We present the implementation and validation of an Inter-layer Traffic Engineering (TE) architecture based on a hierarchical path computation element (PCE), where the parent PCE is notified of established optical layer Label Switched Paths that induce packet traffic engineering (TE) links, thus not requiring full topology visibility. We summarize the architecture, the control plane extensions and its experimental evaluation in a control plane testbed.
© 2012 OSA
A multi-region network (MRN) , combining a packet-switching layer with an optical circuit-switching one, provides both the bandwidth flexibility and granularity of packet switching – including statistical multiplexing – and the cost-efficiency and high bandwidth capacity of the optical layer. Such a MRN enables advanced aggregation and grooming, and both the MPLS-TP and WSON technologies are mature and well positioned for such network deployments. From the operators’ perspective, a control plane provides dynamic provisioning and recovery, along with an efficient usage of resources. In particular, inter-layer Traffic Engineering (TE) refers to the process of optimizing network resource utilization globally, taking into account all layers rather than optimizing resource utilization at each layer independently .
In this context, GMPLS defines a framework for a distributed, multi-layer control plane. Within GMPLS, the preferred approach relies on a layered hierarchy mechanism by which a lower-layer (e.g., optical or Lambda Switch Capable, LSC) LSP is advertised as a TE link to be used in the upper-layer (e.g., packet or Packet Switch Capable, PSC). Such LSP is referred to as a hierarchical LSP or H-LSP, and when that announcement is within the same instance used for the establishment of the LSP such logical TE link is called a Forwarding Adjacency (FA), since it is assumed that no routing adjacency is established between its endpoints. The topology formed by lower-layer LSPs and advertised to the higher layer as TE links is called a Virtual Network Topology (VNT) .
In a MRN it is often stated that a single and unified control plane instance, with a common vision of all the switching layers constituting the network, provides the required inter-layer cooperation . Since the path computation function has full visibility of the topology and network resources, it is thus able to benefit from multi-layer TE strategies. In practice, a unified control plane in a multi-vendor setting may present scalability and/or inter-operability problems, commonly addressed by segmenting the network or defining different domains (e.g. OSPF-TE areas), which may preclude such Inter-layer cooperation. Fortunately, inter-layer TE does not require a unified control plane with full topology visibility. In general, Inter-Layer TE relies on both an optimal inter-layer path computation (I) and the automated provisioning of all involved layers (II).
2. Inter-layer path computation element
Inter-layer TE is an application domain for Path Computation Elements  in collaborative settings, with augmented functionality (e.g. a Virtual Net-work Topology Manager or VNTM), although the exact VNTM interfaces and protocols are to be defined. Requirement (I) relies on coordinated path computation and/or full topology visibility, by means of either a single PCE with topology visibility of all layers or a per-layer PCE. In the later case, each PCE knows its layer topology and relies on PCEP procedures to ensure that the optimal region boundary nodes are selected. (II) requires to provision all (server/client) layers either by: triggered hierarchical signaling, where the establishment of a client LSP triggers a server layer connection at region boundaries (Fig. 1), or layered provisioning, in which an entity (e.g., NMS or VNTM) is able to coordinate the ordered, layered establishment of server segments and finally the client layer.
3. Inter-layer traffic engineering with a hierarchical PCE
The novelty of this paper is that we propose, for the first time, to extend the framework of a two-level H-PCE  combining: i) end-to-end optimality with domain sequence selection in a multi-domain context, ii) aggregated topology management with virtual links representing intra-domain connectivity and iii) VNT management of H-LSP induced TE links at the parent PCE. The deployment model (Fig. 4) involves a child PCE (c-pce) at each domain, attached to a parent PCE (p-pce). C-pces perform path computation within their domain, and PCEP adjacencies are used for in-band notifications, topology updates, and segment expansions.
In our approach, overall, the p-pce manages two kinds of intra-domain virtual link constructs: virtual links resulting from topology aggregation mechanisms that reflect potential connectivity between border nodes, typically used by the p-pce for optimal domain sequence selection; as well as virtual links that reflect an established LSC H-LSP/FA at the lower layer that can be used for grooming. The p-pce topology (TED) is constructed by c-pces announcements: inter-domain TE links defining domain connectivity, if any, are forwarded to the parent. Virtual links for intra-domain topology aggregation are periodically computed by c-pces and forwarded to reflect mesh intra-domain connectivity . Finally, when a border node establishes a (server) LSC H-LSP and it induces a PSC TE link, such link is forwarded to the p-pce. To avoid flooding the FA within the LSC domain, the ABR may choose not to disseminate it via OSPF-TE. The p-pce differentiates both virtual types by their switching capability: H-LSP TE links are PSC and aggregated TE links are LSC. Path computation is as follows (Fig. 3): upon request for a PSC LSP, the ingress node requests a computation to its local PCE, which forwards it to p-pce. The p-pce is able to compute a multi-layer path, including requesting Virtual Shortest Path Trees (VSPTs)  at the source and destination domains for an end to end topology. If a virtual link representing an established H-LSP is selected, the PSC request will be groomed over it. Otherwise, the p-pce will select a virtual link representing topology aggregation/connectivity and it will request its corresponding segment expansion. This implies that, during signaling, the border node will set up a new optical layer LSP. Consequently, the ERO within the PCEP response contains either the unnumbered interface ID (LTII) corresponding to the H-LSP (Fig. 2(a)) or, once expanded, links from both layers (Fig. 2(b)). For hierarchical signaling, border nodes detect region changes either by inspecting the switching capability of the links in the TED, or by configuring the PCE to insert marker (SERVER-LAYER-INFO) ERO sub-objects . In summary, the layered network is isolated so no entity has the knowledge of the whole topology (note that ABRs which, by definition, have visibility of the areas they are connected to), although path optimality is insured by H-PCE procedures, and automated provisioning by triggered signaling.
4. Control plane extensions
Standard MLN/MRN RSVP-TE and OSPF-TE procedures are used in this work, extensions covering the PCEP protocol: parent/children domain relationships and mappings rely on PCE ID and DOMAIN ID TLVs included in the PCEP OPEN object. Endpoint reachability information is aggregated and announced using TLVs that contain CIDR IPv4 prefix sub-objects. For topology summarization, within a domain, domain border nodes (i.e., ABRs) are learnt from Summary and External OSPF LSAs and forwarded to the parent. In all cases, PCEP Notifications are used to wrap topology updates in the vertical direction, using TLVs (Fig. 4, top).
5. Experimental performance evaluation
The system has been implemented and validated using the GMPLS/PCE control plane platform of the ADRENALINE testbed with 14 nodes (Fig. 4). MPLS-TP links have a maximum bandwidth of 10 Gbps, and optical TE links have 8 wavelengths @10 Gbps. ABRs have 2 transceivers, allowing up to 2 active LSC LSPs. The c-pces are co-located at nodes 1, 4, and 13 and the p-pce is co-located at node 12. The inter-arrival process is Poisson with average IAT of 3s (high dynamic) or 9s (low dynamic). Holding time (HT) follows a negative exponential distribution with average depending on the selected traffic load. Each experimentation involves 104 PSC requests, selecting a source from the set 1,2,3 and a destination from 4,5 (inter-area) for a traffic demand of 1 Gbps.
Path computation latency depends on several factors; a notable one being the synchronization / mutual exclusion to access the TED (Fig. 5(a)): path computation may be delayed significantly, resulting from 25ms with no contention to 50ms when also processing topology updates.
The detail in Fig. 5(a) shows 2 lobes around 52ms, (52.3, 52.4) the former which correspond to the cases where an FA is reused or an expansion for the LSC layer is requested. Fig. 5(b) shows the obtained blocking probability. Path computation errors are triggered by the PCEs (e.g., no unreserved bandwidth) and signaling errors are due to outdated information in highly dynamic scenarios: the p-pce may construct an ERO with an FA which was removed shortly after. Control plane scalability is achieved by reducing OSPF-TE overhead, compared to 14 node/48 link TE LSAs (plus dynamic TE links) in a single domain, at the expenses of deploying the H-PCE and increased setup delay.
Finally, a simplified version of the H-PCE architecture for multi-layer path computation has been implemented and deployed in a multi-partner control plane testbed, extending the STRONGEST Distributed Control Plane infrastructure as explained in . The testbed interconnects five European research institutions, located in Madrid (Telefónica I+D), Barcelona (CTTC), Pisa (CNIT and Nextworks) and Munich (Nokia Siemens Networks, NSN), in the scope of a joint collaboration between the FP 7 IP STRONGEST Project , integrating the Optical Burst Switching (OBS) Path Computation Elements from the FP 7 MAINS Project . The path computation scenario was successfully demonstrated in , with a network environment made of several interconnected WSON and OBS domains. The WSON domains are able to provide lambda switched paths. The OBS domains are based on the Optical Burst Switching paradigm, a sub-wavelength technology able to fill wavelengths. End-to-end OBS paths starting in one domain and ending in a different domain are possible using wavelengths provided by the multi-domain WSON network (see Fig. 6).
We have presented a functional architecture to enable inter-layer TE minimizing topology visibility requirements and improving scalability. Our approach uses an H-PCE model in which the parent PCE ensures domain sequence and end-to-end path optimality leveraging existing FAs while triggering new ones using hierarchical signaling. We have implemented the solution, qualitatively validated in a testbed.
This work has been partially funded by Spanish MINECO project DORADO (TEC2009-07995), and by the EC FP7 IP project STRONGEST grant no 247674. The authors would like to thank R. Vilalta, O. González de Dios, G. Bernini, F. Paolucci, G. Carozzo, G. Landi and C. Margaria for their support of this work and the multi-partner testbed validation.
References and links
1. K. Shiomoto, D. Papadimitriou, JL. Le Roux, M. Vigoureux, and D. Brungard, “Requirements for GMPLS-based multi-region and multi-layer networks (MRN/MLN),” IETF RFC 5212 (2008), http://tools.ietf.org/rfc/rfc5212.txt.
2. E. Oki, T. Takeda, JL. Le Roux, and A. Farrel, “Framework for PCE-based inter-layer MPLS and GMPLS traffic engineering,” IETF RFC 5623, (2009), http://tools.ietf.org/rfc/rfc5623.txt.
3. R. Martinez, R. Casellas, and R. Muñoz, “Experimental validation and evaluation of a GMPLS unified control plane in multi-layer (MPLS-TP/WSON) networks,” NFOEC paper NTu2J.1 (2012), http://www.opticsinfobase.org/abstract.cfm?URI=NFOEC-2012-NTu2J.1.
4. R. Casellas, R. Martinez, R. Muñoz, L. Liu, T. Tsuritani, I. Morita, and M. Tsurusawa, “Dynamic virtual link mesh topology aggregation in multi-domain translucent WSON with hierarchical-PCE,” Opt. Express 19, B611–B620 (2011), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-19-26-B611. [CrossRef]
5. F. Zhang, D. Li, F. Javier, Jimenez Chico, O. Gonzalez de Dios, and C. Margaria, “GMPLS-based hierarchy LSP creation in multi-region and multi-layer networks,” IETF work in progress, http://tools.ietf.org/html/draft-zhang-ccamp-gmpls-h-lsp-mln-04.
6. F. Paolucci, O. Gonzalez de Dios, R. Casellas, S. Duhovnikov, P. Castoldi, R. Munoz, and R. Martinez, “Experimenting hierarchical PCE architecture in a distributed multi-platform control plane testbed,” OFC paper OM3G.3 (2012) http://www.opticsinfobase.org/abstract.cfm?URI=OFC-2012-OM3G.3.
7. ICT FP7 STRONGEST Project, “Scalable, tunable and resilient optical networks guaranteeing extremely-high speed transport,” www.ict-strongest.eu.
8. IST FP7 MAINS Project, “Metro architectures enabling sub-wavelengths,” www.ist-mains.eu.
9. Workshop on control plane architectures for new optical switching technologies enabling flexibility in time, frequency and space domains, collocated with ECOC(2012).