Quality of Service of Internet Service Provider Networks: State of Art and New Trends Armand Toguyéni, Ouajdi Korbaa LAGIS – CNRS UMR 8146, Ecole centrale de Lille BP 48, 59651 Villeneuve d’ASCQ, France Tel: +33-3-20-33-54-49, Fax: +33-3-20-33-54-18, e-mail:
[email protected] * Ouajdi Korbaa, ISITC, University of Sousse, Tunisia e-mail:
[email protected]
ABSTRACT These ten last years, the emergences of new applications through Internet such as multimedia applications (VoIP, video broadcasting …) or networked control application, users need more and more quality of service (QoS) when using Internet. However, the requested QoS is not the same depending on the application. Some of them request a lot of bandwidth like video broadcasting others need short guaranteed delays … Several new models to manage internet traffic have been proposed. Most of them focus on specific QoS criteria that they want to optimize. This paper focus on approaches based on MPLS that are more adapted to the configuration of ISP networks. The paper first introduces the different mechanisms that can be used to improve the QoS of a network such as the models of traffic classification (IntServ or DiffServ) or multipath routing. It also justifies the choice of MPLS technology versus a simple use of IP. After that, the paper presents main MPLS approaches such as MATE, LDM or LBWDP that try to guaranty a certain quality with respect to specific criteria. In the last part, it presents new models such as PEMS that try to adapt the offered quality depending on the class of the routed traffic. Keywords: traffic engineering, multipath routing, quality of service, differentiated service, QoS routing 1. INTRODUCTION Today a lot of new applications are developed through Internet. Most of them need a guaranty bandwidth or delay. It is typically the case of multimedia applications such as video on demand or applications such voice on IP. But, Internet is also more and more use to control real time industrial system such as power plants or car production chains. All these applications need to guaranty some features of the network with regard to the quality of flows’ transmissions but with different criteria: it is what is called quality of service (QoS) of the network. This paper concerns the way we can achieve QoS in an ISP (Internet Service Provider) network. ISP networks are essential for QoS because they assume the transit of flows at the network core. The problem is that very often the ISP must increase the capacity of its network resources because of the increase of users’ flows. The ISP also notices that some parts of their networks are often congested while other parts are less. The idea developed is to propose load balancing approaches to allow better performances of ISP networks. The paper will be dividing as it follows. The second section will present the main concepts to understand QoS in Internet. The third section will concern QoS models to classify flows inside routers. The fourth section is about Traffic Engineering based on MPLS. Finally, the last section will present PEMS (Periodic multi-step algorithm) a new algorithm to ingrate Diffserv and Traffic Engineering. 2. QUALITY OF SERVICE IN INTERNET Classical dynamic routing protocols in the current Internet always forward packets to the shortest path. Shortest path routing leads to imbalanced traffic distribution. This can cause congestions in some parts of the network even if traffic load is not particularly heavy. This can cause problems for flows with a need for QoS (Quality of Service) guarantees if the shortest path does not have enough resources to meet the requirements. 2.1 QoS metrics The growth of multimedia applications over wide area networks has increased research interest in QoS. The communication delay and synchronization needed for voice, data and images are major concerns. Internet telephony (Voice over IP) and other multimedia applications such as video conferencing, video-on-demand and media streaming require service guarantees and have strict timing requirements. The size and quality of display devices, and resources such as CPU (Central Processing Unit), battery power and bandwidth are always limited. QoS can be parameterized as throughput, delay, delay variation (jitter), loss and error rates, security guarantees and so on, that are acceptable in an application. As such, QoS depends on characteristics of applications. For instance, the variation in delay, the difference between the largest and the smallest delay, is called delay jitter
and jitter is an important quality for IP (Internet Protocol) telephony, which can tolerate a certain percentage of packet loss without any degradation of quality. For data transfer, loss is a crucial QoS parameter. QoS control requires an understanding of the quantitative parameters at the application, system and network layers. 2.2 QoS techniques Two main techniques are developed to implement QoS in IP networks: Constraint-Based Routing and Traffic Engineering. Constraint-Based Routing (CBR) is a routing scheme that considers the QoS requirements of a flow when making routing decisions. As opposed to traditional shortest path routing which only considers the hop-count, CBR is designed to find feasible paths that satisfy multiple constraints, for example bandwidth, QoS requirements such as delay and jitter (QoS-based routing), or other policy requirements on the paths (Policybased routing). While determining a route, CBR considers not only network topology but also requirements of the flow, resource availability of the links, and possibly other policies specified by the network administrators. Therefore, CBR may find a longer but lightly loaded path better than the heavily loaded shortest path. Network traffic is thus distributed more evenly. Traffic Engineering (TE) is a solution to install and check the flows of traffic inside the network, taking full advantage of all the potentialities offered by the available resources and avoiding uneven network utilization. TE is needed in the Internet mainly because current IGPs (Interior Gateway Protocols) always use the shortest paths to forward traffic. By performing TE in their networks, ISPs can greatly optimize resource utilization and network performance. The common optimization objectives include: • Minimizing congestion and packet losses in the network, • Improving link utilization, • Minimizing the total delay experienced by packets, • Increasing the number of customers with the current assets.
3. SERVICE MODELS BASED ON FLOWS’ CLASSIFICATION The architectures and mechanisms developed for enabling QoS have two key QoS issues: resource allocation and performance optimization. For resource allocation in the internet, Integrated Service (IntServ) [1] and Differentiated Service (DiffServ) [2] are proposed by IETF. But Intserv has been proved to be not scalable because it requires maintaining per-microflow state and signaling at every router. Intserv has been replaced by Diffserv which is less complicated and more scalable. The main goal of DiffServ is to provide a scalable framework for offering a range of services in the Internet with Quality of Service support and without the need to maintain per-flow state in every router. DiffServ has only a limited number of service classes indicated by the DS field. Since resources are allocated in the granularity of class, the amount of state information is proportional to the number of classes rather than the number of flows. So DiffServ is more scalable than IntServ. In a DiffServ network, the boundary nodes (or edge nodes) and interior nodes (or core nodes) have different tasks. DiffServ achieves scalability through performing complex QoS functions such as classification, marking, and conditioning operations using the DiffServ Code Point (DSCP) into a limited number of traffic aggregates or classes only at the edge nodes. In the core routers, scheduling and queuing control mechanisms are applied to the traffic classes based on the DS field marking: all traffic conditioning and dropping is intelligently handled at the network layer using IP DiffServ QoS mechanisms. DiffServ has three kinds of services. • EF (Expedited Forwarding) : premium service with reliable, low delay and low jitter delivery, • AF (Assured Forwarding) : assured service with reliable and timely delivery, • Default (Best-effort): it is the normal service offered by IP networks. DiffServ is a scalable solution that does not require per-flow signalling and state maintenance in the core. But it also has some limitation notably regardless resource use. This model suggests only mechanisms for relative packet forwarding treatment to aggregate flows, traffic management and conditioning. However it does not provide architecture for end-to-end QoS. Furthermore, there is no traffic engineering provision in DiffServ. As a result some links in the domain might experience congestion while other links go unutilized.
4. TRAFFIC ENGINEERING IN A MPLS NETWORK Schematically an ISP network can be considered as composed of several pairs of ingress-egress routers and core routers that enable to transport the traffic between the pairs of peripheral routers. To achieve this function, this requires some kinds of connections in the connectionless IP networks. MPLS (Multi-Protocol Label Switching) supports efficiently, the explicit route setup and can give this capacity to establish connection in IP network. Routing control in traffic engineering essentially balances traffic load over multiple paths in the network to minimize congestion. It concerns how to select paths and to distribute traffic among those paths such that given QoS constraints are met or close to the target at a large possible extent while at the same time optimizing some global network-wide objectives such as utilization. To allow comparing the different approaches, [3] proposes a generic framework of multipath routing based on two main steps: computation of multiple paths and traffic splitting among these multiple paths. A multipath algorithm uses first a Candidate Paths Selection algorithm to calculate the paths that can be used for load balancing. There is a subset of all the paths between a pair of ingress-egress routers at the borderline of the ISP network. These candidate paths are defined with respect to a cost function. According to the nature of the cost function, different algorithms can be applied to determine these candidate paths. To specify the cost function, the authors consider various criteria such as the bandwidth, hop count, delay, error rate, and so on. Generally this step uses static criteria. So, the first problem in this step can be summarized by finding the appropriate cost function or metric that can take some of the previous criteria into account. To perform a global optimization of network QoS is a multi-criteria problem. This multi-criteria problem is an NP-complete problem [4] and consequently requires heuristics to be solved. Hierarchical criteria approaches such as SWP (Shortest Widest Path) [5] or WSP (Widest Shortest Path) [6] have been proposed for this problem. But this type of algorithm can perform individual optimization of flows’ requirements but not a global optimization. The second step of multipath routing algorithms consists in splitting traffic among multiple candidate paths. The paths for this step are qualified of candidate paths because they are not necessary all used at each time. Traffic splitting ratio are decided depending on the evaluation of dynamic criteria such as blockages, packet loss rate, measured delay, jitter and so on. In this case, the problem remains the same as the first step: how to define a cost function or metric that combines different previous criteria. Several models are proposed in the literature to perform Traffic Engineering. In this paper, we consider models such as MATE, LDM and LBWDP. Before presenting theses models, we present traffic bifurcation model that is a mathematical formulation of route optimization problem. It is a reference to compare the different propositions. 4.1 Traffic Bifurcation The main objective of Traffic Engineering can be reformulated as obtaining low congestion and optimizing the utilization of network resources. Practically, this means minimizing the utilization of the most heavily used link in the network. Link utilization (Ul) is calculated as follows with cl the link capacity and rl its residual bandwidth. c −r (1) U (l ) = l l with 0 ≤ rl ≤ cl cl When translating this into a mathematical formulation, the objective is in essence to minimize the maximum link utilization in a network. Intuitively the hot spots are the points with the highest link utilization. Reducing the link utilization at these points balances the traffic distribution across the network. The traffic bifurcation problem (TB hereafter) consists of finding multiple paths carrying each part of or all the traffic between ingress and egress node which minimize the maximum link utilization α. When splitting a traffic demand to multiple paths, the granularity of load splitting, g (0 n The first condition implies that LDM utilizes the LSPs with more extra hops if they have lower utilization than the LSP that has the highest utilization among the LSPs in the current Aij. Moreover, since the second condition keeps the links with utilization of ηm or higher from being included in the LSPs with the path length (h(shortest LSP) + m) or longer, and ηm < ηn for m > n, LSPs with more extra hops are applied with more strict utilization restriction when they are considered for the candidate LSP set eligibility. The second condition actually implies a multi-level trunk reservation. That is, links with the utilization higher than ηm can only be used by the LSPs with less than m extra hops. Note that with the candidate LSP set expansion condition i.e., U(Aij) ≥ ρ and the first of the two eligibility conditions given above, a longer LSP is added to the candidate LSP set if the utilization of the current candidate LSP set is higher than ρ and the utilization of this longer LSP is relatively lower than the LSP with the maximum utilization among the LSPs in the current candidate paths set. With these two conditions, longer detour using an LSP with more extra hops is allowed even when that longer LSP’s absolute level of congestion is already significant. The second eligibility condition enables to prevent this. The candidate path set could either be pre-computed when there are some significant changes in the dynamic network status or be computed on demand for a new arriving user flow request. This is done in a O(n²) time in the worst case, the number of iterations growing with the utilization of the network. Here n refers to the number of paths available for the considered ingress-egress pair of nodes. In terms of convergence, the number of iterations has an upper limit defined by a given parameter δ, so the number of iterations is bounded. The traffic splitting is done using a heuristic to determine a repartition policy for incoming flows at an ingress edge router. For each incoming traffic flow, LDM randomly selects an LSP from the candidate LSP set according to the probability distribution that is a function of both the length and the utilization of the LSP. Let Pl denote the probability that LSP l is selected for sending current demand, and let us define several additional notations to explain how to compute Pl. For a candidate LSP set A, let A = {l1, l2, . . . lNa}, where Na is the number of LSPs in A. C0 is the constant to make the sum of the probabilities that are inversely proportionate to the hop count of an LSP. That is,
C0 =
1 Na
∑ i =1
(13)
1 h (l i )
Let E = max[u(sp), ∀sp ∈ shortest LSPs in A] and d(li) = E - u(li), for 1≤ i ≤ Na. Then C1 is the variable to make the sum of the probabilities that are proportionate to d(li). That is,
C1
=
Na
∑ d (l i =1
i
)
(14)
The a0 and a1 factors are to be defined by the administrator to fit its needs. Pl is defined as the following, P (l ) = a 0
C0 d (l ) + a1 h (l ) C1
with
a 0 + a1 = 1 (15)
Once probabilities are defined, each incoming flow is directed to its route selected randomly. The complexity of the whole splitting procedure is clearly O(n). Here n refers to the number of paths selected at the end of the previous step. Instability can affect LDM because of oscillations due to candidate path selection. This oscillation problem can be solved using two thresholds [1].
4.4 Load Balancing over Widest Disjoints Paths algorithm (LBWDP) It is a hybrid algorithm that uses the selection path algorithm proposed by WDP (Widest Disjoint Paths algorithm) [9] and a splitting algorithm called PER (Prediction of Effective Repartition) [10]. PER is an improvement of LDM splitting algorithm. Its first stage performs a calculating of a distribution probability based on equation (16).
ri = p0
H b(i) with p +p =1 + p1 0 1 h(i) B
(16)
with B the sum of the residual bandwidth of selected candidate paths and H is equivalent to the C0 of equation (13). The second stage of PER selects one LSP using a gradient method based on constraint bandwidth traffic. This stage considers ei is the effective repartition ratio on LSPi and Si the probability of selection for each candidate path. ni
ei =
∑d j =1
ji
where
n
∑d j =1
k
ni
i =1 j =1
(17)
n
∑∑ d =∑d ji
j =1
j
j
with dji is the traffic amount of the ji th demand assigned on LSPi. Si =
ri − e i ri
(18)
Figure 1 summarizes how PER works.
Figure 1. Flowchart of PER algorithm 4.5 Traffic Scenario for multi-users is as follows. In order to make an objective comparison, the previous algorithm have been simulates with ns2 based on the same network architecture (Figure 2) and same traffic scenarios. The features of the simulation are the following: • The volume of each individual demand is 300kb per second fixed. • The source and destination pairs are Src0-Dst0, Src1-Dst1 and Src2-Dst2. These pairs are selected randomly among them every demand generation interval. • One flow generated in certain time is stopped in a random time. • We adapt a time based triggering as a triggering policy and update link state every 3 seconds.
•
The delay of each simulation is 150 seconds.
Figure 2. Simulation topology In this context, the best results are obtained with LBWDP which is very close of the ideal distribution illustrated by TB (Figure 3).
rate rate
Maximum MaximumLink LinkUtilization Utilization
TB TB MATE MATE LDM LDM LBWDP LBWDP
2 2 6 6 10 10 1 14 4 1 18 8 2 22 2 2 26 6 3 30 0 3 34 4 3 38 8 4 42 2 4 46 6 5 50 0 5 54 4 5 58 8 6 62 2 6 66 6 7 70 0 7 74 4 7 78 8 8 82 2 8 86 6 9 90 0 9 94 4 9 98 8
s se ec c
100 100 90 90 80 80 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0
sec sec
Figure 3. Simulations results 5. PERIODIC MULTI-STEP ROUTING ALGORITHM FOR DS-TE (PEMS) [11] DiffServ aware MPLS Traffic Engineering (DS-TE) mechanisms operate on the basis of different DiffServ classes of traffic to improve network performance and extend the base capabilities of TE to allow route computation and admission control to be performed separately for different classes of service. With DS-TE, it is possible to define explicit routes with different performance guarantees on each route to ensure QoS constraints are met. It can also give network designers the flexibility to provide differential treatment to certain QoS classes that need path-protection and the strict QoS guarantees while optimizing the network resources utilization. PEriodic Multi-Step Routing algorithm (PEMS) is a DS-TE algorithm. It is composed of three stages as illustrated by (Figure 4).
Figure 4. Three stages of PEMS In the pre-processing phase, it extracts good paths of all possible paths which can include every link at least once within them for each source-destination pairs using only topology information, in the offline mode. These paths are kept until the topology is changed.
When a traffic demand arrives, it uses PER algorithm to select one LSP to carry current flow. Many QoS metrics such as hop count, available bandwidth and delay constraints are considered before the path selection to assign. In PEMS, hop-count and disjointedness are used in the pre-processing phase and available bandwidth, measured delay are used in the cost function to establish splitting ratios. PEMS aims to minimize the maximum link utilization like LBWDP algorithm basically, and additionally to give different service quality to each class, especially to guarantee the low delay to EF class. Figure 5 gives PEMS flowchart to summarize how it works. The meaning of its notations is as follows: − dei : delay of LSPi − bi : residual bandwidth of LSPi − CPEF, CPAF, CPBE : candidate path set for EF class, AF class and BE class respectively − dkcc : k-th demand with class cc − CPcc: current class (one in CPEF, CPAF or CPBE) − CPccpotential : subset of CPcc corresponding to LSPi that can process the requested demand dkcc
Figure 5. PEMS flowchart In the online mode, when link state information are updated, new candidate paths for each class are calculated based on updated information such as measured delay and residual bandwidth. At this point, we use metric ordering by delay and residual bandwidth. This phase selects multiple low-delayed paths in the ordered paths set as candidate paths of delay-sensitive traffic and selects multiple paths having more residual capacity for the traffic to which the bandwidth is important for multipath routing to each traffic class. Simulations results [11] have proved that algorithm like PEMS that integrates Diffserv model are more efficient for premium and assured traffics but less efficient for best effort traffics.
6. CONCLUSIONS MPLS offers many advantages to service providers. In order to support today’s various kinds of applications, the system needs to guarantee the Quality of service. However, MPLS is incapable of providing differentiated service levels in a single flow. Hence MPLS and DiffServ seem to be a perfect match and if they can be combined in such a way to utilize each technology’s strong points and counter the others’ weaknesses, it can lead to a symbiotic association that can make the goal of end to end QoS feasible. DiffServ-aware Traffic Engineering mechanisms operate on the basis of different Diffserv classes of traffic to improve network performance and extend the base capabilities of TE to allow route computation and admission control to be performed separately for different classes of service. Algorithms like PEMS seem to be a good compromise between improvement of resource utilization and the QoS required by end users. REFERENCES [1] R. Braden, D. Clark, S. Shenker: Integrated Service in the Internet Architecture: an Overview, RFC1633, Jun. 1994, IETF [2] D. Black, M. Carlson, E. Davies, Z. Wang, W. Weiss: An Architecture for Differentiated Service, RFC2475, Dec 1998, IETF
[3]
K. Lee, A. Toguyeni, A. Rahmani: Comparison of multipath algorithms for load balancing in a MPLS Network, in Proc. ICOIN2005, Seoul, South Korea, January 2005. [4] C. Shigang: Routing Support for Providing Guaranteed End-to-End Quality-of-Service", Ph.D. thesis, UIUC, 207 pages, 1999. [5] Z. Wang, J. Crowcroft: QoS Routing for Supporting Multimedia Applications, in IEEE Journal of Selected Areas in Communications, n° 14, pp. 1228-1234, 1996. [6] R. Guerin, A. Orda., D. Williams: Qos Routing Mechanisms and OSPF Extensions, In Proc. of the Global Internet Miniconference, Phoenix, USA, 1997. [7] A. Elwalid, C. Jin, S. Low, and I. Widjaja: MATE: MPLS Adaptive Traffic Engineering, in Proc. INFOCOM’2001, Alaska, USA, pp. 1300-1309, 2001. [8] J. Song, S. Kim, M. Lee: Dynamic Load Distribution in MPLS Networks, in LNCS Vol. 2662, pp. 989-999, 2003. [9] N. Srihari, Z. Zhi-Li: On Selection of Paths for Multipath Routing, In Proc. IWQoS'01, Karlsruhe, Germany, 2001. [10] K. Lee, A. Toguyéni, A. Rahmani: Hybrid multipath routing algorithms for Load balancing in MPLS based IP network, in Proc. AINA’2006, Vienna, Austria , pp. 165-170, April 2006 [11] K. Lee, A. Toguyéni, A. Rahmani: Periodic Multi-Step routing algorithm for DS-TE: PEMS, International e-Conference on Computer Science 2006 (IeCCS’2006), July 2006