Regenerator Placement Algorithms for Cloud-Ready Elastic Optical Networks Michał Aibin, Krzysztof Walkowiak* Wrocław University of Technology, Wybrzeże Wyspiańskiego 27, 50-370 Wrocław, Poland * Tel: (4871) 320 35 39, e-mail:
[email protected] ABSTRACT This paper focuses on the problem of regenerator placement in cloud-ready Elastic Optical Networks (EONs). We assume that the EON supports different modulation formats (MFs) under impairment constraints that are represented as the maximum transmission reach of the MFs. High spectrum effective MFs provide shorter transmission range what in the case of larger distances triggers the need to use regenerators. We propose new algorithms for regenerator placement in EONs. To address the fact that cloud computing services are provisioned by data centers (DCs) with the use of anycasting, the proposed algorithms include information on DC location and ratio of anycast traffic in the network. Evaluation of proposed algorithms is made in the context of dynamic traffic scenario with the objective to minimize blocking probability experienced in a network with regenerator location yielded by a particular placement algorithm. Results of experiments show that the cloud-aware location of regenerators can significantly reduce the blocking probability as well as reduce the number of used regenerators comparing to classical approaches. Keywords: elastic optical networks; dynamic routing, optimization, regenerator placement 1. INTRODUCTION Driven by the increasing communication needs worldwide, the next-generation optical networks are expected to provide huge bandwidths as well as support diverse service demands. At present, the surge of data transfer is unavoidable, posing evident obstacles for transport networks. New technology needs to be mobilized to address the rising requirements in backbone networks, introducing up to 400 Gb/s [1]. Elastic Optical Networks (EONs) uses OFDM technology [2] to allocate subcarrier slots in the optical spectrum; this results in high bandwidth efficiency. The escalating abilities of computer networks gain the requirements for dynamic routing. In EONs, the dynamic routing involves the establishment of lightpaths for individual connection requests. The process is accompanied by solving the problem of Routing, Modulation and Spectrum Allocation (RMSA) [3], [4]. In this paper, we focus on the regenerator placement problem in the context of dynamic routing of anycast and unicast traffic. The location of regenerators in the network is crucial to reduce the load on the most frequently used links and in consequence affects the key performance metric defined in the context of dynamic routing, namely, blocking probability. We focus on anycast traffic due to the growing popularity of services provided by various DCs placed in the network (e.g., content delivery networks, cloud computing, etc.) [5]. In order to satisfy the principle of anycast routing in EONs, we assume the following conditions as in [6]. The main contributions of this paper are new algorithms of regenerator placement in EONs provisioning both anycast and unicast traffic. Moreover, we show how the location of regenerators provided by these algorithms influences performance of dynamic provisioning of anycast and unicast demands in terms of blocking probability and relationship between network size and the number of required regenerators to achieve a particular level of blocking probability. According to the best of our knowledge, this paper is the first one that addresses the regenerator placement problem in the context of cloud-ready EONs and proposes algorithms that adaptively combine information on network utilization, transmission distances location of DCs and traffic type. Earlier regenerator placement problems were raised in the context of offline optimization and static routing, and the placement of the regenerator was decided upon routing and spectrum allocation [7], [8]. In our case, we place regenerators separately, prior to running the dynamic routing. The rest of the paper is organized as follows. Section 2 presents algorithms. In Section 3, we present our results. Lastly, in Section 4, we conclude our work. 2. REGENERATOR PLACEMENT ALGORITHMS In this Section, we propose several algorithms for the placement of regenerators in EONs with both anycast and unicast requests. The objective of these algorithms is to provide a placement of regenerators guaranteeing good performance of dynamic routing algorithms in terms of Bandwidth Blocking Probability (BBP) defined as the volume of rejected traffic divided by the volume of all traffic offered to the network. Therefore, the presented algorithms will be evaluated under dynamic traffic scenarios using various RMSA algorithms. 2.1 Notation We use similar notations as in [9]. The physical network is modeled as graph G(V,E,B,L) where V denotes a set of nodes, E is a set of fiber links, each fiber link can accommodate B frequency slices (slots) at most, and L =
[l(1),l(2),...,l(|E|)] represents link lengths for each e ∈ E. We assume that R DCs are already located at some nodes of the network. Nodes with DC are labeled as vR whereas on the on the other hand nodes without DC are labeled as v. Moreover, we assume that set E(v) includes all links adjacent to node v. 2.2 Notation First algorithm – called Fixed Regenerator (FR) – uses a fixed regenerator placement approach that is a very common method in the literature. To be more precise, a specific (fixed) number of regenerators are placed at each node in the network. This approach results in equal treatment of all nodes in the network, but does not consider additional information following from anycast traffic features or the network topology information in terms of link lengths. This algorithm is used as a reference method in our experiments. Algorithm 1. Fixed Regenerator Placement Algorithm
1 nregs = number of available regenerators 2 for each v ∈ V do 3 nregsv = ⎣nregs / V ⎦ 4 end for
Algorithm 2. Distance Adaptive Regenerator Placement Algorithm
1 Set x = ∑e E l(e), xv= ∑ e E(v) l(e), nregs = number of available regenerators 2 for each v ∈ V do 3 nregsv = ⎣ nregs × (xv / x) ⎦ 4 end for ∈
∈
The second algorithm takes into account a particular network and automatically calculates location of regenerators. It is named Distance Adaptive Regenerator Localization Algorithm (DA) and is based on the length of paths between network nodes. In the first step (line 1), we set the basic information about the network, i.e., we save x as the sum of the lengths of all links in the network, xv as the sum of link lengths connected to a given node and the number of regenerators nregs available to be deployed. For each network node (lines 2-4), we count the number of regenerators to allocate as the floor of the total number of available regenerators nregs multiplied by the division of the lengths of paths xv and x. The last two algorithms implement an innovative approach, which facilitates additional information on network traffic. In more detail, we assume that the AR (anycast ratio) defined as the ratio of anycast traffic in the whole traffic is known. Parameter AR for a particular network can be obtained by analysis of the network traffic in some period preceding the moment when the regenerator placement problem is solved. As in the DA algorithm, we take into account the network topology information. A new prioritization mechanism for nodes hosting DC is used, when calculating the amount of regenerators for each node in the network. We use AR parameter to calculate nregs (lines 3 and 6). First, we calculate number of regenerators for nodes with DC (lines 2-4), then for other nodes in the network (lines 5-7). Algorithm 3. Increase Data Center Priority Regenerator Placement (IDCP)
1 Set x = ∑e E l(e), xv= ∑ e E(v) l(e), nregs = number of available regenerators 2 for each vR ∈ V do 3 nregsv = ⎣ [nregs × (xv / x)] × (100% + AR) ⎦ 4 end for 5 for each v ∈ V do 6 nregsv = ⎣ [nregs × (xv / x)] × (100% - AR) ⎦ 7 end for ∈
∈
R
Algorithm 4. Decrease Data Center Priority Regenerator Placement (DDCP)
1 Set x = ∑e E l(e), xv= ∑ e E(v) l(e), nregs = number of available regenerators 2 for each vR ∈ V do 3 nregsv = ⎣ [nregs × (xV / x)] × (100% - AR) ⎦ 4 end for 5 for each v ∈ V do 6 nregsv = ⎣ [nregs × (xV / x)] × (100% + AR) ⎦ 7 end for ∈
∈
R
To evaluate the proposed algorithms, we use two RMSA algorithms for dynamic routing – Shortest Path Aware Algorithm [9] and Maximum Network Connectivity [10]. Furthermore, we compare the results achieved by these algorithms for different placement algorithms in terms of Bandwidth Blocking Probability (BBP) and the use of regenerators, which affects the cost of the network. 3. NUMERICAL RESULTS 3.1 Simulation setup In the simulations, we use a pan-European Nobel-EU network that contains 28 nodes and 82 directed links (called Euro28), a US national backbone network consists of 26 nodes and 84 directed links (called US26) and DT14, which is German National network, containing 14 nodes and 46 links. Note that the average link length is 610 km, 754 km, and 154 km for Euro28, US26 and DT14 networks, respectively. We assume that 7 DCs are located in each network for Euro28 and US26; for DT14 there are 4 DCs. Each network has three interconnection points to other networks (e.g., locations for submarine cable landing stations) used to carry the international traffic. The data provided by [11] determines the location of DCs and interconnection points. We take into consideration the physical impairment of links and we use regenerators to amplify the signal in the links that require higher MFs. The location of the regenerators is set before simulation of the dynamic routing. Number of pre-calculated number of regenerators is 1000 for US26 and Euro28 networks and 100 for DT network.
We assume that various MFs can be used in the EON. Let M denote a set of available MFs, where M = {BPSK, QPSK, m-QAM} and m belongs to {8,16,32,64}. Moreover, the MFs can use different spectral efficiency (expressed in [bit/s/Hz]). It should be noted that with a larger value of the spectral efficiency, less resources of spectrum are required to provision a particular lightpath. Nevertheless, a higher value of an optical signal to noise ratio (OSNR) is required to detect the transmission. This significantly reduces the transmission reach of the modulation format that triggers the need to use regenerators and increases CAPEX/OPEX network cost. We use a transmission model proposed in [12], which estimates the transmission distance in a function of the selected modulation level and transported bit-rate. Finally, we do not allow grooming in the regenerators, therefore one regenerator serves one request at the same time.
1,0E+00
Bandwidth Blocking Probability
Bandwidth Blocking Probability
3.2 Results At the beginning, we compared the performance of two dynamic routing algorithms: SPA and MNC in terms of BPP for different regenerators’ placement algorithms. Fig. 1 presents performance of SPA algorithm in Euro28 Network.
1,0E-01 SPA$%$FR$ SPA$%$DA$ 1,0E-02 SPA$%$DDCP$ SPA$%$IDCP$ 1,0E-03 300
400
500
600
700
800
900
1000
1,0E+00'
MNC'%'FR'
1,0E%01'
MNC%'DA' MNC'%'DDCP' MNC'%'IDCP' 1,0E%02' 300'
400'
500'
600'
Erlangs
700'
800'
900'
1000'
Erlangs
Figure 1. BBP for SPA Algorithm, Euro28 Network Different Regenerator Placement Algorithms
Figure 2. BBP for MNC Algorithm, Euro28 Network Different Regenerator Placement Algorithms
1,0E+00
1,0E+00
1,0E-01
Bandwidth Blocking Probability
Bandwidth Blocking Probability
As we can observe in Fig. 1, the best algorithm is DDCP. Due to the small distance between client nodes and DC nodes, the DDCP algorithm allocates more regenerators to nodes without DC, taking into consideration that the nodes with DC are end nodes of anycast requests that do not need frequent regeneration of the signal due to shorter path lengths used for anycast requests. In addition, the observation that allocating more regenerators to nodes without DCs is more effective is confirmed by performance of the algorithm IDCP, which uses an approach opposite to DDCP. As we can see in Fig. 1, increasing the number of regenerators in the nodes with DC causes a significant increase in BBP, even exceeding the results of performance of the FR algorithm. The second best algorithm is DA, which confirms our earlier prediction of the distribution of regenerators in the network. As shown by the observations of previous papers, deterioration for higher values of FR is caused by a lack of effective use of regenerators in the network. The same trend as in Fig 1 can be observed in Fig. 2, which shows performance of various regenerators placement methods verified by the MNC algorithm, i.e., DDCP again performs the best. Next, we checked the effectiveness of the four aforementioned regenerator placement algorithms for a larger network – US26. In Fig. 3, we can see that the improvement resulting from the use of DDCP instead of FR is around 20%, which significantly affects the efficiency of the network by reducing ratio of BBP, and hence reduces the cost of maintaining such a network. Results for algorithm MNC for US26 are similar. Moreover, we should underline that for the gap between DDCP and other reported algorithms is larger for US26 comparing to Euro28. This follows from the fact that in US26 network link lengths and in consequence also lengths of routing paths are larger comparing to the Euro28 network. Therefore, the role of regenerators increases due to need to use more regeneration as the transmission distances grow.
SPA$%$FR$ SPA$%$DA$ SPA$%$DDCP$ SPA$%$IDCP$
1,0E-02 300
400
500
600
700
800
900
Erlangs
Figure 3. BBP for SPA Algorithm, US26 Network Different Regenerator Placement Algorithms
1,0E-01 1,0E-02 1,0E-03 1,0E-04
SPA - FR SPA - No Regs
1,0E-05
SPA - DDCP
1,0E-06
1000
300
400
500
600
Erlangs
700
800
900
1000
!
Figure 4. BBP for SPA Algorithm, DT14 Network Different Regenerator Placement Algorithms
Finally, in Fig. 4 we report performance of our algorithms for a small network - DT14. We checked 3 cases regenerator placement calculated by FR and DDCP algorithms as well as where there are no regenerators in the network (No Regs). In DT14 network, for low traffic (500 ER) we can observe an interesting trend that FR
yields results very close compared to the No Regs scenario. The DDCP algorithm outperforms other methods – similarly as for larger networks. Next, we report the usage of regenerators by different algorithms. In Figs 5 and 6, we present the number of used regenerators and gain in regenerators’ usage with the use of DDCP instead of FR (i.e. 5% gain means 5% less used regenerators in network) for Euro28 and US26 networks, respectively.
350
SPA - DDCP
10%
Gain
8%
300 250
6%
200 4%
150 100
2%
50 0
300
400
500
600
700
800
900
1000
0%
Erlangs
Figure 5. Gain in regenerators’ usage, Euro28 network
600 500 400
7%
SPA - FR SPA - DDCP
6%
Gain
5% 4%
300 3% 200
2%
100
1%
0
Percentage Improvement
400
12%
SPA - FR
Percentage improvement
450
Number of used regenerators
Number of used regenerators
500
0% 300
400
500
600
700
800
900
1000
Erlangs
Figure 6. Gain in regenerators’ usage, US26 network
Observing the performance of the SPA algorithm in Fig. 5, we are able to notice that with the increase of the network traffic, the number of used regenerators increases. Therefore, when we plan to expand the network and traffic forecasts show that traffic load will be high, it is better to use adaptive algorithms of regenerators’ placement such as DDCP, because gain in regenerators usage increases with the traffic load. DDCP uses less regenerator and also achieves better BBP (refer to Fig. 1). Fig. 6 presents improvement in regenerators usage, however results are much more balanced, maintaining similar values of improvement for various traffic. 4. CONCLUSIONS In this paper, we have focused on regenerator placement problem in cloud-ready elastic optical networks. We have proposed regenerators’ placement algorithms as well as we run extensive simulations to evaluate performance of presented methods. We can conclude that the best regenerators’ placement algorithm utilizes information on DC location and ratio of anycast traffic in the network. This approach provides relatively good impact on routing algorithms in both main performance metrics: blocking probability and usage of regenerators. In the future, we plan to develop and test more sophisticated algorithms for dynamic routing of anycast requests in EONs including a case with a limited number of available regenerators. Moreover, we would like to introduce a multicast flows in dynamic routing scenarios. ACKNOWLEDGEMENTS This work was supported by The Polish National Science Centre (NCN) under Grant DEC-2012/07/B/ST7/01215 and statutory funds of the Department of Systems and Computer Networks, Wroclaw University of Technology.
REFERENCES [1]
L. Zong, G. N. Liu, A. Lord, Y. R. Zhou and T. Ma, “40/100/400 Gb/s mixed line rate transmission performance in flexgrid optical networks,” in Proceedings of OFC, 2013. [2] J. Armstrong, “OFDM for optical communications,” IEEE/OSA Journal of Lightwave Technology, vol. 27, no. 3, pp. 189-204, Feb. 2009. [3] K. Christodoulopoulos et al., "Elastic bandwidth allocation in flexible OFDM based optical networks", IEEE J. Lightw. Technol., vol. 29, no. 9, pp. 1354 – 1366, May 2011. [4] M. Klinkowski and K. Walkowiak, "Routing and spectrum assignment in spectrum sliced elastic optical path network", IEEE Commun. Lett. vol. 15, no. 8, pp. 884 – 886, August 2011. [5] M. Klinkowski and K. Walkowiak, "On the Advantages of Elastic Optical Networks for Provisioning of Cloud Computing Traffic", IEEE Network, vol. 27, no. 6, 2013, pp. 44-51 [6] K. Walkowiak, „QoS Dynamic Routing in Content Delivery Network”, Lecture Notes in Computer Science, vol. 3462, pp. 1120-1132, 2005. [7] M. Klinkowski, “On the Effect of Regenerator Placement on Spectrum Usage in Translucent Elastic Optical Networks”, ICTON, Coventry, England, July 2012. [8] T. Zami, “Co-Optimizing Allocation of Nyquist Superchannels and Physical Impairments Aware Placement of Regenerators in Elastic WDM Networks”, Journal of Lightwave Technology, vol. 32, no. 16, August 15, 2014. [9] M. Aibin, K. Walkowiak, „Dynamic Routing of Anycast and Unicast Traffic in Elastic Optical Networks with Various Modulation Formats - Trade-Off Between Blocking Probability and Network Cost”, HPSR, Vancouver, Canada, July 2014. [10] N. Wang, J. P. Jue, „Holding-Time-Aware Routing, Modulation, and Spectrum Assignment for Elastic Optical Networks”, IEEE Globecom, Texas, USA, 2014. [11] DataCenterMap. [Online]. Available: http://www.datacentermap.com [12] E. Palkopoulou et al., "Quantifying Spectrum, Cost, and Energy Efficiency in Fixed-Grid and Flex-Grid Networks," IEEE/OSA J. of Opt. Commun. and Netw., vol. 4, no. 11, pp. B42-B51, 2012.