The Application of Bandwidth Optimization Technique in SLA Negotiation Process

Transcription

1 The Application of Bandwidth Optiization Technique in SLA egotiation Process Srecko Krile University of Dubrovnik Departent of Electrical Engineering and Coputing Cira Carica 4, Dubrovnik, Croatia Tel , Fax , e-ail Djuro Kuzuilovic University of Dubrovnik Departent of Electrical Engineering and Coputing Cira Carica 4, Dubrovnik, Croatia Tel , Fax , e-ail Abstract - In this paper we propose an efficient resource anageent technique for MPLS-based DiffServ networks The service provider (SP) can predict sufficient bandwidth resources on the link for traffic caused by contracting of SLAs (Service Level Agreeent) It could be an iportant part of adission control in the SLA negotiation process (SLA-creation) Decisions of optial capacity arrangeents fro network provider (P) have to be done interactively and dynaically It can significantly iprove QoS end-to-end routing in the oent of service invocation The proble of bandwidth optiization can be seen as an expansion proble of link capacities Such explicit traffic engineering technique provides the possibility to intelligently tailor the route foe each SLA traffic flow such that different parts of the network reain equally loaded can be done dynaically but such resource anageent approach has to be the result of all previous contracted SLAs, that are participating with definite bandwidth in the sae tie period Such bandwidth optiization technique helps to avoid the creation of bottleneck links on the path and aintains high network resource utilization efficiency The iportant difference fro existing traffic routing techniques is that optial path for SLA need not to be the shortest path solution as we have in eg OSPF protocol (often used for traffic routing in the oent of service invocation) Index Ters - dynaic bandwidth, adission control in DiffServ/MPLS networks, SLA creation, end-to-end QoS routing, traffic routing of aggregate flows I ITRODUCTIO In DiffServ networks the classification of the aggregated flows is perfored according to the SLA (Service Level Agreeent) signed between a custoer and the service provider (SP) operator Each SLA contract specifies a tie period for service utilization of defined service class and appropriate bandwidth (speed how traffic ay be sent) In general, any eleents define service class or QoS level To obtain quantitative end-to-end guarantees in DiffServ/MPLS architecture, based on traffic handling echaniss with aggregate flows, soe kind of bandwidth optiization technique (on the link) has to be introduced Multi-protocol label switching (MPLS) has gained popularity as a technology for anaging network resources and providing of perforance guarantees In this paper we are looking for optial path provisioning for traffic flows caused by all existing SLAs participating in the sae tie period (forer contracted SLAs); see fig and fig 2 The ain condition is: the sufficient network resources ust be available at any oent of that period In the worse case it ust be sufficient for the first service class (full satisfied) Any virtual network (V) is a subset of physical network resources purchased by an SP fro the IP transport network provider (P) to build a logical service delivery infrastructure It can be ade dynaically, using soe TE technique If we can predict sufficient link resources (without shortages) in the oent of SLA creation, the possibility of link congestions in V will be significantly reduced in the oent of service invocation The probles of dynaic bandwidth optiization technique are investigating Such proble can be seen as the link capacity expansion proble (CEP) fro the coon source with expansion values in allowed liits Explanation of the atheatical odel for link bandwidth QoS service clasess guaranteed (fully satisfied) adaptive (alost satisfied) best-effort SLA_3 SLA_4 SLA_ SLA_5 critical period SLA_2 tie Figure An exaple of nuber of SLAs that participate in the sae period of tie; see figure 2 The bandwidth optiization and the congestion control has to be done for that critical period (duration of new SLA)

2 SLA_3 ingress SLA_6 egress Edge SLA_ 3 ingress SLA_ ingress SLA_2 ingress SLA_4 egress s SLA_ egress SLA_2 egress SLA_4 ingress SLA_5 egress SLA_6 ingress d SLA_ 4 SLA_ 6 Edge SLA_ 3 5 Edge Sla core network SLA_ 5 SLA_ egress SLA_3 egress SLA_5 ingress Edge Figure 2 An exaple of nuber of SLAs in context of new SLA creation In service specifications (details about service activation and duration) they share bandwidth on the link in the sae period of tie; see fig expansion technique is given in section II Heuristic algorith is described in section III uerical exaples, testing results and algorith application are discussed in section IV The service provider (eg ISP) wants to accept new SLA (traffic flow) that generates the traffic flow between edge s In DiffServ/MPLS architecture we call the LER (Label Edge Router) Traffic deand on exit of LER represents the su of all ingress and egress flows Interior s in core network, capable to forward traffic in equivalent classes (FEC), are called LSR (Label Switching Routers) The network exaple that is shown in figure can be a good representation of such architecture Such aggregated flow is coing to LSR and has to be routed to destination (egress ) Packets of the sae FEC are assigned the sae label and generally traverse trough the sae path across the MPLS network A FEC ay consist of packets that have coon ingress and egress nodes, or the sae service class and sae ingress/egress nodes or any other cobination In this case, the FEC of appropriate service class aggregates traffic deands (new SLA and forer contracted SLAs) of the sae QoS level A path traversed by an FEC is called a label switching path (LSP) In that sense network operator has to find the optial path for the FEC without any congestion in the network Possibility of congestion on soe link (with liited bandwidth) exists, specially for definite period of tie; see fig For each counication link in the DiffServ network given traffic deands can be satisfied on any (usually three) different QoS levels (with appropriate bandwidth) At the SLA level, the transport network resources are shared by set of per-class, per ingress/egress-pair SLAs Traffic can be satisfied on appropriate QoS level or higher, but not on lower QoS level In fact, different LSP is used for different service class But SP doesn t want that transport network be divided into ultiple virtual networks, one per service class The end effect is that SP can give to preiu traffic ore resources, but exactly that is necessary, no less no ore In the sae tie SP strongly needs the optial utilization of liited capacity resources with inial bandwidth resources arrangeent fro P The optial resource anageent proble can be seen as the link capacity expansion proble (CEP) in allowed liits; see [4] and [6] With congestion control/expansion algorith we can predict link resources to satisfy traffic deands caused by all accepted SLAs If the optial routing sequence has any link expansion with value that exceeds allowed liits, it eans that link capacity on the path cannot be sufficient for such traffic It eans that new SLA cannot be accepted and ust be redefined through negotiation process For exaple, the custoer can decide to take the adaptive QoS service class instead of fully satisfied (guaranteed) service class Soe iportant papers about that proble are [], [2] and [3] In paper [] such bandwidth optiization technique is a part of TEQUILA architecture based on autonoic coputing In the paper [4] such algorith is a part of service anageent architecture In papers [8] and [9] siilar network engineering technique is related on path provisioning

3 II THE MATHEMATICAL MODEL The proble explained above can be solved as the capacity expansion proble (CEP) without shortages Partially expansions for each link are ade fro coon source in given liits Transission link capacities (bandwidth) on the path between s are capable to serve traffic deands for different QoS levels (service class) for i =,2,, Fig 3 gives an exaple of network flow representation for ultiple QoS levels () and M internal (core) s included in the path Let G (A, E) denote a network topology, where A is the set of nodes and E the set of links The source and destination nodes (eg edge s in IP-doain) are denoted by s and d respectively Between the there are M interior s on the path; see fig 3 For M interior s on the path we have: M links + 2 = M + links { =,, M+} In the atheatical odel the following notation is used: j and k = QoS level We differentiate n service classes (QoS levels)the levels are ranked fro i =,2,,, fro higher to lover = the order nuber of the link on the path, connecting two successive LSR s, =,, M+ u,v = the order nuber of capacity points in the subproble, u,, v M+ r = traffic deand increent for additional capacity I = the relative aount of idle capacity I i = 0, I M+ = 0 x = the aount of used capacity for facility L = upper bound for link capacity y = the aount of capacity for facility i on the link, redirected to satisfy the traffic of lower level j Consider a network G (A, E) where each link is characterized by z-diensional link weight vector, consisting of z-nonnegative QoS weights The nuber of QoS easures (eg bandwidth, delay) is denoted by z Weights are liited with appropriate constraints for each easure QoS easures can be roughly classified into additive (eg delay) and non-additive (eg available bandwidth) In case of an additive easure, the QoS value of the path is equal to the su of the corresponding weights of the link along that path For a non-additive easure, as we have in our proble, the QoS value of the path is the iniu (or axiu) link weight along the path In this atheatical odel it is assued that the network state inforation (eg traffic deands, link capacities, delay liits) are teporary static In general we have ulti-constrained proble (MCP) but in this paper we talk about one-diensional link weight vectors for M+ links on the path {w, E, i =,, } with only one constraint denoted with L Definition of the single-constrained proble is to find a path P fro node s to node d such that: M ( +, i, = i = w P ) = in w i ( I ) () where: I L (2) for i =,, ; =,, M+ A path obeying the above condition is said to be feasible ote that there ay be ultiple feasible paths between s and d Capacity situation I of each link depends of expansion solutions (x, y, ) Bandwidth constraints for link capacity values on the link and for appropriate service class are denoted with L (L, L 2,, L, ) Such proble is special case of the MCP O R, X, X,2 X,M I, =0, I,2,2 I,3 I,M,M I,M + = 0 r, r,2 r,m QoS le ve l I 2, =0 I,=0 y 2, 2, y 2, I 2,2 y 2,2 2,2 y 2,2 I,2,,2 I 2,3 r 2, y, r 2,2 y,2 r 2,M I,3 I 2,M I,M y 2,M 2,M y 2,M,M y,m I 2,M + = 0 I,M += 0 edge r, in te rio r r,2 r, M lin k lin k 2 lin k M + in te rio r M edge Figure 3 A network flow representation of the CEP odel applied for congestion control in SLA creation process Service classes are differentiated but possible traffic conversion is possible only in direction toward higher QoS level

4 proble, known as Restricted Shortest Path (RSP) proble We can say that each weight w depends of capacity situation on the link between two s (capacity expansion or conversion, inventory or shortage etc) That weight usually corresponds to the link expansion cost It eans that the link weight (cost) is the function of used capacity: lower aount of used capacity (saller bandwidth with delay in acceptable liits) gives lower weight (cost) The ain condition is that given traffic deands ust be fully satisfied without shortages onlinear cost function is necessary if link weights are not positively correlated III ALGORITHM DEVELOPMET The proble of the optial bandwidth on the link for given traffic deands with different QoS service classes can be seen as the Miniu Cost Multi-Coodity Flow Proble (MCMCF) in the single (coon) source ultiple destination network It is very coplex proble (-coplete) Instead of nonlinear convex optiization we used two-stage network optiization technique that is ore appropriate for such non-linear proble Looking for objective function (5) on the first stage of network optiization we are calculating the inial link weights between all pairs of capacity points Calculation of each weight value we call it: sub-proble Algorith is looking for the best expansion solution aong all possible (inial cost) On the second stage of optiization we are looking for the shortest path in the network with forer calculated link weights between node pairs; see fig 4 Link capacity that is capable to serve traffic deands of service class i we call facility It is used priarily to serve deands for QoS level but it can be used to satisfy traffic deands for QoS level j (j > i) Rerouting of traffic deands towards higher QoS level is the sae thing as facility conversion to lower QoS level In this odel conversion of traffic deand is peritted only in the direction toward higher QoS level It eans that alost satisfied service class (adaptive or best effort) can be treated as fully satisfied (guaranteed) service class, but not vice versa Explanation of diagra in fig 3: the -th row of nodes represents a possible link capacity state of each transission link between s for i-th QoS level Generalizing the concept of the capacity states for each quality level of transission link in which the capacity states of each link are known within defined liits we define a capacity point - α α = (I,, I 2,,, I, ) (3) α = α M+ = (0, 0,, 0) (4) In forulation (3) α denotes the vector of capacities I for all QoS levels (facility types) on link, and we call it capacity point Each colun on the flow diagras (fig 3) represents a capacity point, consisting of capacity state values Forulation (4) iplies that idle capacities or capacity shortages are not allowed on the link between edge and interior Total capacity value on the link can be positive only and shortages are not allowed Horizontal links between the represent the traffic flow between s Coon node O is the source for used capacity (expansions), introducing the new traffic on the link Vertical links represent facility conversion that is equal to rerouting of traffic deands to higher quality level (in opposite direction) If the link expansion cost corresponds to the weight of used capacity, the objective is to find optial routing policy that iniizes the total cost incurred over the whole path between edge s (M interior s and M + transission links) and to satisfy given traffic deands The flow theory enables separation of these extree flows, which can be a part of an optial expansion solution, fro those which cannot be With such heuristic approach we can obtain the near-optial result with significant coputational savings; see [4] The objective function for CEP proble can be forulated as follows: M + in c + ( ) + ( ) x h I + g y (5) = i= so that we have:, d(,) d(,c2) I = I + x y ri, j= i+ = I M + = 0 2, d2(,) d2(,c3) 2,C 2 + (6) d3(,) d3(,c4) d22(,) dm(,) d22(c2,c3) I (7) for =, 2,, M+; i =, 2,, ; j = i +,, In the objective function the total cost (weight) on the path fro edge to edge includes soe costs: the cost for capacity expansion c (x ), the idle capacity cost 3, d23(,) d22(,c3) 3,C 3 d2m(,) d2m(c2,) d23(,c4) d2m(c2,) d33(,) M+, Figure 4 Suppose that all links in the network are known, the optial solution for CEP can be found by searching for the optial sequence of capacity points and their associated link values 4, d33(c3,c4) 4,C 4 d3m(,) d3m(c3,) d4m(,) d4m(c4,)

5 D i x,v x,u x,u+ Guaranted (fully satisfied) I,u,u I,u+,u+ I,u+2,v I,v+ r,u r,u+ r,v+ QoS service classes y 2,u y 2,u+ y 2,v adaptive (al ost satisfied) I 2,u 2,u I 2,u+ 2,u+ I 2,u+2 2,v I 2,v+ r 2,u y 3,u r 2,u+ y 3,u+ r 2,v+ y 3,v y 23,u y 23,u+ y 23,v best-effort I 3,u I 3,u+ 3,u 3,u+ I 3,u+2 3,v I 3,v+ r 3,u r 3,u+ r 3,v+ links Figure 5 A network flow representation of a sub-proble for =3 h (I + ) as penalty cost to force the usage of the iniu link capacity (prevention of unused/idle capacity) For the link expansion in allowed liits we can set the expansion cost to zero or we can introduce different cost to each service class Also we can introduce facility conversion cost g (y ) that can control non-effective usage of link capacity Costs are often represented by the fix-charge cost or with constant value We assue that all cost functions are concave and nondecreasing, reflecting econoies of scale, and they change for appropriate link With different cost paraeters we can influence to the optiization process, looking for the ost appropriate routing solution A Capacity Expansion Sub-proble (CES) Let the value d u,v (α u, α v+ ) represents the iniu cost between two capacity points Calculation of that value is denoted as CES (Capacity Expansion Subproble) The expansion sub-proble for facilities i =, 2,, on the path between s u and v is as follows (see fig 5): v in c ( x ) + hi, ( I + ) + g ( y ) (8) = u i= j= + where I v+ = I u + Di ( u, v) Ri ( u, v) (9) v R i u, v) = r i, = u v D i (u,v)= = u ( (0) j = x i, y ; i j () for =, 2,, M+; i =, 2,, ; j = i +,, Let C be the nuber of capacity point values at position (for link between core s) Only one capacity point for the link that connects to the edge : C = C M+ = The total nuber of capacity points is: S I, 0 I 2, = x = D y 23 = 2 ; (- D 2 ) I 3, = 0 Internal M = + p C = C (2) In the CEP we have to find any cost values d u,v (α u, α v+ ) that eanate two capacity points, fro each node (u, α u ) to node (v+, α v+ ) for v u; see fig 4 The total nuber of all connections between capacity points is: = M d i M + Ci C j = j= i+ D 0 D 2 = -2 D 3 = 3 r, x 3 = ; (D 2 + D 3 ) r 2, = - r 3, = 3 Figure 6 An expansion solution I,+ I 2,+ = 0 I 3,+ = 0 + (3) The ost of the coputational effort is spent on coputing the sub-proble values The nuber of all possible d u,v values depends on the total nuber of capacity points It is very iportant to reduce that nuber

6 (C p ) and that can be done through iposing of appropriate capacity bounds or by introduction of adding constraints Lot of expansion solutions is not acceptable and they are not part of the optial sequence Any of the if it cannot be a part of the optial sequence is set to infinity This is the key of the heuristic approach The approach described in [7] requires solving repeatedly a certain single location expansion proble (SLEP) Many different expansion and conversion solutions (rerouting) can be derived, depending on D i polarity An exaple of the expansion solution can be seen in fig 6 In that case preiu class (fully satisfied) has to be expanded independently Traffic deands of lowest service class can be satisfied with capacity of higher class (conversion) because traffic deands of second class has negative polarity and idle capacity (surplus) on the link is present With such sharing approach of link bandwidth aong different quality levels we can solve traffic deands efficiently B Algorith coplexity Suppose that all links are known, the optial solution for CEP can be found by searching for the optial sequence of capacity points and their associated link state values for each location On that network optiization level proble can be seen as a shortest path proble for an acyclic network in which the nodes represent all possible values of capacity points and links represent sub-proble values; see fig 4 Than Dijkstra s algorith or any siilar algorith can be applied It has to be noted that the shortest path in such network need not to be the shortest path solution for routing of traffic flow (new SLA) between edge s Also, on this level of algorith calculation it is very easy to introduce delay liits on the path The nuber of all possible d u,v values depends on the total nuber of capacity points It is very iportant to reduce that nuber (C p ) and that can be done through iposing of appropriate capacity bounds or by introduction of adding constraints The required effort for one sub-proble is O( 2 M) The nuber of all possible d u,v values depends on the total nuber of capacity points If there are no liitations on capacity state (I ) and expansion aount (x ) the coplexity of such heuristic approach is pretty large and increases exponentially with Proble requires the coputation effort of O(M 3 4 R i 2(-) ) In real application Figure 7 In this nuerical test-exaple link capacities are sufficient to satisfy traffic deands (SLAs) Figure 8 In this nuerical test-exaple the lack of capacity is obvious and new SLA cannot be accepted

7 we norally apply definite granularity of capacity values It reduces the nuber of the capacity points significantly IV TESTIG RESULTS We tested our algorith on any nuerical testexaples, looking for the best routing sequence on the path Between edge s there are axiu six LSR (core s) and they are connected with seven links Traffic deands (SLAs) are overlapping in tie The heuristic algorith in all test-exaples can achieve nearoptial expansion sequence with inial total cost, that is equal or very close to that one we can get with algorith based on exact approach (calculating all possible capacity points and all sub-proble values) If the expansions of link capacities are possible (in allowed liits) or the capacity surplus is obvious, the traffic deands can be satisfied In exaple fro figure 7 we can see that traffic deands can be fully satisfied and capacity surplus for all links are positive or zero It eans that there are no negative values; see lower diagra in fig 7 In exaple fro figure 8 we can notice the lack of capacity on the second link (for second QoS level) It eans that traffic deands cannot be fully satisfied although the surplus is obvious on third (lower) QoS level There is no free capacity on the first QoS level (for the preiu service class) to be used instead It eans that adding capacity (fro P) cannot be arranged In that case new SLA cannot be accepted or ust be redefined through negotiation process With such anageent tool we can predict sufficient bandwidth on the link and possible congestions on the path very efficiently Also, we can introduce new link capacity resources (bandwidth) in optial way Optial bandwidth on the link and bandwidth arrangeent for given SLAs can be done fro P dynaically Such explicit traffic engineering technique provides the possibility to intelligently tailor the route foe each SLA traffic flow such that different parts of the network reain equally loaded Also, it helps to avoid the creation of bottleneck links on the path and ensures optial network resource utilization V COCLUSIOS In the process of SLA anageent in DiffServ/MPLS networks service provider (SP) needs bandwidth optiization technique to ensure sufficient resources fro P to satisfy traffic deands caused by forer contracted SLAs In this paper we proposed new bandwidth optiization technique that can ensure sufficient bandwidth on the link It could be ade dynaically through SLA negotiation process With such tool we can significantly reduce the possibility of traffic congestions in the oent of service invocation Proposed heuristic algorith is based on atheatical odel for the capacity expansion proble (CEP) The first aspect of algorith is that load balancing leads to inial cost for used bandwidth and to higher resource usage efficiency It can prevent critical network resources, not to be exhausted early and becoing a bottleneck for the network in the oent of service invocation It eans that such heuristic approach can be successfully applied for adission control in the SLA negotiation process, that is in fir correlation with resource reservation echaniss and adission control process in DiffServ networks The iportant difference fro existing traffic routing techniques is that optial path for SLA need not to be the shortest path solution as we have in eg OSPF protocol (often used for traffic routing in the oent of service invocation) REFERECES [] Yu Cheng, R Farha, A Tizghada and all, Virtual etwork Approach to Scalable IP Service Deployent and Efficient Resorce Manageent, IEEE Counicattion Mag, 2005, Vol 43, o 0, pp76-84 [2] K Gopalan, Tzi-cker Chiueh, Yow-Jian Lin, Load Balancing Routing with Bandwidth-Delay Guarantees, IEEE Counicattion Mag, 2004, Vol 42, o 6, pp08-3 [3] S Giordano, S Salsano, G Ventre, Advanced QoS Provisioning in IP etworks : The European Preiu IP Projects, IEEE Counications Mag, 2003, Vol 4 o, pp30-36 [4] S Krile, ew SLA Creation and Optial Resource Manageent, Proc of 3rd Conf CSDSP (Counication Systes, etworks and Digital Signal Processing), ew Castle (UK), 2004, pp52-55 [5] D Kagklis, C Tsakiris, Liapotis,, Quality of Service: A Mechanis for explicit activation of IP Services Based on RSVP, Journal of Electrical Engineering, Vol 54, o 9-0, Bratislava, 2003, pp [6] S Krile, S Saric, SLA Creation in Relation to Resource Manageent, Journal of Electrical Engineering (JEE), 2003, Vol 54, o 9-0, pp [7] S Krile, M Kos, A Heuristic Approach for Path Provisioning in Diff-Serv etworks, Proc 7 th ISSSTA (Int Syp On Spread-Spectru Tech & Application) - IEEE, Prag, 2002, pp [8] K Biswas, S Ganguly and R Izailov, Path provisioning for service level agreeents in differentiated services networks, ICC IEEE International Conference on Counications, vol 25, o, 2002, pp [9] E Bouillet, D Mitra, G K Raakrishnan, The structure and anageent of service level agreeents in networks, IEEE Journal on Selected Areas in Counications, Vol 20, o 4, 2002, pp69-699