Toolbox, hybrid IP/MPLS optimisation method and fairness Research Unit in Networking EECS Department University of Liège 13 September 005
Outline 1 3 4 5
Outline MPLS principles 1 MPLS principles 3 4 5
MPLS principles Today, most operators overprovision their network. But increasing demand and quality of service requirements means this approach is less and less tenable economically. Traffic Engineering (TE) involves adapting the routing of traffic to the network conditions, with the joint goals of good user performance and efficient use of network resources.
MPLS principles MPLS principles Multi Protocol Label Switching (MPLS) allows to establish tunnels (LSPs, label switched paths) in the topology, along explicit routes (e.g. using RSVP-TE). Routing is not only destination-based any more LSPs can be bandwidth guaranteed Advantages We can freely choose the path and the granularity of these LSPs These are independent from each other
Outline Objectives Architecture Applications 1 Objectives Architecture Applications 3 4 5
Toolbox objectives Objectives Architecture Applications Lots of research in TE domain but not used by operators because of their integration and usage complexity. Our TOTEM toolbox has two objectives: Allow researchers to promote their new TE methods and compare them with other existing ones. Allow operators to test these methods on their networks.
Architecture requirements Objectives Architecture Applications The architecture must meet several requirements: Minimise new algorithm integration effort Be interoperable with existing tools Allow the integration of algorithms written in multiple languages Allow different execution modes on-line in a real network off-line for simulation purposes
Toolbox integration Objectives Architecture Applications On-line integration Off-line simulation
Toolbox architecture Objectives Architecture Applications IGP config Methods repository MPLS config Topology IGP metric optimiser Link load analysis BGP config Link load SNMP Netflow traces Traffic Matrix LSP path computation Backup LSP path computation Path delay analysis Inter domain traffic analysis BGP decision BGP dump BGP routing table process simulation CONTROL Simulation Scenario Create a model of the netwok Optimise and simulate Analyse report
Toolbox applications Objectives Architecture Applications Analyse link load associated with a traffic matrix Simulate link failure Compare link load based on several traffic matrices Optimise the IP metric Compute near-optimal MPLS full mesh using DAMOTE Compute hybrid IP/MPLS solution using SAMTE Simulate inter-domain routing using C-BGP simulator Compute traffic matrix from netflow monitoring data Generate topology and traffic matrix using Topgen generator
Outline Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE 1 3 Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE 4
Objective Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE Objective: find a small number of LSPs to optimize a given operational need like minimize max load, load balancing,... Advantages Scalable: just a few LSPs, i.e. not a full mesh Incremental: you know exactly which traffic will be routed on the LSPs Generic: you can combine different LSPs for different objectives such as delay, max load,...
Hybrid IP/MPLS routing models Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE 5 1 6 5 1 6 5 1 6 3 4 LSP 3 4 LSP 3 4 LSP 1 F,5 F,6 1 F,5 F,6 1 F,5 F,6 F1,6 F1,6 Overlay Basic IGP Shortcut IGP shortcut All flows entering the network at node and exiting it at node 6 will be forwarded on the LSP All flows exiting the network at node 6 will be forwarded on the LSP (if they cross node before) F1,6 All flows crossing node and then node 6 will be forwarded on the LSP
SAMTE : hybrid IP/MPLS method Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE Compute a solution of K LSPs to optimise your objective Generate a candidate path list with P shortest path for each pair origin/destination. Based on simulated annealing meta heuristic Initial solution : Generate a set of K LSPs at random Neigbourhood : replace a LSP of the solution by a LSP of the candidate list Objective functions (bottleneck) : minimize maximum link utilisation load balancing (combined with shortest path)
Simulated annealing execution Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE 0.75 0.7 Simulated annealing execution Current solution Best solution Temperature 0.65 Maximum link load 0.6 0.55 0.5 0.45 0.4 0 000 4000 6000 8000 10000 Iterations
GÉANT network Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE Summary European research network 30 countries and 6 national networks 3 nodes 38 links
Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE Minimize maximum link utilisation on GEANT 0.8 0.7 Link load with differents number of LSPs Maximum link load Mean link load Std of the link load Percentile 10 of the link load 0.6 0.5 Link load 0.4 0.3 0. 0.1 0 0 5 10 15 0 5 Number of LSPs
Comparaison with other TE methods Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE Method #LSP Max Per10 Mean Std CPU time in % in % in % in % MCNF 506 41.9 - - - days SPF-GÉANT 0 70.7 3.0 7.1 11.5 0 SPF-InvCap 0 46.3.3 6.9 9.6 0 IGP-WO 0 45.1. 7. 9.6 315 s DAMOTE α= 506 41.9 16.1 8.5 7.5.5 s GÉANT + SAMTE fml 4 4.1 4.5 7.5 10.7 1.0 s GÉANT + SAMTE flb 1 4.5 18.9 7.8 8.4 3.1 s GÉANT + SAMTE flb 3 4.5 16.5 7.8 8.0 13.8 s
Description Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE Objective : When and how to re-engineer (update the set of LSPs) the network according to traffic evolution? Maximize the operational objective Minimize the number of re-optimisations Simulations done on one month of traffic matrices of Geant (computed with Netflow)
Two strategies evaluated Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE The utopic strategy: At the beginning of each period of 15 minutes, we consider that the next traffic matrix is known exactly (which is obviously impossible) and SAMTE is used to set up 5 new LSPs to replace the former 5. [Reference] The TM-max strategy: we build TM-max which is a traffic matrix composed of the maximal traffic over one month for each pair. Then TM-max will be used as input to SAMTE to compute 5 LSPs for the next month. [Proposition]
Comparison of the two strategies Objective Hybrid IP/MPLS routing models SAMTE Simulations Re-optimisation strategy for dynamic TE % 50 40 30 0 10 0-10 30/04 07/05 14/05 1/05 8/05 04/06 Time Summary Worst max load: 66.9 % - TM-max 45. % - each TM 48 % of increasing Average increase : 3.8 % Each TM rerouting In average Maximum 3.5 % of the pairs 1.6 % of the volume 11.46% of the pairs 5.3% of the volume
Outline Objectives Max-min fairness Distributed algorithm Simulation 1 3 4 Objectives Max-min fairness Distributed algorithm Simulation 5
Objectives Objectives Max-min fairness Distributed algorithm Simulation Our goal : sharing the available bandwidth among all the LSPs according to their weights The classical max-min rate allocation policy has been accepted as an optimal network bandwidth sharing criterion among user flows. Extension with a weight : WPMM (Weighted Proportional Max-Min)
Objectives Max-min fairness Distributed algorithm Simulation Weighted Proportional Max-min fairness Notations L : a set of links S : a set of LSPs Each LSPs has : a reserved rate RR s a fair rate FR s a maximal rate MR s a weight w s A fair share allocates a LSP with a small demand what it wants, and distributes unused resources evenly to the big LSPs according to their weights
Objectives Max-min fairness Distributed algorithm Simulation Proposed distributed WPMM algorithm Periodically, the ingress sends a PATH packet Each router computes a local fair share for the LSP and updates a new RSVP field (called ER, explicit rate) if its local fair rate is less than the actual fair rate Upon receiving a PATH packet, the egress router sends a RESV packet. Each router updates its information with the RESV parameters.
Simulation Objectives Max-min fairness Distributed algorithm Simulation For stabilizing 90% of the LSPs, our solution takes 4 iterations (16 with Hou s one). In the worst topology (among 63 topologies), it takes 36 iterations to converge (84 with Hou s solution)
Outline 1 3 4 5
The toolbox is now really pertinent for researchers and operators. It could become a tool to accelerate the developement of recent traffic engineering methods The proposed hybrid IP/MPLS method is a scalable and very flexible method to traffic engineer a network Other contraints could be added, like resilience, combination with IP-FRR, re-optimisation strategy This work, characterised by its pragmatic approach, focuses on real needs in current high speed networks
Publications G. Leduc, H. Abrahamsson, S. Balon, S. Bessler, M. D Arienzo, O. Delcourt, J. Domingo-Pascual, S. Cerav-Erbas, I. Gojmerac, X. Masip, A. Pescapè, B. Quoitin, S. P. Romano, E. Salvadori, F. Skivée, H. T. Tran, S. Uhlig, and H. Ümit. An open source traffic engineering toolbox. To appear in Computer Communications, 005. (18 pages) F. Skivée, S. Balon, O. Delcourt, J. Lepropre, and G. Leduc. Architecture d une boîte à outils d algorithmes d ingénierie de trafic et application au réseau GÉANT. Actes de Colloque Francophone sur l Ingénierie des Protocoles (CFIP), pages 317-33, Bordeaux, France, 9 Mar.-1 Avr. 005. Hermès Lavoisier. F. Skivée and G. Leduc. A Distributed Algorithm for Weighted Max-Min Fairness in MPLS Networks. Proc. of 11th IEEE International Conference on Telecommunications (ICT 004), 1-6 Aug. 004, Fortaleza, Brazil, J. Neuman de Souza, P. Dini, P. Lorenz (eds.), Telecommunications and Networking, LNCS, 314, pp. 644-653, Springer Verlag. S. Balon, F. Skivée, G. Leduc. Comparing traffic engineering objective functions. Proc. of the 1sh ACM CoNEXT student workshop, October 4-7, 005, Toulouse, France. ( pages)
Submitted papers S. Balon, O. Delcourt, J. Lepropre, F. Skivée and G. Leduc. A traffic engineering toolbox and its application to the GEANT network. IEEE etransactions on Network and Service Management. (11 pages) F. Skivée, S. Balon and G. Leduc. A scalable heuristic for hybrid IGP/MPLS traffic engineering - Case study on the GEANT network. IEEE International Conference on Communication (ICC). (8 pages)