Ample an adaptive traffic engineering system based on virtual routing topologies.bak


Published on

1 Like
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Ample an adaptive traffic engineering system based on virtual routing topologies.bak

  1. 1. WANG2 LAYOUT_Layout 1 2/22/12 3:28 PM Page 185 ACCEPTED FROM OPEN CALL AMPLE: An Adaptive Traffic Engineering System Based on Virtual Routing Topologies Ning Wang, University of Surrey Kin Hon Ho, Hong Kong Polytechnic University George Pavlou, University College London ABSTRACT published traffic traces dataset in the GEANT network as an illustration. The actual maximum Handling traffic dynamics in order to avoid link utilization (MLU) dynamics is substantial on network congestion and subsequent service dis- a daily basis, varying from less than 40 percent ruptions is one of the key tasks performed by during off-peak time to more than 90 percent in contemporary network management systems. busy hours [2]! As such, using one single traffic Given the simple but rigid routing and forward- matrix as input for offline computing a static TE ing functionalities in IP base environments, effi- configuration is not deemed as an efficient cient resource management and control solutions approach for resource optimization purposes in against dynamic traffic conditions is still yet to such dynamic environments. be obtained. In this article, we introduce Traffic engineering for plain IP-based net- AMPLE — an efficient traffic engineering and works (we will be referring to these as IGP- management system that performs adaptive traf- based networks, as is common in the literature fic control by using multiple virtualized routing since they route traffic based on the Interior topologies. The proposed system consists of two Gateway Protocol, OSPF or IS-IS) has received complementary components: offline link weight a lot of attention in the research community [3, optimization that takes as input the physical net- 4]. Existing IGP-based TE mechanisms are only work topology and tries to produce maximum confined to offline operation and hence cannot routing path diversity across multiple virtual cope efficiently with significant traffic dynamics. routing topologies for long term operation There are well known reasons for this limitation: through the optimized setting of link weights. IGP-based TE only allows for static traffic deliv- Based on these diverse paths, adaptive traffic ery through native IGP paths, without flexible control performs intelligent traffic splitting traffic splitting for dynamic load balancing. In across individual routing topologies in reaction addition, changing IGP link weights in reaction to the monitored network dynamics at short to emerging network congestion may cause rout- timescale. According to our evaluation with real ing re-convergence problems that potentially dis- network topologies and traffic traces, the pro- rupt ongoing traffic sessions. In effect, it has posed system is able to cope almost optimally been recently argued that dynamic/online route with unpredicted traffic dynamics and, as such, it re-computation is to be considered harmful even constitutes a new proposal for achieving better in the case of network failures [5], let alone for quality of service and overall network perfor- dealing with traffic dynamics. mance in IP networks. In recent years, the concept of virtual net- works has received increasing attention from the INTRODUCTION research community, with the general spirit being to enable virtualized network resources on top of Traffic Engineering (TE) is an essential aspect of the same physical network infrastructure. Such contemporary network management. Offline TE resources not only include physical elements such approaches aim to optimize network resources in as routers or links, but also soft resources such as a static manner, but require accurate estimation logical network topologies through configurations of traffic matrices in order to produce optimized that allow them to coexist gracefully. Our motiva- network configurations for long-term operation tion differs from the existing proposals focusing (a resource provisioning period each time, typi- on virtual network provisioning to support ser- cally in the order of weeks or even longer). How- vice differentiation, resource sharing or co-exist- An earlier version of this ever, these approaches often exhibit operational ing heterogeneous platforms [6]. Instead, we article has been published inefficiencies due to frequent and significant traf- consider how multiple “equivalent” virtual net- in [1]. fic dynamics in operational networks. Take the work topologies, each having its own routing con- IEEE Communications Magazine • March 2012 0163-6804/12/$25.00 © 2012 IEEE 185
  2. 2. WANG2 LAYOUT_Layout 1 2/22/12 3:28 PM Page 186 For example, if the link between Kansas City Seattle and Houston is highly loaded, some traffic origi- nally carried through the green path (in VRT 1) {3, 6, 6} Chicago New York can be shifted to the other two (i.e. the blue and {1, 9, 1} {1, 1, 9} pink paths in VRTs 2 and 3, respectively) by Denver {1, 5, 1} adjusting the traffic splitting ratio across the Kansas City {1, 1, 5} Indianapolis three VRTs at Sunny Vale. The ultimate goal is Sunny Vale {1, 5, 9} to intelligently adjust traffic assignment through {1, 8, 7} {3, 7, 1} {1, 1, 4} splitting across multiple routing topologies at Washington {4, 1, 1} {1, 8, 1} individual source PoP nodes in reaction to the {1, 5, 9} monitored traffic conditions. In order to achieve LA this, the underlying MT-IGP link weights need {7, 9, 2} Atlanta {1, 8, 3} to be carefully computed offline and set for max- imizing path diversity, based on which adaptive Houston traffic control is performed. From a system point of view, AMPLE consists of two major components. The Offline Link IGP path in VRT1 IGP path in VRT2 IGP path in VRT3 Weight Optimization (OLWO) component focus- es on the static dimensioning of the underlying Figure 1. Providing path diversity in the Abilene network topology. network, with MT-IGP link weights computed for maximizing intra-domain path diversity across multiple VRTs. Once the optimized link weight figuration (such as IGP link weight setting), can configuration has been enforced onto the network, be used for multi-path enabled adaptive traffic the Adaptive Traffic Control (ATC) component engineering purposes in IP-based networks. performs short timescale traffic splitting ratio Multi-topology aware Interior gateway routing adjustment for adaptive load balancing across protocols (MT-IGPs) [7] are used as the underly- diverse IGP paths in the engineered VRTs, ing platform for supporting the coexistence of according to the up-to-date monitored traffic con- multiple virtual IGP paths between source-desti- ditions. Given the fact that traffic dynamics are nation (S-D) pairs on top of the physical network both frequent and substantial in today’s ISP net- infrastructure. works, our proposed TE system offers a promising In our proposal we introduce AMPLE (Adap- solution to cope with this in an efficient manner. tive Multi-toPoLogy traffic Engineering), a holis- tic system based on virtualized IGP routing topologies for dynamic traffic engineering. The SYSTEM OVERVIEW fundamental idea behind this scheme follows the Figure 2 presents an overall picture of the pro- strategy of offline provisioning of multiple diverse posed AMPLE TE system, with Offline MT-IGP paths in the routing plane and online spreading Link Weight Optimization (OLWO) and Adap- of the traffic load for dynamic load balancing in tive Traffic Control (ATC) constituting the key the forwarding plane, as advocated in [5]. The components. As previously mentioned, the ulti- approach can be briefly described as follows. mate objective of OLWO is to provision offline MT-IGPs are used as the underlying routing maximum intra-domain path diversity in the protocol for providing traffic-agnostic intra- routing plane, allowing the ATC component to domain path diversity between all source-desti- adjust at short timescale the traffic assignment nation pairs. With MT-IGP routing, customer across individual VRTs in the forwarding plane. traffic assigned to different virtual routing A salient novelty is that the optimization of the topologies (VRTs) follows distinct IGP paths MT-IGP link weights does not rely on the avail- according to the dedicated IGP link weight con- ability of the traffic matrix a priori, which plagues figurations within each VRT. existing offline TE solutions due to the typical Figure 1 depicts an illustration of how path inaccuracy of traffic matrix estimations. Instead, diversity can be achieved for S-D pairs in the our offline link weight optimization is only based Point-of-Presence (PoP) level Abilene network on the characteristics of the network itself, i.e. topology with three VRTs, by considering as an the physical topology. The computed MT-IGP example from Sunny Vale to Washington. The link weights are configured in individual routers, ith number in the bracket associated with each and the corresponding IGP paths within each link is the IGP weight assigned to it in the ith VRT are populated in their local routing infor- VRT. As illustrated in the figure, with each net- mation bases (MT-RIBs). While OLWO focuses work link assigned distinct IGP link weights in on static routing configuration in a long the three VRTs, completely non-overlapping timescale (e.g. weekly or monthly), the ATC paths can be provisioned between the S-D pair. component provides complementary functionali- As such, the key task of the offline configuration ty to enable short timescale (e.g. hourly) control is to compute MT-IGP link weights for providing in response to the behavior of traffic that cannot maximum path diversity for every S-D pair. be usually anticipated. Once these link weights have been configured in As shown in the figure, the input for ATC the network, an adaptive algorithm in the for- includes: warding plane performs traffic splitting ratio • The diverse MT-IGP paths according to the adjustment for load balancing across diverse IGP link weights computed by OLWO. paths in short timescale (e.g. hourly or even • Monitored network and traffic data such as more frequently) according to the monitored incoming traffic volume and link utiliza- network and traffic conditions. tions. 186 IEEE Communications Magazine • March 2012
  3. 3. WANG2 LAYOUT_Layout 1 2/22/12 3:28 PM Page 187 At each short-time interval, ATC computes a new traffic splitting ratio across individual VRTs Physical network Traffic topology dynamics for re-assigning traffic in an optimal way to the diverse IGP paths between each S-D pair. This functionality is handled by a centralized TE manager who has complete knowledge of the network topology and periodically gathers the up-to-date monitored traffic conditions of the Network Offline MT-IGP operating network. These new splitting ratios are link weight monitoring then configured by the TE manager to individual optimization source PoP nodes, who use this configuration for remarking the multi-topology identifiers (MT- Incoming Link IDs) of their locally originated traffic according- MT-IGP traffic volume utilizations link weights ly. The TE manager function can be realized as a dedicated server, but for robustness and resilience it can be implemented in a distributed MTR Adaptive replicated manner for avoiding the existence of a configuration traffic control single point of failure. In the next section we Static MT-IGP present the detailed design of individual compo- path sets nents in the AMPLE system. MT-ID MT-RIBs COMPONENT SPECIFICATION remarking OFFLINE LINK WEIGHT OPTIMIZATION Forwarding First of all, a fundamental issue in OLWO is decisions how to determine the definition of “path diversi- ty” between PoPs for traffic engineering. Let’s consider the following two scenarios of MT-IGP Figure 2. AMPLE system overview. link weight configuration. In the first case, highly diverse paths (e.g. end-to-end disjoint ones) are available for some PoP-level S-D pairs, while for across multiple VRTs (see [1] for details), and some other pairs individual paths are completely our evaluation based on two operational net- overlapping with each other across all VRTs. In works shows good path diversity performance: the second case, none of the S-D pairs have dis- only three VRTs are sufficient to avoid any criti- joint paths, but none of them are completely cal link for the GEANT network topology, while overlapping either. Obviously, in the first case if the Abilene topology needs four VRTs to any “critical” link that is shared by show later, all paths achieve the same goal. As we will becomes congested, its load cannot be alleviated even without necessarily creating high path through adjusting traffic splitting ratios at the diversity for every S-D pair, there is a high associated sources, as their traffic will inevitably chance of achieving near-optimal TE perfor- travel through this link no matter which VRT is mance based on the MT-IGP link weight setting used. Hence, our strategy targets the second sce- in OLWO. nario by achieving “balanced” path diversity across all S-D pairs. NETWORK MONITORING Toward this end, we define the binary metric Network monitoring is responsible for collecting of Full Degree of Involvement (FDoI) to evaluate up-to-date traffic conditions in real-time and the overall path diversity for a given MT-IGP link plays an important role for supporting the ATC weight configuration. More specifically, the FDoI operations. AMPLE adopts a hop-by-hop based value for a link with respect to an S-D pair is set monitoring mechanism that is similar to the pro- to 1 if this link is shared by the shortest IGP posal of [8]. The basic idea is that a dedicated paths across all VRTs for that S-D pair; otherwise monitoring agent deployed at every PoP node is it is set to 0. Let’s take Fig. 1 as an example again. responsible for monitoring: The FDoI value for the link from Sunny Vale to • The volume of the traffic originated by the LA with regard to the S-D pair (Seattle, LA) is 1, local customers toward other PoPs (intra- as this link is part of all the shortest IGP paths PoP traffic is ignored). between Seattle and LA across the three VRTs. • The utilization of the directly attached In comparison, the FDoI value for the same link inter-PoP links. with regard to the S-D pair (Sunny Vale, Wash- As shown in Fig. 3, this monitoring agent gathers ington) is 0, as alternate routes are available via data on the locally originated traffic volume Denver in other VRTs. The optimization objec- from all the access routers (ARs) attached to tive of OLWO is to minimize the sum of FDoI customers at the PoP. Meanwhile the agent also values across all network links with regard to all collects the utilization of the directly attached S-D pairs. If this sum is equal to 0, then no criti- inter-PoP links from individual backbone routers cal link is formed given the underlying MT-IGP (BRs). link weights, which means that at least one source In a periodic fashion (e.g. hourly), the central in the network will always be able to find alterna- TE manager polls individual monitoring agents tive path(s) to bypass the over-loaded link given within each PoP and collects their locally moni- any single link congestion scenario. tored traffic volume and link utilizations. These Our solution is based on an offline optimiza- statistics are then used by the central TE manag- tion algorithm for maximizing path diversity er for updating its maintained traffic engineering IEEE Communications Magazine • March 2012 187
  4. 4. WANG2 LAYOUT_Layout 1 2/22/12 3:28 PM Page 188 Given the optimized TE manager MT-IGP link weights TIB produced by OLWO, Monitoring agent adaptive traffic control (ATC) can be Inter-PoP link Polling Local traffic Reporting invoked at short-time volume utilization intervals during operation in order to re-optimize the utilization of network AR BR resources in reaction to traffic dynamics. To neighboring To local customers PoPs PoP node Intra-PoP structure Figure 3. Network monitoring and ATC. information base (TIB, to be introduced in the TIB, which consists of two inter-related reposito- next section) and computing traffic splitting ries, namely the Link List (LL) and the S-D Pair ratios for the next interval. Such a hop-by-hop List (SDPL). The LL maintains a list of entries based paradigm works efficiently in a TE system for individual network links. Each LL entry with a central manager. The main reason is that records the latest monitored utilization of a link new traffic splitting ratios are computed by the and the involvement of this link in the IGP paths TE manager who is able to have the global view between associated S-D pairs in individual VRTs. of the network, enabling it to achieve a global More specifically, for each VRT, if the IGP path optimum in traffic control. between an S-D pair includes this link, then the ID of this S-D pair is recorded in the LL entry. It ADAPTIVE TRAFFIC CONTROL is worth mentioning that this involvement infor- Given the optimized MT-IGP link weights pro- mation remains static after the MT-IGP link duced by OLWO, adaptive traffic control (ATC) weights have been configured (static information can be invoked at short-time intervals during is presented in black in Fig. 4, while dynamic operation in order to re-optimize the utilization information that needs to be updated periodically of network resources in reaction to traffic at short timescale is shown in red). On the other dynamics. The optimization objective of ATC is hand, the SDPL consists of a list of entries, each to minimize the maximum link utilization for a specific S-D pair with the most recently (MLU), which is defined as the highest utiliza- measured traffic volume from S to D. Each SDPL tion among all the links in the network. The entry also maintains a list of subentries for differ- rationale behind ATC is to perform periodic and ent VRTs, with each recording the splitting ratio incremental traffic splitting ratio re-adjustments of the traffic from S to D, as well as the ID of the across VRTs based on traffic pattern “continu- bottleneck link along the IGP path for that S-D ity” at short a timescale, but without necessarily pair in the corresponding topology. performing a global routing re-optimization pro- During each ATC interval, the TIB is updat- cess from scratch every time. In this section, we ed upon the occurrence of two events. First, present a lightweight but efficient algorithm that upon receiving the link utilization report from can be applied for adaptive adjustment of the the network monitoring component, the TE traffic splitting ratio at individual PoP source manager updates the link utilization entry in the nodes to achieve this goal. In a periodic fashion, LL and the ID of the bottleneck link for each the following two operations are performed: S-D pair under each VRT in SDPL. Second, • Measure the incoming traffic volume and when the adaptive traffic control phase is com- the network load for the current interval as pleted and the new traffic splitting ratios are described in the previous section. computed, the splitting ratio field in SDPL is • Compute new traffic splitting ratios at indi- updated accordingly for each S-D pair under vidual PoP source nodes based on the split- each VRT. ting ratio configuration in the previous ATC is performed based on the up-to-date interval, according to the newly measured data maintained in the TIB. We start the ATC traffic demand and the network load for algorithm description by defining the following dynamic load balancing. parameters: To fulfill the second task, a traffic engineering • t(u,v) — traffic from the source PoP node u information base (TIB) is needed by the TE man- to the destination PoP node v. ager to maintain necessary network state based • φu,v(r) — traffic splitting ratio of t(u,v) at u on which new traffic splitting ratios are computed. on routing topology r, 0.0 ≤ φu,v(r) ≤ 1.0. Figure 4 presents the structure of our proposed The algorithm consists of the following three 188 IEEE Communications Magazine • March 2012
  5. 5. WANG2 LAYOUT_Layout 1 2/22/12 3:28 PM Page 189 steps. We define an iteration counter y which is set initially to zero. Link ID Utilization (%) S-D pair ID Traffic volume t Step 1: Identify the most utilized link lmax in the network, which can be simply achieved by S-D pair ID Splitting Bottleneck visiting the updated LL in the TIB. VRT l ... VRT l ratio link ID Step 2: For the set of S-D pairs whose traffic S-D pair ID flows are routed through lmax in at least one but ... ... ... ... ... not all the routing topologies (i.e. FDoI = 0), S-D pair ID Splitting Bottleneck consider each one at a time and compute its new VRT P ... VRT P ratio link ID traffic splitting ratio among the VRTs until the S-D pair ID first feasible one is identified (see details in the (a) (b) follow-up description). A feasible traffic flow means that, with the new splitting ratios, the uti- Figure 4. Traffic engineering information base structure: a) entry structure of lization of lmax can be reduced without introduc- LL; and b) entry structure of SDPL. ing new hot spots with utilization higher than the original value. To support this operation, all fea- sible S-D pairs that meet the above requirement ⎛ ⎞ are identified from the entry of lmax in the LL. ⎜ 1 − μr ⎟ φu , v (r )′ = φu , v (r ) + ⎜ × δ ⎟ ∀r ∈ R − Step 3: If such a feasible traffic flow is found, ⎜ ∑− 1 − μr ⎟ accept the corresponding new splitting ratio ⎝ r ∈R ⎠ adjustment. Increment the counter y by one and go to Step 1 if the maximum K iterations have The lower (higher) the bottleneck link utiliza- not been reached (i.e. y ≤ K). If no feasible traf- tion, the higher (lower) the traffic splitting ratio fic flow exists or y = K, the algorithm stops and will be increased. the latest resulting values for the traffic splitting An important issue to be considered is the ratio are configured in the corresponding entry value setting for δ. If not appropriately set, it in the SPDL in order to be executed by individu- may either lead to slow convergence or over- al source PoP nodes. shoot the traffic splitting ratio, both of which are The parameter K controls the algorithm to undesirable. On one hand, too large value of δ repeat at most K iterations in order to avoid may miss the chance to obtain desirable splitting long running time. The value of K can be care- ratios due to the large gap between each trial. fully determined by taking into account the On the other hand, too small (i.e. too conserva- trade-off between the TE performance and sys- tive) value of δ may cause the algorithm to per- tem complexity. In Step 2, the task is to exam- form many iterations before the most ine the feasibility of reducing the load of the appropriate value of δ is found, thus causing most utilized link by decreasing the splitting slow convergence to the equilibrium. Taking this an algo- ratios of a traffic flow assigned to the routing consideration into account, we apply topologies that use this link, and shift a propor- rithm to perform an exponential increment of δ tion of the relevant traffic to alternative paths starting from a sufficiently small value. If this with lower utilization in other topologies. More adjustment is able to continuously reduce the specifically, the adjustment works as follows. utilization of lmax without introducing negative First, a deviation of the traffic splitting ratio, new splitting ratios on R+, the value of δ will be denoted by δ where 0.0 < δ ≤ 1.0, is taken out increased exponentially for the next trial until no for trial. For the traffic flow t(u,v) under consid- further improvement on the utilization can be eration, let R+ be the set of routing topologies made or the value of δ reaches 1.0 (i.e. the maxi- in which the IGP paths from u to v traverse lmax. mum traffic splitting ratio that can be applied). The main idea is to decrease the sum of traffic splitting ratios on all the routing topologies in WORKING AS A WHOLE SYSTEM R+ by δ and at the same time to increase the After presenting the detailed information on sum of the ratios on other topologies that do individual components, we now briefly describe not use lmax by δ. (We denote this set of topolo- how they work in unison as a whole TE system. gies by R – where R – = RR +.) Specifically, for First, optimized MT-IGP link weights are config- all the topologies in R+, which share a common ured on top of the underlying MT-IGP platform link with the same (maximum) utilization, their and remain static until the next offline OWLO traffic splitting ratios are evenly decreased. cycle. During this period, ATC plays the major Hence, the new traffic splitting ratio for each role for adaptively re-balancing the load accord- routing topology in R+ becomes: ing to the traffic dynamics in short-time intervals. As a bootstrap procedure, the initial traffic split- φu,v(r)’ = φu,v(r) – δ/�R+� ∀r ∈ R+ ting is evenly distributed across VRTs, but this will be recomputed based on follow-up traffic On the other hand, let μ r be the bottleneck monitoring results. In response to the periodic link utilization of the IGP path in routing topol- polling requests by the TE manager, the monitor- ogy r ∈ R–. To obtain μr, the TE manager should ing agents attached to individual PoP nodes first identify the ID of the bottleneck link along report back the incoming traffic volume (from the IGP path between the associated S-D pair access routers) and inter-PoP link utilizations from the SDPL, and then refer to the LL to (from backbone routers). The TE manager obtain its utilization. The traffic splitting ratio of accordingly updates the traffic volume between each routing topology in R – increases in an each S-D pair in the SDPL and link utilization inverse proportion to its current bottleneck link information stored in the LL of the TIB. Accord- utilization, i.e. ing to the obtained link utilization information, IEEE Communications Magazine • March 2012 189
  6. 6. WANG2 LAYOUT_Layout 1 2/22/12 3:28 PM Page 190 the bottleneck link ID along the IGP paths [10] provides traffic traces measured every five between individual S-D pairs in each VRT is also minutes, for consistency with the GEANT sce- updated in the SDPL. Based on the updated nario, we also use seven-day long traffic matrices information, the TE manager computes the new at the interval of every 15 minutes. In this article traffic splitting ratio for each S-D pair across we compare the following optimization methods: individual routing topologies. These new splitting Actual: The actual static link weight setting in ratios are configured in the SDPL and the TE the current operational networks. MT-IGP rout- manager, then instructs all the source PoP nodes ing is not used. within the network to use these new values for Multi-TM: We use the TOTEM toolbox to compute a set of static link weights for multiple traffic splitting during the next interval. In addi- tion, these values in the SDPL will also be used traffic matrices. The objective is to make the as the starting point for the future computation IGP TE robust to traffic demand uncertainty [3]. of the splitting ratios in the next interval. OnceSpecifically, the link weights are computed at the each source PoP node has received the new val- beginning of each day based on the sampled ues for traffic splitting from the central TE man-traffic matrices (one per hour) on the same day ager, it enforces them by remarking the MT-ID of the previous week. MT-IGP routing is not values carried by the locally originated traffic used. packets in the new proportions across individual AMPLE-n: Our proposed adaptive TE algo- routing topologies. Such a MT-ID remarking rithm that runs based on n MT-IGP routing operation follows the same style as the technique topologies with their link weights computed by the OLWO. The ATC operations are performed described in [9] for enabling IP fast reroute func- tions based on MT-IGP platforms. at 15 minute intervals according to the latest measured traffic conditions. Optimal: As the baseline for our compar- EXPERIMENTAL RESULTS isons, we use the GNU Linear Programming Kit In order to evaluate the performance of (GLPK) function in the TOTEM toolbox to AMPLE, we use the real topologies and traffic compute the optimal MLU for each distinct traf- traces from the GEANT and Abilene networks fic matrix associated with the given topologies. that are provided by the TOTEM Project [10]. Figure 5 plots the MLU versus the time inter- We present results based on a seven-day long vals of traffic traces for the GEANT and Abi- traffic traces dataset. Although the dataset in lene networks. From the figure we can have an overall glance of the traffic dynamics pattern in both networks during the sampling period. In a further evaluation, Table 1 shows results of the 1.2 Actual following additional statistics that are derived Multi-TM from Fig. 5: 1 maximum link utilization (AMU): AMPLE-3 Optimal • Average The average value of the MLU across all the 0.8 traffic traces during the seven-day period. • Highest maximum link utilization (HMU): MLU 0.6 The highest value of the MLU across all the traffic traces during the period. 0.4 • Proportion to near-optimal performance (PNO): The percentage over all the traffic 0.2 traces in which AMPLE can achieve near- optimal performance. We define here the 0 1 101 201 301 401 501 601 meaning of near-optimality to be the MLU Interval that is within 3 percent of the gap from (a) Optimal. An overall observation is that AMPLE can 1 substantially reduce the MLU for most of the traf- Actual fic traces. For example, in the GEANT network, Multi-TM AMPLE-3 the Actual link weight approach produces AMU 0.8 Optimal that is 86 percent higher than that of the optimal value, whereas with AMPLE the value varies 0.6 between 0.1 percent and 43 percent, depending on the number of routing topologies that are MLU 0.4 used. In general, the larger the number of rout- ing topologies used, the closer to the optimal performance can be achieved. Similar results are 0.2 also observed for the HMU performance. For the PNO metric in Table 1, if AMPLE is 0 based on two routing topologies, the value is only 1 101 201 301 401 501 601 13.1 percent but it still performs significantly bet- Interval ter than all the other approaches. We can now (b) start to see the practical usefulness of our approach for improving network utilization: Figure 5. MLU comparison between schemes with seven days of traffic traces: When the number of routing topologies increases a) the GEANT network MLU performance; and b) the Abilene network to three, the PNO boosts up to 78.3 percent. MLU performance. With 99.6 percent of all the traffic traces, 190 IEEE Communications Magazine • March 2012
  7. 7. WANG2 LAYOUT_Layout 1 2/22/12 3:28 PM Page 191 AMPLE achieves near-optimal performance with GEANT (%) Abilene (%) four routing topologies. These results reveal that, Optimization for the GEANT network, AMPLE has a very Method high chance of achieving near-optimal TE perfor- AMU HMU PNO AMU HMU PNO mance under any scenario of traffic traces with four routing topologies. For the Abilene network, Optimal 30.05 52.82 — 12.2 33.42 — our experimental results reach conclusions that Actual 55.74 96.91 0 19.59 63.24 1.19 are similar to that of the GEANT network. Multi-TM 48.56 104.15 0.44 53.2 230 0.15 Another observation is that the Multi-TM approach does not achieve good performance in AMPLE-2 42.9 94.61 13.08 18.61 60.96 64.14 minimizing the MLU according to Fig. 5. There AMPLE-3 31.95 60.36 78.34 12.36 33.44 88.69 are two reasons for this. First, the ultimate AMPLE-4 30.08 52.88 99.56 12.4 49.6 97.77 objective of Multi-TM is to minimize the cost Table 1. MLU performance statistics. represented by a piece-wise linear function [3] rather than specifically on the MLU. Second, even if multiple traffic matrices with different [3] B. Fortz and M. Thorup, “Optimizing OSPF/IS-IS Weights pattern characteristics are considered in link in a Changing World,” IEEE JSAC, vol. 20, no. 4, May 2002, pp. 756–67. weight optimization, unexpected traffic spikes [4] D. Xu, M. Chiang, and J. Rexford, “Link-State Routing may still introduce poor TE performance. This is With Hop-By-Hop Forwarding Can Achieve Optimal especially the case in the Abilene scenario whose Traffic Engineering,” Proc. IEEE INFOCOM, Apr. 2008. traffic pattern is less regular compared to the [5] M. Caesar et al., “Dynamic Route Computation Consid- ered Harmful,” ACM Comp. Commun. Rev. (CCR), vol. GEANT traffic pattern. 40, no. 2, Apr. 2010, pp. 66–71. [6] N. M. Mosharaf Kabir Chowdury and R. Boutaba, “A Survey of Network Virtualization,” Computer Networks, SUMMARY vol. 54, issue 5, Apr. 2010, pp. 862–76. [7] P. Psenak et al., “Multi-Topology (MT) Routing in In this article we have introduced AMPLE, a OSPF,” RFC 4915, June 2007. novel TE system based on virtualized IGP rout- [8] A. Asgari et al., “Scalable Monitoring Support for ing that enables short timescale traffic control Resource Management and Service Assurance,” IEEE against unexpected traffic dynamics using multi- Network Mag., vol. 18, issue 6, Nov. 2004, pp. 6–18. [9] A. Kvalbein et al., “Multiple Routing Configurations for topology IGP-based networks. The framework Fast IP Network Recovery,” IEEE/ACM Trans. Net., vol. encompasses two major components, namely, 17, no. 2, 2009, pp. 473–86. Offline Link Weight Optimization (OLWO) and [10] S. Balon et al., “Traffic Engineering and Operational Adaptive Traffic Control (ATC). The OLWO Network with the TOTEM Toolbox,” IEEE Trans. Net- work and Service Management, vol. 4, no. 1, June component takes the physical network topology 2007, pp. 51–61, Project website (including GEANT/Abi- as the input and aims to produce maximum IGP lene Network Topology and Traffic Dataset): path diversity across multiple routing topologies through the optimized setting of MT-IGP link weights. Based on these diverse paths, the ATC component performs intelligent traffic splitting BIOGRAPHIES adjustments across individual routing topologies NING WANG ( is a Lecturer at the Cen- in reaction to the monitored network dynamics tre for Communication Systems Research (CCSR), University of Surrey in UK. He received his B.Eng. (Honors) degree at short timescale. As far as implementation is from the Changchun University of Science and Technology, concerned, a dedicated traffic engineering man- P.R. China in 1996, his M.Eng. degree from Nanyang Uni- ager is required, having a global view of the versity, Singapore in 2000, and his PhD degree from the entire network conditions and being responsible University of Surrey in 2004 respectively. His research inter- ests include Internet traffic engineering, network virtualiza- for computing optimized traffic splitting ratios tion, QoS provisioning and content-aware networks. according to its maintained TE information base. Our experiments based on the GEANT KIN HON KO ( was an Instructor and Abilene networks and their real traffic at the Department of Computing, The Hong Kong Poly- technic University (PolyU). He holds a B.Sc. (Hons) in Com- traces have shown that AMPLE has a high puter Studies from City University of Hong Kong, a M.Sc. chance of achieving near-optimal network per- (Eng.) in Data Communications from the University of formance with only a small number of routing Sheffield, and a Ph.D. from University of Surrey. Before topologies, although this is yet to be further veri- joining PolyU, he was a postdoctoral research fellow at the CCSR, University of Surrey, and a full-time teaching profes- fied with traffic traces data from other opera- sional at City University of Hong Kong. His current research tional networks when available. A potential interests cover network operations and management, traf- direction in our future work is to consider a fic engineering, dependable and secure network design, holistic TE paradigm based on AMPLE, which is P2P/IPTV traffic management and Green Internet design. able to simultaneously tackle both traffic and GEORGE PAVLOU ( holds the Chair of network dynamics, for instance network failures. Communication Networks at the Dept. of Electronic & Elec- trical Engineering, UCL. Over the last 25 years he has REFERENCES undertaken and directed research in networking, network management and service engineering, having extensively [1] N. Wang, K-H. Ho, and G. Pavlou, “Adaptive Multi- published in these areas. He has contributed to ISO, ITU-T topology IGP Based Traffic Engineering with Near-Opti- and IETF standardization activities and has been instrumen- mal Performance,” Proc. IFIP Networking 2008. tal in a number of key European and UK projects that pro- [2] S. Uhlig et al., “Providing Public Intradomain Traffic duced significant results. His current research interests Matrices to the Research Community,” ACM Sigcomm include traffic engineering, content-based systems, auto- Comp. Commun. Rev. (CCR), vol. 36, no. 1, Jan. 2006, nomic networking, policy-based management and infra- pp. 83–86. structure-less wireless networks. IEEE Communications Magazine • March 2012 191