3.6 Online Algorithm for Dynamic Admissions of NFV-enabled Requests
3.7.4 Impact of different parameters
The rest of this section is to study the impact of important parameters: lmax, βandσ
We start with investigating the impact of the maximum numberlmaxof network functions in a service chain on the performance of algorithmsUNICAST,UNICAST_DELAY,
ONE_DC, andONE_DC_DELAYin terms of the operational costs, by varyinglmax from 3 to 6 while fixing the network size at 200. It can be seen from Fig. 3.5(a) and Fig. 3.5(b) that the operational costs of algorithms UNICASTand UNICAST_DELAYincrease with the growth oflmax, because more network functions mean more computing resource demands and higher computing resource usage cost.
We then evaluate the impact of parameterβon the performance of the two pro-
posed online algorithmsONLINEandONLINE_DELAY, by varying the value of βfrom
2|V| to 8|V|. It can be seen from Fig. 3.6(a) and Fig. 3.6(b) that algorithms ONLINE
andONLINE_DELAY will admit less requests with the increase ofβ. For instance, the
number of requests admitted whenβ=8 is no greater than 70% of the number of re-
quests admitted when β=2. This is due to the fact that the larger the value of β, the
higher the cost of using an overloaded resource will be, leading to more conservative resource usage.
We finally evaluate the impact of the admission thresholdσon the performance
of the two proposed online algorithmsONLINEandONLINE_DELAYwith and without adopting the threshold, to highlight the importance of admission control. It can be seen from Fig. 3.7 that both algorithms ONLINE and ONLINE_DELAY will admit less requests if no admission control is applied. In addition, the larger the network size, the greater the impact of the admission control threshold σ on the number
of requests admitted. This is because in large networks, the distance between the source and the destination of a request can be very large, thus requiring much more bandwidth resources to admit the request. With a given threshold, the proposed online algorithms are able to reject such a request, thereby being able to admit future requests and achieving higher network throughput.
3.8
Summary
In this chapter we first studied NFV-enabled unicasting in a Software-Defined Net- work (SDN) with and without the end-to-end delay constraints, for which we pro-
72 Routing Cost and Throughput Optimization of Requests in the Remote Cloud 3 4 5 6
l
max 0 2 4 6 8implementation cost
UNICAST ONE_DC(a) Impact oflmaxon the operational costs of algorithmsUNICASTandONE_DC
3 4 5 6
l
max 0 1 2 3 4implementation cost
UNICAST_DELAY ONE_DC_DELAY(b) Impact of lmax on the operational costs of algorithms UNICAST_DELAY and ONE_DC_DELAY
Figure 3.5: The impact of lmax on the performance of algorithms UNICAST,
UNICAST_DELAY,ONE_DCandONE_DC_DELAY.
10 50 100 150 200
network size n
100 150 200 250# of admitted requests
beta=2n beta=4n beta=6n beta=8n(a) AlgorithmONLINE
10 50 100 150 200
network size n
50 100 150 200# of admitted requests
beta=2n beta=4n beta=6n beta=8n (b) AlgorithmONLINE_DELAYFigure 3.6: The number of requests admitted by online algorithms ONLINE and
ONLINE_DELAYwith different values ofβ.
10 50 100 150 200
network size n
80 100 120 140 160 180 200# of admitted requests
with threshold without threshold(a)Algorithm ONLINE
10 50 100 150 200
network size n
50 100 150 200# of admitted requests
with threshold without threshold (b) AlgorithmONLINE_DELAYFigure 3.7: The number of requests admitted by online algorithms ONLINE and
posed a generic optimization framework. We then devised efficient algorithms for the admission of a single NFV-enabled request with the aim to minimize its imple- mentation cost in terms of both computing and bandwidth resource consumption. We also investigated the dynamic admissions of NFV-enabled requests without the knowledge of future request arrivals, with the objective of maximizing the network throughput, for which we devised efficient online algorithms. We finally evaluated the performance of the proposed algorithms through experimental simulations. Sim- ulation results demonstrate that the proposed algorithms are promising, and outper- form other heuristics.
QoS-Aware Virtual Network
Service Deployment in the
Edge-Cloud
4.1
Introduction
The emergence of Augmented Reality (AR) applications and IoT computing promises many rich location-based services such as location-aware advertising and route rec- ommendation for driverless vehicles. As a result many users in a local area may request the same type of location-based services in the near future. Instead of al- locating resources to each individual offloaded task, an instance for the service can be instantiated into a VM, known asVirtualized Network Functions (VNF), and subse- quent tasks requesting the same service can be served by the existing VNF instance, saving resource allocation costs and allowing the service to be carried out immedi- ately for new tasks. However, to meet the various demands of offloaded tasks from users and their stringent QoS requirements, network service providers must face a trade-off between over-provisioning VNF instances, or allowing new users to expe- rience a delay while resources are allocated. On the one hand, instantiating VNFs ahead of anticipated demand can save time in creating new VNF instances avoiding a delay for impatient users, but on the other hand, maintaining idle VNF instances absent demand will incur on-going operation costs for the service provider.
Provisioning network services with different types of VNFs in the edge-cloud
76 QoS-Aware Virtual Network Service Deployment in the Edge-Cloud
poses several challenges: How many instances are to be instantiated at which cloudlets such that the computing resource of cloudlets can be maximally utilized while the cost and delay of instance instantiation can be minimized? How should offloaded tasks be assigned to different cloudlets while meeting their individual QoSs? and how can the admission cost of offloading task requests be minimized by utilizing the existing VNF instances of their requested network function services? Finally, how can the number of required VNF instances be predicted? and how should the cre- ation and removal of VNF instances be managed in the network? In this chapter we will address these challenges, by comprehensively studying the problem of task offloading with network function service requirements in a mobile edge cloud. To the best of our knowledge, we are the first to explore the possibility of fully utiliz- ing existing VNF instances of network function services in cloudlets for cost-effective task offloading while meeting the QoS requirements of individual offloaded tasks, by formulating a novel QoS-aware task offloading optimization problem and providing an optimization framework. To respond to the dynamic changes of offloading re- quest patterns over time to further reduce request admission costs, we also develop an effective prediction mechanism to predict the instance demands in future through the removal and creation of numbers of VNF instances at each cloudlet.
The remainder of the chapter is arranged as follows. Section 4.2 discusses the related works to this topic. Section 4.3 will introduce the system model, notations and problem definition. Section 4.4 will devise algorithms for the problem. Sec- tion 5.5 will provide some experimental results on the performance of the proposed algorithm, and Section 4.6 concludes the chapter.