To overcome network ossification, network slicing has been proposed as a promising method in 5G network. Software defined networking (SDN) and network function virtualization (NFV) as emerging enabled technology support multiple logical networks sharing physical infrastructure. Virtual network embedding (VNE) is one of the main challenges for network slicing. To avoid link congestion and satisfy the latency requirement of virtual network requests, latency model of network and load balancing link weight formulation are proposed. Then we design a load balancing link reconfiguration embedding algorithm and latency-aware load balancing VNE algorithm. Compared with the baseline algorithms, simulation results show that our proposed algorithm has better performance.
Network slicing[
The corresponding virtual network (VN) is essentially the deployment of resource allocations
optimization by considering network and computing resources. The mapping from the virtual networks
to the substrate infrastructure is one of the main challenges referred as virtual network embedding
(VNE)[
5G emerging high-bandwidth and low latency applications have driven efficient VNE strategy to
satisfy the Quality of Service(QoS) of user requests. Virtual link scheme without considering load
balancing may cause link congestion. Authors in [
Some critical nodes and links as primary selection to satisfy the latency will cause link congestion, thereby the literatures above addressed the problem by load balancing embedding or reconfiguration. However, they don’t consider latency affect for sensitive applications. In this work, we design latency model to satisfy the requirement of latency-sensitive virtual network requests and formulate load balancing link weight. Then, we have proposed a load balancing link reconfiguration embedding algorithm (VLR-E) and latency-aware load balancing VNE algorithm (LALB-VNE).
The remainder of this paper is organized as follows. Section 2 introduces the network model and formulates load balancing link weight. In Section 3, we detail the proposed VLR-E and LALB-VNE algorithm. We evaluate our proposed algorithm by simulation in Section 4. Finally, we conclude the paper in Section 5.
Substrate network use optical connected by reconfigurable optical add-drop multiplexer (ROADM) to accommodate online virtual network requests. The network uses OXC and data center links establish transparent all-optical without considering router to reduce network latency. In this section, we describe network model and load balancing link weight formulation.
The substrate network is directed by an undirected graph G = (N, E),where N is the node set {1, . . n}and E is the link set(m, n). Each node has Com(n) computing resource, and the location of each node is (locX(n), locY(n)). Each optical fiber (m, n) has W wavelength and C is the capacity length of link (m, n) and D represent the distance of two EDFA. of each wavelength. is the number of EDFA, where, = ⌊ ⁄ − 1⌋ + 2, Node process and transmission of link latency are considered as link latency, is the = 2( + ) + ∙ + ∙
+ 2 converts request into an optical signal for transmission. . is the delay of a transponder that is the delay of FEC coder-decoder processing module. is the delay of ROADM node. Optical fiber transmission delay /km is the main components of link delay. The delay of EDFA to enhance signal during transmission is . The delay of regenerator and dispersion compensation is not considered in this section.
The ℎvirtual network request is directed by an undirected graph = ( , ), where is the virtual node set and is the virtual link set. For virtual node s in , the computing resource request is , and the location is (vlocX(s), vlocY(s)). The bandwidth request of virtual link (s, d) ∈ , where the maximize acceptance latency is
.
To avoid link congestion, we design a novel load balancing link weight update mechanism by utilizing standard deviation of link bandwidth resource consumption. The bandwidth resource occupied is mapped onto substrate link (m, n). of substrate link (m, n) is described in Eq.1, where equals 1 if virtual link (s, d) of request r
Standard deviation of link load is described in Eq.2, where ̅ is the mean value of link load occupation. where links
The link weight is defined as Eq.3 both considering standard deviation of link load and link latency, is the initial wavelength and capacity of link (m, n), α and β represent the weight of link unitization and latency ratio, 0 < α, β < 1, α + β = 1. The Maximum link latency for all substrate is shown in Eq.4.
= ∑ ∈ ∑( , )∈
∙ σ = √(∑( , )∈ −̅)2
| | = + + = {
, ∀( , ) ∈ }
In this section, for online virtual network requests, latency-aware load balancing VNE algorithm
(LALB-VNE) is proposed, which is one-stage embedding strategy. For virtual node embedding, we use
Global Topology Resource (GTR) node ranking for substrate node and virtual node embedding
according to Ref.[
Since virtual node embedding results may cause virtual link embedding failure, in this section, VLRE algorithm is proposed to improve the link embedding performance by considering path splitting, shown in Algorithm 1.
Algorithm 1
VLR-E Input: G = ( , ), (s, d), m, n, , ; Output: virtual link embedding results 1. update link weight, establish link auxiliary graph; 2. search the shortest path satisfy the bandwidth
and latency requirements; 3. if successful then 4. update bandwidth state of substrate network and embedding state of virtual link;
return ; 5. 6. else 7.
set n; 8. 9. 10. 11. 12. 13.
calculate satisfied candidate path
between substrate node m and else search split path satisfying latency in ; if successful then
update bandwidth state of substrate network and embedding state of virtual link; return ;
return link embedding 14. 15 end if
failure; end if
For online virtual network requests, we proposed a LALB-VNE shown in Algorithm 2. First, generate a set of upcoming requests according to time window, and release resource if there exists request leave. Then, for each request execute one-stage virtual network embedding according GTR virtual node and VLR-E embedding strategy. resource; do generate a set of virtual network requests if there exists requests leave, release for all virtual network requests ∈
sort all virtual nodes according to GTR, obtain ; for all virtual node in
do sort substrate node according to GTR; calculate the location and computing resource for all substrate nodes, and establish candidate set;
select the first node in candidate set for embedding; if virtual node embed successful then for virtual link do
get the corresponding substrate node
m and n for successful embedding;
execute Algorithm 1 for virtual link embedding;
update network resource;
end for refuse the request else end if end for end for 21. end for Algorithm 2
do
In this section, we present simulation result to investigate the performance of the proposed
LALBVNE algorithm. In the simulation, we consider NSFNET and USNET as network topology. The number
of wavelength W is in range [50, 80] and computing resource Com(n) is in range [400, 500] unit.
Substrate network parameter is setting according to Ref.[
We assume VN request arrival follows Poisson distribution with the mean of [
Virtual nodes number of per request is generated in the range of 2 to 4. And the computing
requirement is in range [
In this paper, online GTR-based two-stage VNE algorithm (GTR-T-VNE) and GTR-based one-stage VNE algorithm (GTR-O-VNE) are designed as benchmark.
Figure 1-3 shows the simulation results in NSFNET. The acceptance ratio is shown in Figure 1, proposed LALB-VNE is 20% improvement than GTR-O-VNE and GTR-T-VNE. Revenue to cost ratio(R/C) is shown in Figure 2, proposed LALB-VNE is 1%-2% increased. Average latency is shown in Figure 3, proposed LALB-VNE can decrease 2-2.5ms latency.
Figure 4-6 shows the simulation results in USNET. Figure 4 shows that the acceptance ratio of proposed LALB-VNE has about 2% improvement than GTR-O-VNE and GTR-T-VNE. Revenue to cost ratio of proposed LALB-VNE achieve up to 6% improvement shown in Figure 5. The average latency decrease about 1.35ms for proposed LALB-VNE shown in Figure 6.
This paper has investigated the load balancing VNE problem by considering latency sensitive applications. Specifically, we have described network model and proposed a latency model to calculate the network latency. To avoid link congestion, load balancing link weight formulation has been proposed. Further, VLR-E algorithm has been proposed based proposed load balancing link weight. Then a LALB-VNE algorithm is developed to satisfy the latency sensitive applications of VN. Simulation results have demonstrated that our proposed algorithm has better performance compared with the baseline algorithms. In the future work, the intelligent algorithm such as particle swarm optimization or ant colony optimization algorithm to improve network performance furtherly.
The authors thank the Power China Huadong Engineering Corporation Limited Research project. And thank the contribution of related colleague.