Mo.3.E.6.pdf
Application-aware and Adaptive Virtual Data Centre Infrastructure Provisioning over Elastic Optical OFDM Networks Shuping Peng, Reza Nejabati, Mayur Channegowda, and Dimitra Simeonidou High Performance Networks Group, EEE, University of Bristol, UK {
[email protected]} Abstract A novel application-aware virtual data centre (VDC) provisioning method for distributed data centres (DCs) enabled by coordinated virtualization of optical OFDM network and DCs is proposed. Furthermore, an adaptive VDC replanning method, supporting virtual topology shifting for accommodating DCs traffic variations, is proposed. Introduction Data centre (DC) providers, in order to accommodate the ever-growing highperformance network-based applications over their geographically distributed DCs, must either lease network infrastructures from network operators or build their own internal networks, e.g. Google G-scale network [1]. In order to support multi-tenant cloud platform (i.e. multiple co-existing hardware and software instances being created for different applications), Software-Defined Networking (SDN) and Virtualization are considered as key enablers [2]. Google has deployed an OpenFlow powered SDN solution in its G-scale network for carrying the inter-DC traffic [1]. Furthermore, to satisfy the high-performance requirements (e.g. highbandwidth, low-latency, etc.) of the inter/intra DC connectivity, optical network empowered by novel optical technologies is being considered as a strong candidate for transport platform [3]. In this paper, we target future cloud platform that deploys advanced optical transport technologies for interconnecting remote DCs. We propose an application-aware virtual data centre (VDC) infrastructure provisioning method over the elastic optical OFDM networks [3]. The coordinated virtualization of optical network and IT resources in DCs is developed as a key part of the provisioning method. Considering the applications and/or traffic variations, an adaptive VDC replanning method is proposed, which is able to dynamically adjust the existing VDCs for provisioning new VDC by taking advantage of the benefits achieved from SDN/Virtualization and the underlying optical technologies. Coordinated virtualization of OFDM-based optical network and distributed DCs A generic high performance cloud scenario is modelled as the lower part of Figure 1, in which distributed DCs and sets of users are connected to the transport network through access points. In order to support multi-tenancy, multiple application-specific VDC infrastructures comprising both virtual network (interconnecting
the distributed DCs) and virtual IT resources (computing and storage inside each DC) need to be created by using VDC provisioning methods.
Figure 1: Virtual DC infrastructure provisioning over the modelled Cloud scenario
The coordinated virtualization of network and IT resources is essential in the designing of the VDC provisioning methods. Some work has been done in this area. In [4], the authors have conducted extensive studies on the virtual infrastructure provisioning over single-line-rate and mixed-line-rate optical WDM networks. However, to the authors’ best knowledge, the study on VDC provisioning over OFDM-based optical network is still missing, wherein the VDC provisioning can significantly benefit from the high capacity and flexibility provided by the optical OFDM transport technology. The optical OFDM network consists of bandwidth-variable transponders (BVT) at the edge and bandwidth-variable OXCs (BV-OXC) in the core, which has been described in SLICE (Spectrum-sLICed Elastic optical network) [3]. To provision an end-to-end connectivity, a BVT will generate optical signals with just enough number of subcarriers and appropriate modulation format, while each BV-OXC along the route will allocate a switching window with corresponding optical filter width to establish an end-to-end lightpath with suitable and sufficient spectrum. Optical layer constraints (e.g. the spectrum continuity) should be taken into
Mo.3.E.6.pdf account during the establishment of lightpaths. Several routing and spectrum assignment (RSA) algorithms [5] have been proposed for setting up lightpaths in optical OFDM networks. However, for the VDC provisioning, all the virtual links (each one can be taken as a lightpath) of the requested VDC as well as the demands (i.e. the resource kinds and the capacities) on the IT resources in DCs should be considered in a coordinated manner. Based on the above analysis, we proposed a VDC provisioning method enabled by coordinated virtualization of Net + DCs over OFDM-based optical network, as illustrated in Figure 2 left. The workflow of the provisioning method is given in Figure 2 right.
Figure 2: Coordinated virtualization of network + DCs
Among all the virtual IT nodes (stored in the vector VN) of a VDC request, the one that requests the most resources will be selected first and mapped to the physical DC that has the most available resources considering the load balancing. The mapped virtual IT nodes will be put into VN_mp, while the allocated physical resources will be put into PN_mp. The adjacent IT nodes of the VN_mp are put into Adj_nd. Again the virtual IT node that requests the most resources (e.g. vn is Figure 2) will be selected, and all the potential nodes for mapping vn will be put in Pot_nd. Then the connections between vn and the nodes in VN_mp will be calculated, and the requirements on bandwidth (B) and latency (L) need to be satisfied as well as the spectrum continuity and contiguity constraints. The modulation format will also be properly selected. If all the connections can be established with enough resources and satisfied performance, vn will be put into VN_mp, while the physical resources will be put into PN_mp. If the mapped node is the last one in VN, the VDC provisioning is successful. If the connections between vn and VN_mp cannot be found, the next node in Pot_nd will be checked and goes through the path calculation and allocation process. If there is no node left in Pot_nd, the provisioning is failed. The proposed provisioning method enables the virtual IT and network resources to be mapped in a single step, e.g. vn and its two
connections to VN_mp (see Figure 2 left). Replanning with topology shifting capability for application-aware VDC Provisioning Using the above proposed method, multiple VDC infrastructures can be created and operated in parallel. Due to the dynamic nature of DC applications, the created VDC infrastructures need to be replanned (e.g. capacities to be scaled up/down, topologies to be shifted) according to the changing demands in order to improve the resource utilization efficiency. One of the most common cases would be the DCs’ backup during nights, wherein the data in all other DCs will be delivered and stored in one Backup DC, which requests to create a new VDC with big connection capacities between the Backup DC (node 4 in Figure 3 left) and all other DCs (nodes 1, 2, and 3). Since fewer resources/ capacities are needed in all the running VDCs during night compared to daytime, the reserved capacities in the VDCs can be scaled down, and the released capacities can be used to create the new VDC for the backup data transfer. Meanwhile, it is desirable if the topology of the Backup VDC can be derived from the existing VDCs in order to reduce the interactions with the physical infrastructures because of the VDC provisioning, which can significantly save the provisioning time and resources to be used. The workflow of the replanning method with topology shifting capability is depicted in Figure 3 right. The Backup DC vBDC is first located. The connections between the vBDC and all other DCs vn {NDC -vBDC} will be checked in terms of the existence and the spare capacity. If there is no existing connection or not enough capacity, the new connection with the required capacity will have to be established.
Figure 3: Replanning for VDC provisioning
Network parameters and Simulation results The COST239 European Optical Network (EON) topology is used as the physical topology. The locations of DCs (i.e. Amsterdam, Paris, Zurich, Berlin, and Copenhague) are selected referring to Interoute’s DC locations [6] across Europe, as shown in Figure 4. The capacity of each DC (computing) is set 150 to 300 units. The number of subcarriers per link is 128 or 256.
Mo.3.E.6.pdf FR DC w/o Multi-MF FR Net w/o Multi-MF FR Net with Multi-MF
Failure rate of VDC provisioning
0.6 0.5 0.4 0.3 0.2 0.1 0 150
200 250 Each DC capacity (units)
300
Figure 4: COST239 EON with 11 nodes and 26 links. DCs are marked in green. The length of each physical link is labelled, and the unit is km.
The VDC topologies are randomly generated (20 in total) with controllable parameters: the number of virtual nodes (3 or 4), the network degree (2 or 3), the probability of interconnecting virtual nodes (0.5), the bandwidth of virtual links (10 to 75 Gbps), and the requested capacity of each virtual node (20 to 30 units). In this study, we adopted four modulation formats, i.e. BPSK, QPSK, 8QAM, and 32QAM. The parameters such as subcarrier capacity (C) and spectrum (S) and the maximum transmission distance (D) with acceptable quality of transmission are given in Table I. Table I: Parameters of different modulation formats Modulation BPSK QPSK 8QAM 32QAM
C=2.5 Gbps, S=5 GHz C 2C 3C 5C
D (km) 3,000 1,500 750 375
We first evaluate the performance of the coordinated Net + DCs virtualization in terms of the failure rate (FR) of VDC provisioning caused by lack of IT resources in DCs (“FR DC” in Figure 5 top) or available spectrum (“FR Net”), and the acceptance rate affected by the total number of subcarriers (“Subcar”) and multiple modulation formats (“Multi-MF”) availability (Figure 5 bottom). In Figure 5 top, we can see that as the DC capacities increase, the constraints coming from the limited IT resources becomes relieved (black), while the network capacity becomes the dominant factor for damaging the overall performance (red). However, as we can see from the results, after the capability for choosing modulation format is enabled, the bottleneck arisen from the network is almost eliminated (blue). In Figure 5 bottom, it shows that the overall acceptance rate is increasing as the DCs’ capacities grow, and the Multi-MF capability push the improvements forward, while the effect of the further increased network capacity is negligible under the given bandwidth demands. The results proved a key point which is essential in VDC provisioning, that is, the DCs and network need to be coordinately taken into account, since they may alternately become the performance bottleneck.
Acceptance rate of VDC provisioning
1 Subcar = 256, with Multi-MF Subcar = 128, with Multi-MF Subcar = 128, w/o Multi-MF
0.8 0.6 0.4 0.2 0
150
200 250 Each DC capacity (units)
300
Figure 5: The results of coordinated Net + DCs virtualization over optical OFDM networks
In order to evaluate the effectiveness of replanning the existing VDCs for provisioning new VDC, we created the Backup VDC based on three existing VDCs that are randomly generated. 75% of the resources reserved by the existing VDCs can be used for the backup purpose. The results in terms of the requested bandwidth (Gbps) with/without replanning are compared in Table II, which were collected from five independent test cases. We can see that by using the proposed replanning method, around 62.24% (average) resources can be saved, which also means the reduced interactions with physical infrastructures when provisioning the new VDC as well as the provisioning time. Table II: VDCs Replanning (75% can be spared) Req. B. w/o Repl. Req. B. w. Repl. Res. Saving (%)
152 66 56.6
196 47 76
176 93 47.2
188 82 56.4
128 32 75
Conclusions Elastic optical OFDM network is adopted for interconnecting the distributed DCs. The impact of its rich capabilities on the VDC provisioning is investigated. The necessity of the coordinated virtualization of network and DCs is also reflected through this study. The proposed replanning strategy provides an applicationaware and adaptive method for provisioning VDC with reduced interactions with physical infrastructure and provisioning time. Acknowledgements The work has been supported by STRAUSS, PATRON, and Photonic Hyperhighway projects. References [1] Google Inc., 2012. [2] WCP Research, S. Blakley, Jan 2013. [3] J. Masahiko, et al., IEEE Mag., Aug. 2010. [4] S. Peng, et al, JOCN, March 2013. [5] K. Christodoulopoulos, et al, JLT, May, 2011. [6] https://www.linx.net/join/lfa/interoute.html
