side of the equal sign is the degree of adaptation to ECS expressed by frequency. The
Greek letters on the right are regression coefficients and constants
. The historical frequency records from t to 2t are listed in equation
1. What does it mean should be the coefficient of Y.
I understand that it is to solve the equation
. G is the
actual value of the augmented matrix Y, and the estimated calculated value of the GX T formula
EX expectation. The
derivative is equal to 0 and the minimum is taken
Unit task processing time minimum (minimize unit task processing time)
The control plane receives a large number of calculation requests in a short period of time. Computing tasks that are sensitive to delay should be appropriately allocated as soon as possible. According to the above scheme, a group of ECSs that are good at handling different types of tasks are obtained. How to assign tasks to the most suitable ECS among multiple candidate ECSs.
选择最合适的ECS的目的：
如果任务被分配给它，最小化ECS未来的单位任务处理时间。
同时，分配的任务应该在截止日期前尽快完成。
ST unfinished task set
TCC total computing power (amount)
TPD task deadline
RCC task calculation required
The purpose here is: to minimize the future completion unit
G represents the function of the evaluation task processing evaluation
numerator representationProcessing evaluation value of k types of services (value unit is time)withEstimated value of unprocessed tasksSum and
denominatorUnprocessed service quantity setwithk service types quantity setThe sum
of the additional overhead time WD waits for a new task. The
Greek letter represents the percentage of the current task processing progress.
Through eq 12 is the ECS that selects the earliest task to complete.
Task processing time estimation
As mentioned above, the task evaluation function must be proposed. How is it calculated?
First consider two factors
- One factor is the usual processing time of the task itself in the past t period of time,
- Another factor is the possibility of time fluctuations caused by handling exceptions
The first one on the right side of the equation represents the average processing time of K tasks in the t period,
C is the influencing factor. The
right side is the difference between the average normal time and the abnormal situation,
p is the abnormal period
n is the total number of processing in the m period, q It is the number of exception handling in m time period.
The brackets are abnormal minus normal
Estimation time overhead optimization
The situation mentioned above is that in the last round of allocation, some tasks may not be successfully allocated to an ECS. However, their processing time in ECS has been estimated. This obviously leads to unnecessary time overhead of re-estimating the processing time of these tasks in a new round of allocation. We designed a method to reduce the estimation time overhead and further improve the allocation efficiency, as shown below:
To put it simply: how to reduce unnecessary overhead
for tasks that have been estimated by FUT but not successfully assigned . The superscript r indicates the previous round of allocation, and the superscript c indicates the current round of allocation. If in the previous round of work assignment, the future unit task processing time increment assigned to ECSi SCTk is higher than the future unit task processing time increment assigned to ECSi by one of the other tasks waiting to be assigned in the current work assignment, then In the work distribution, it will not be estimated for ECS.
Algorithm complexity analysis (skip)
Performance evaluation (including some experimental implementation details)
In this part, we evaluate the performance of the proposed ACTE mechanism and analyze the computational tasks assigned by collaboration. These schemes are implemented in Python , and all experiments are performed on a computer with an Intel®Core™i7-6700 CPU @ 3.40 GHz and 16 GB RAM. The network topology used in the simulation is Geant and Interroute obtained from Internet Topology Zoo [28] , as shown in Figure 3. Specifically, Geant is a network topology with 41 nodes and 65 links, and Interroute is a network topology with 110 nodes and 148 links. We divide Geant and Interroute into 4 and 8 edge network domains respectively. Each domain contains about 10 switching devices and an ECS that controls these switching devices.
One type of task is set to occupy 2 to 5 units of computing power , and its normal calculation time is randomly set to 10 to 20 units of time , and it happens in ECS that is better at handling this type of task It is half of the above time (ie, 5 to 10 units of time). ECS is set to have a computing power of 4000 units. We also assume that computing tasks are divided into 10 types, and each ECS is only better at handling 2 or 3 of them.
We compared the processed ACTE with two recent related schemes. These two schemes are distributed task offloading strategy (DTOS) and delayed-aware task allocation . DTOS is a method of offloading distributed tasks to low-load base station groups in a mobile edge computing environment, which is mainly simulated based on related work [22]. LATA is a method of making task allocation decisions through SDN to reduce task processing delay, and it is mainly simulated based on related work [27]. We use the following performance indicators to compare these three methods and evaluate their performance. The performance indicators used in the evaluation are unit task processing delay (UTPD) , task processing time optimization rate (TPTOR) , ECS load balance (ELBD) , task Migration efficiency (TME) and task migration success rate (TMSR) .