Load Control for Overloaded MPLS/DiffServ Networks during SLA Negotiation

Transcription

1 Int J Counications, Network and Syste Sciences, 29, 5, doi:14236/ijcns Published Online August 29 ( Load Control for Overloaded MPLS/DiffServ Networks during SLA Negotiation Srećko KRILE 1, Dragan PERAKOVIĆ 2 1 University of Dubrovnik, Departent of Electrical Engineering and Coputing, Dubrovnik, Croatia 2 University of Zagreb, Faculty of Transport and Traffic Engineering, Zagreb, Croatia Eail: sreckokrile@uniduhr, draganperakovic@fpzhr Received April 16, 29; revised May 2, 29; accepted July 2, 29 ABSTRACT In end-to-end QoS provisioning soe bandwidth portions on the link ay be reserved for certain traffic classes (and for particular set of users) so the congestion proble of concurrent flows (traversing the network siultaneously) can appear It eans that in overloaded and poorly connected MPLS/DS networks the CR (Constraint-based Routing) becoes insufficient technique If traffic engineering is supported with appropriate traffic load control the congestion possibility can be predicted before the utilization of guaranteed service In that sense the initial (pro-active) routing can be pre-coputed uch earlier, possible during SLA (Service Level Agreeent) negotiation In the paper a load siulation technique for load balancing control purpose is proposed It could be a very good solution for congestion avoidance and for better load-balancing purpose where links are running close to capacity To be acceptable for real application such coplicated load control technique needs very effective algorith Proposed algorith was tested on the network with axiu M core routers on the path and detail results are given for N=3 service classes Further iproveent through heuristic approach is ade and results are discussed Soe heuristic options show significant coplexity savings that is appropriate for load control in huge networks Keywords: End-To-End Qos Provisioning, Traffic Engineering In MPLS/Diffserv Networks, Constraint-Based Routing, Load Control during SLA Creation 1 Introduction The core network for NGN (New Generation Network) is evolving to MPLS/DiffServ based network With capability in service differentiation techniques the network operator can ensure the traffic priorization, specialy to quality voice (VoIP) and video calls (preiu traffic), as sae as for truly differentiated data services It eans that DiffServ classifies individual flows in a sall nuber of service classes (at network edges) Also it enables soft reservation (allocation) of resources and special handling of packets in the core Together, MPLS (Multi Protokol Label Switching) and DiffServ provide a scalable QoS solution for the core of NGN; see [1] and [2] MPLS uses extensions to Resource Reservation Pro- tocol (TE-RSVP) and the MPLS forwarding paradig to provide explicit routing; see [3-5] But end-to-end provisioning of coexisted and aggregated traffic in networks is still deanding proble, specially if over- provisioning is not possible All traffic flows in doain are distributed aong LSPs (Label Switching Path) related to N service classes, but we know the IGP (Interior Getaway Protocol) uses siple on-line routing protocols (eg OSPF, IS-IS) based on shortest path ethodology With OSPF (Open Shortest Path First) soe paths ay becoe congested while others are underutilized Constraint-based routing (CR) as an extension of explicit routing ensures traffic engineering (TE) capabilities CSPF (Constrained Shortest Path First) allows an originating (ingress) router to copute a path (LSP) to egress router (sequence of interediate LSRs), taking care of constraints such as bandwidth, delay and adi- nistrative policy; see [12] It Copyright 29 SciRes Int J Counications, Network and Syste Sciences

2 LOAD CONTROL FOR OVERLOADED MPLS/DIFFSERV NETWORKS DURING SLA NEGOTIATION 423 can find out a longer but lightly loaded path, better than the heavily loaded shortest path With constraint-based label distribution protocol (CR-LDP) the bandwidth provisioning dire- ctives and other inforation can be ensured (list of router's neighbors, attached networks, actual resource availability and other relevant inforation) It can be distributed for each service class and for each link along the path (LSP); see [6] CR process can be incorporated into each ingress router and co-exists with the conventional routing technique MPLS/DiffServ aware TE (DS-TE) allows constraintbased routing of IP traffic with final task to adjust class load to actual class capacity But the routing approach explained above can be effective in under loaded networks or in fully connected networks only For the the RED or WRED (Weighted Rando Early Detection) are effective congestion avoidance techniques But in soe networks dropping packets can lead to custoer dissatisfaction and SLA violation In the context of siultaneous flows (forer contracted SLAs + new SLA creation) bandwidth overbooking is possible and congestion proble during service utilization can appear For overloaded and poorly connected networks we need prediction of congestion probability uch earlier Load balancing has to be done uch before the oent of service utilization, possibly during SLA negotiation process To obtain quantitative end-to-end guarantees the QoS provisioning has to be in fir correlation with bandwidth anageent; see [7,8] Also, such load control as a part of DS-TE can help in optial bandwidth reservation, to predict sufficient resources and to ensure better end-to-end QoS provisioning, see [14] TE involves the anageent of existing bandwidth resources to suit trafic deands and to eet the growing deands of custoers in the network New approach in load control during SLA creation is given in Section 2 Load control can be seen as the capacity expansion proble (CEP) of N capacity types Expansions are possible in given liits for each bandwidth portion (sub-pool), influencing on each other The atheatical odel explanation for CEP is given in the Section 3 In the Section 4, we have CEP algorith developent and heuristic approach The coparison of results for different algorith options we can see in the Section 5 2 LSP Creation during SLA Negotiation The service provider for doain (eg ISP) wants to accept a new SLA that results with priority traffic flow between edge routers A traffic trunk is defined as a logical pipeline within an LSP, with reservation of certain aount of capacity to serve the traffic associated with a certain SLA So it is clear that LSP between an ingress/egress pair ay carry ultiple traffic trunks associated with different SLAs; see [1] In Figure 2 we have situation on the path for the exaple of siultaneous SLA flows fro Figure 1 All traffic flows on the path are participating possibly in the sae tie (the worst case) In that sense the network operator (eg ISP) has to find the optial LSPs for aggregated flows without any possible congestion in the core network; see [9] Each traffic deand can be satisfied on appropriate or higher QoS level, using appropriate bandwidth portion fro a sub-pool The ain condition is: the sufficient network resources ust be available for the priority traffic at any oent SLA_3 egress SLA_6 ingress Edge router ingress SLA_1 ingress SLA_2 ingress SLA_4 ingress s Edge router SLA_1 1 2 SLA_ egress SLA_2 egress SLA_4 egress SLA_5 egress SLA_6 egress d Edge router Sla core network SLA_1 egress SLA_3 egress SLA_5 ingress SLA_ 5 Edge router Figure 1 An exaple of nuber of flows (forer SLAs) in the context of new SLA creation Copyright 29 SciRes Int J Counications, Network and Syste Sciences

3 424 S KRILE ET AL SLA_ 1 SLA_ 3 SLA_ SLA_ 5 SLA_1 SLA_ 3 SLA_1 SLA_ 5 SLA_ 5 Figure 2 Siultaneous flows with possibly congestion on the path So we propose the next scenario: during SLA negotiation process the RM (Resource Manager) odule has to deterine the ain paraeters that characterize the required flow (ie, bandwidth, QoS class, ingress and egress IP router addresses, tie of service utilization); see [13] At first RM can apply off-line routing to get initial LSP, taking care of all other siultaneous flows It can be done with any shortest path-based routing algorith (eg OSPF) Fro RM we can get statistical details for each SLA flow traversing the network siultaneously, eg ingress node, egress node, service class, bandwidth etc) We need very effective algorith to siulate such traffic load and detect congestion possibility on the path Such optiization is a ulti-constrained proble (MCP) The BB (Bandwidth Broker) will check if there are enough resources on the calculated path to satisfy the requested service class If such calculated path has any link that exceeds allowed capacity liits (axial bandwidth for appropriate service class) possible congestion exists; see [15] It eans that link capacity on the path cannot be sufficient for new traffic load Alternatively, adding capacity arrangeent (dynaic bandwidth reservation) for congested link has to be done but it ay produce significant extra cost Abut dynaic bandwidth reservation echanis we can see in [2] and [21] If calculation finds that proposed path has no any congestion the new SLA can be accepted and related LSP is assigned to that traffic flow (SLA) and stored in database of RM In opposite the new SLA cannot be accepted or ust be re-negotiated In the oent of service invocation such calculated (and stored) LSP can be easily distributed fro RM to the MPLS network to support explicit routing, leveraging bandwidth reservation and prioritization; see [16] In that way the LSP creation should be in co-relation with SLA, to enable better load-balancing and to ensure congestion avoidance in doain Also, such load balan- cing technique is appropriate for inter-doain end-to-end path provisioning in the part of optial bandwidth reservation fro neighbor ASes (Autonoous Syste) Capacity reservations have to be done in the ost effective way in order to provide bandwidth guarantees for predicted traffic; see [11] Having purchased access to sufficient bandwidth fro downstrea ASes, hoe AS can utilize both: purchased bandwidth and its own network capacity 3 Matheatic Model of CEP for Congestion Control and Load Balancing Purpose The congestion control technique explained above can be seen as the capacity expansion proble (CEP) on the path with or without shortages If the fulfillent of traffic deand is the ain condition we talk about CEP without shortages Transission link on the path are capable to serve traffic deands for N different QoS levels (service class) for i = 1, 2,, N For each traffic load we need appropriate bandwidth aount, so it looks like bandwidth expansion Bandwidth portions on the link can be assigned to traffic flow of appropriate service class up to the given liit of bandwidth sub-pool (axial capacity for defined service class) Used capacity can be increased in two fors: by expansion or by conversion Expansions can be done separately for each service class or through conversion (redirected aount) to lower quality class It eans that capacity can be reused to serve the traffic of lover quality level under special conditions For exaple, if capacity predefined for priority traffic is unused it can be redirected to support best effort services Bandwidth usage for each service class (sub-pool) can be a part of resource reservation strategy but su of all sub-pools has to be equal/less than link capacity Figure 3 gives an exaple of network flow representation for network flow for ultiple QoS levels (N) and M core routers (LSR) and M+1 links on the path Copyright 29 SciRes Int J Counications, Network and Syste Sciences

4 LOAD CONTROL FOR OVERLOADED MPLS/DIFFSERV NETWORKS DURING SLA NEGOTIATION 425 R (1,M) O X 1 1 X 2 X M 2 M x 1,2 x 1,M I 1,1 = x 1,1 1,1 I 1,2 1,2 I 1,3 I 1,M 1,M I 1,M+1 = 1,M+1 x 1,2 r 1,1 x 2,2 r 1,2 x 2,M r 1,M Neighbor AS QoS levels (service classes) I 2,1 = I N,1 = y 1,2,1 2,1 y 2,N,1 x 3,2 I 2,2 y 1,2,2 2,2 y 2,N,2 I N,2 N,1 N,2 I 2,3 I N,3 x 3,M I 2,M x 1,3 r 2,1 y 1,N,1 r 2,2 y 1,N,2 r 2,M I N,M y 1,2,M 2,M y 2,N,M I N,M+1 = N,M I 2,M+1 = y 1,N,M 1,M+1 1,M+1 Destination is reachable through downstrea ASes Offered bandwidth guarantee (possible SLA) edge router r N,1 interior router 1 r N,2 r N,M link 1 link 2 link M interior router M link M +1 edge router Figure 3 The network flow representation of the CEP for Russian Dolls bandwidth allocation odel Network has V core routers, M V, and A links connecting all of the, including edge routers In the CEP odel the following notation is used: j and k = QoS level We differentiate n service classes (QoS levels)the N levels are ranked fro i = 1, 2,, N, fro higher to lower = on the path, connecting two successive routers 1 and 2, = 1,, M+1 r = traffic deand increent for additional capacity on the link fro appropriate sub-pool i For convenience, the r i is assued to be integer For the flow going out fro the path r is negative The su of traffic deands for capacity type i between two routers on the path: 2 i ( 1, 2) r i, 1 R (1) x = the aount of adding capacity for appropriate service class i on the link Possible negative values decrease It eans that we have idle (sufficient) capacity If dynaic bandwidth anageent is possible such reduction can be realized X N i 1 x i, (2) The su of deands for whole path and for all capacity types has to be positive or zero (including new SLA): N i 1 M R i (1, M ) X (3) 1 Fro this forulation it is obvious that su of traffic deands on the path has to be equal to capacity aount used to satisfy the It eans that we don t expect reduction of total capacity on the path toward egress router, in other words we presue the increase of capacity y j, = the aount of capacity for quality level i on the link, redirected to satisfy the traffic of lower quality level j Any traffic deand can also be satisfied by converted capacity fro any capacity type k < i with higher quality level In Figure 3, such flows are arked with doted lines I = relative aount for the capacity type i on the link, connecting two neighbor routers Idle capacity is represented with positive value If shortages are not allowed negative value cannot exist I i1 =, I M+1 = that eans: no adding capacity is necessary on the link toward edge routers It eans that the capacity for that link is always sufficient Copyright 29 SciRes Int J Counications, Network and Syste Sciences

5 426 S KRILE ET AL L = bandwidth constraints for link capacity values on the link and for appropriate service class i (L 1,, L 2,, L N, ) w = weight for the link and appropriate service class i (QoS level) del = delay on the link for appropriate service class i Maxial delay on the path is denoted with DEL i As we have nonlinear expansion functions (showing the econoy of scale) the CEP can be solved by any nonlinear optiization technique Instead of polynoial optiization (eg nonlinear convex prograing), that can be very coplicated (NP-coplete), the network optiization ethodology is efficiently applied The ain reason on such approach is the possibility of discrete capacity values for liited nuber of QoS classes, so the optiization can be significantly iproved The proble can be forulated as Miniu Cost Multi-Coodity Flow Proble (MCMCF) Such proble can be easily represented by ulti-coodity the single (coon) source ultiple destination network; see Figure 3 Let G (V, A) denote a network topology, where V is the set of vertices(nodes), representing link capacity states and A, the set of arcs, representing traffic flows between routers Each link on the path 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 In general we have ulti-constrained proble (MCP) but in this paper we talk about one-diensional link weight vectors for M+1 links on the path w, A, i = 1,, N Eg the capacity constraint for each link on the path is denoted with L (L 1, L 2,, L N, ) For non-additive easures (eg for bandwidth where the cost-function is concave, and we are looking for iniu) definition of the singleconstrained proble is to find a path fro ingress to egress node with inial link weight along the path In the context of MCP we can introduce easily the adding constraint of ax Delay on the path (end-to-end) As it is an additive easure (ore links on the path cause higher delay) it can be used as criteria to eliinate any unacceptable capacity expansion solution fro calculation The flow situation on the link depends of expansion and conversion values (x, y j, ) It eans that the link weight (cost) is the function of used capacity: lower aount of used capacity (capacity utilization) gives lower weight If the link expansion cost corresponds to the aount of used capacity, the objective is to find the optial expansion policy that iniizes the total cost on the path Definition of the single-constrained proble is to find a path P fro ingress to egress node such that: M 1 N ( j, 1 i 1 w P ) in w ( I, x, y ) (4) where: I L (5) satisfying condition of ax Delay for P: 2 1 del DEL (6) for i = 1,, N ; = 1,, M A path obeying the above conditions is said to be feasible Note that there ay be ultiple feasible paths between ingress and egress node Generalizing the concept of the capacity states for each quality level of transission link between LSRs in which the capacity states for each service class (QoS level) are known within defined liits we define a capacity point - = (I 1,, I 2,,, I N, ) (7) 1 = M+1 = (,,, ) (8) In forulation (37) denotes the vector of capacities I for each service class on link, and we call it capacity point On the flow diagras (Figure 2) each colun represents a capacity point of the node, consisting of N capacity state values (for i-th QoS level) Link capacity is capable to serve different service classes Capacity aount labeled with i is priarily used to serve traffic deands of that service class but it can be used to satisfy traffic of lower QoS Level j (j > i) Forulation (38) iplies that idle capacities or capacity shortages are not allowed on the beginning and on the end of the path It eans that process is starting with new SLA flow that ust be fully satisfied through the network (fro ingress to egress node) The objective function for CEP proble can be forulated as follows: in M 1 1 so that we have: N i1 x h I g y 1 i j, j, c (9) N I 1 I x y j, ri, ji1 (1) I (11) 1 I i, M 1 for = 1, 2,, M+1; i = 1, 2,, N; j = i + 1,, N In the objective function (39) the total cost (weight) includes soe different costs Expansion cost (adding capacity) is denoted with c (x ) For the link expansion in allowed liits we can set the expansion cost to zero We can differentiate expansion cost for each service class We can take in account the idle capacity cost h (I +1 ), but only as a penalty cost to force the usage of the iniu link capacity (prevention of unused/idle capacity) Also we can introduce facility conversion cost g j, (y j, ) that can control non-effective usage of link capacity (eg usage of higher service class capacity instead) Costs are often represented by the fix-charge cost or with constant value We assue that all cost functions Copyright 29 SciRes Int J Counications, Network and Syste Sciences

6 LOAD CONTROL FOR OVERLOADED MPLS/DIFFSERV NETWORKS DURING SLA NEGOTIATION 427 are concave and non-decreasing (reflecting econoies of scale) and they differ fro link to link The objective function is necessarily non-linear cost With different cost paraeters we can influence on the optiization process, looking for benefits of the ost appropriate expansion solution 4 Algorith Developent The network optiization can be divided in two steps At v N N duv, in ci, ( xi, ) hi, ( Ii, 1 ) g j, ( y j, ) u i1 j1 first step we are calculating the inial expansion weight d u,v for capacity expansion between two capacity points of neighbor links It has to be done for all capacity points and for all pairs of neighbor routers (interconnected) u,v = the order nuber of capacity points in the sub-proble for appropriate link, 1 u,, v M+1 The calculation of weight value between two capacity points we call: capacity expansion sub-proble (CES); see (41) The expansion sub-proble for N facilities i = 1, 2,, N on the path between routers u and v is as: (12) 1,1 d 1,2 (1,1) d 1,2 (1,C 2 ) 2,1 d 2,3 (1,1) d 2,3 (C 2,1) 3,1 d 2,3 (1,C 3 ) d 3,4 (1,1) d 3,4 (1,C 4 ) d 3,4 (C 3,1) 4,1 d 4,M+1 (1,1) d 4,M+1 (C 4,1) M+1,1 2,C 2 4,C 4 d 2,3 (C 2,C 3 ) d 3,4 (C 3,C 4 ) 3,C 3 Figure 4 The CEP proble can be seen as the shortest path proble for an acyclic network in which the nodes represent all possible values of capacity points and the links represent CES values where: I I D ( u, v) R ( u, ) (13) v u i i v v R i u, v) r i, u M 1 p C 1 C (16) ( (14) In the CEP we have to find any cost values d u,v ( u, v ) that eanate two capacity points, fro each node (u, u ) to node (v, v ) for v u The total nuber of CES v N can be pretty large: u j, D i (u,v)= x i y ; i j (15), j 1 for = 1, 2,, M+1; i = 1, 2,, N; j = i + 1,, N Let C be the nuber of the capacity point values for link between two neighbor core routers Only one capacity point for the link that connects the edge router: C 1 = C M+1 = 1 The total nuber of capacity points is: N d M 1 C C 1 (17) For every CES the calculation of any different expansion solutions can be derived fro D i value Many cobinations exist fro expansion and conversion aount The ost of the coputational effort is spent on coputing of the sub-proble values The nuber of all pos- Copyright 29 SciRes Int J Counications, Network and Syste Sciences

7 428 S KRILE ET AL sible d u,v values depends on the total nuber of capacity points; see Figure 4 Suppose that all links (sub-probles) are calculated, the optial solution for CEP can be found by searching for the optial sequence of capacity points and their associated link state values On that level the CEP proble can be seen as a shortest path proble for an acyclic network in which the nodes represent all capacity point values, and branches represent CES values; see Figure 4, Then Dijkstra s algorith or any siilar algorith can be applied The nuber of all possible d u,v ( 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 (eg ax delay) Through nuerical test-exaples we ll see that any expansion solutions cannot be a part of the optial expansion sequence 41 Single Location Expansion Proble Approach described in chapter above requires solving repeatedly a certain single location expansion proble (SLEP) in all possible odifications, looking for the best result Let SLEP j (, D i, D j ) be a Single Location Expansion Proble associated with link for facility (capacity) type i+1,, j and corresponding values of capacity change intention D i, D i+1,, D j For exaple, in solving SLEP 1,3 for three different capacity types (bandwidth sub-pools) we have any expansion solutions divided into three different scenarios (expansion strategies): a) capacity changes of one capacity type are not correlated with changes of others; b) capacity changes of two capacity types depend on each other, but change of the third is independent; c) capacity changes for all of three capacity types depend on each other Fro three expansion scenarios (expansion strategy) any different expansion solutions can be derived, depending on D i polarity A lot of the are not acceptable and are not part of optial sequence For this proble an acceptable expansion solution has to satisfy soe basic properties: x D (18) y j, D (19) y j, D j, (2) Property (411) iplies that the expansion (increase) of capacity type i cannot be acceptable if that facility has intention to be reduced on location (link) (D < ) Siilar stays for negative values Expansion (capacity increase) is also possible through conversion, so (18) and (19) iply the siilar restriction as (2) Zero value of any capacity type eans that any change of capacity is allowed In scenario A we have only one possible expansion solution In scenario B we can cobine all three capacity types in couples In scenario C we can see that only one expansion solution exists Totally, we have five different expansion solutions with any variations In scenarios B and C we have expansion solutions with conversions of capacity fro one type to another It can be done as stand-alone expansion or together with expansion That eans that the conversion is just copleentary with the expansion in satisfying of traffic deands Conversions can be applied only when idle capacities are noticed or negative deand increents are present Special case is occurred when conversion y j, eliinates both: eliinating idle capacity of type i plus satisfying traffic deands of capacity type j Also we can ake distinction between two options: the partial expansion and the excessive expansion The partial expansion x j, eans that deands are satisfied by expansion of appropriate capacity type j plus by conversion y j, of capacity type i with higher quality level, but only if shortage of facility i is not occurred The excessive expansion eans that the expansion aount x is used to partially expand facility i and to satisfy deands for lower capacity type j, with conversion aount y j, 42 Adding Properties (the Iproveent of CEP Algorith) The ost of the coputational effort is spent on coputing of the sub-proble values But a lot of expansion solutions are not acceptable and they cannot be a part of the optial expansion sequence The key for this very effective approach is in fact that extree flow theory enables separation of these extree flows which can be included in optial expansion solution fro those which cannot be Any of d u,v value, if it cannot be a part of the optial sequence, is set to infinity It can be shown that a feasible flow in the network given in Figure 3 corresponds to an extree point solution of CEP if and only if it is not the part of any cycle (loop) with positive flows, in which all flows satisfy given properties; see 18 One ay observe that the absence of cycles with positive flows iplies that each node has at ost one incoing flow fro the source node (positive or negative) This result holds for all single source networks That eans that optial solution of d u,v has at ost one expansion (or reduction) for each facility Copyright 29 SciRes Int J Counications, Network and Syste Sciences

8 LOAD CONTROL FOR OVERLOADED MPLS/DIFFSERV NETWORKS DURING SLA NEGOTIATION 429 Using a network flow theory approach, adding properties of extree point solution are identified These properties are used to develop an efficient search for the link costs d u,v Absence of such cycles with positive flows iplies that extree point solutions for CEP satisfy the following properties: I x (21) I y j, (22) I j, y j, (23) I j, x y j, = if x y j, (24) I I j, y k, y j,k, = if y k, y j,k, (25) for: j, k = 1, 2, 3 i k j ; = 1,, M+1 Properties (21) to (25) iply that the capacity of any capacity type is changed through an expansion, reduction or by conversion only if it doesn t ake cycles with positive flows (21) and (22) iply that the capacity of any capacity type can be increased by an expansion or by a conversion only if there is no idle capacity Siilar rule exists for reduction of idle capacity (23) iplies that capacity can be reduced only if there is no capacity shortage (24) iplies that incoing flow of facility, going to be converted (reduced) in partially or excessive expansion solution, has to be zero If not, cycles with positive flows can be occurred; see Figure 5, on that diagra we have idle capacity fro previous link (for first and second class) The third class is satisfied with capacity conversions S eans that such solution cannot be a part of extree solution and has to be avoided fro further calculation One of the capacity values (I 2, or I 3, ) ust be zero Property (25) is used for siultaneous ulticonversion solution fro scenario C Only one incoing flow of converted (reduced) facility can exist It eans that two incoing flows are not allowed in the sae tie In the case of siultaneous conversions, incoing flows have to be zero We can say that any acceptable SLEP 1,3 expansion solution for any CES have to satisfy properties (18)-(2) and (21)-(25) So any expansion solutions are not a part of optial sequence and could be eliinated fro further coputation; see 19 It eans that any of subproble value if it cannot be a part of the optial sequence is set to infinity 5 Testing Results and Coparison of Different Algorith Options The proposed algorith is tested on any nuerical test-exaples, looking for optial expansion sequence on the path Between edge routers there are axiu M core routers (LSR) and the path consists of axiu M+1links Traffic deands (forer contracted SLAs) are given in relative aount for each interior router on the path Deands are overlapping in tie and are defined for each capacity type (service class) Results obtained by iproved algorith (reduction of unacceptable expansion solutions) are copared with results obtained by referent algorith that is calculating all possible expansion solutions for each CES For each test-exaple we know the total nuber of I 1, = 1 D 1 = -3 r 1, = -2 I 1,+1 = Priority traffic Y 1,3, = 3 ; (- D 3 ) I 2, =2 D 2 = -1 I 2,+1 = Y 1,3, = 1 ; (- D 2 ) r 2, = 1 The best effort I 3, =? D 3 = 4 I 3,+1 = 2 r 3, = Figure 5 An exaple of single location expansion solution that cannot be a part of the extree solution of higher classes On that diagra dotted lines ark a cycle with positive flows fro the coon source It Figure 6 Trends of algorith coplexity and coparison of results (inial cost) Copyright 29 SciRes Int J Counications, Network and Syste Sciences

9 43 S KRILE ET AL Table 1 Results of nuerical test-exaple Traffic deands (increent) Routers on the path r 1, r 2, r 3, Algorith option The best result (inial cost) Nuber of capacity points Nuber of sub-probles satisfying properties Coputational savings in perc (%) Full approach Basic_A M_H A_H R_H P_H T_H ,7 (59,3) 59,35 (4,65) 63,1 (36,99) 82,31 (17,69) 96,8 (3,2) 99,99 (,1) c (x ) = f -1 i (A i +B i x ai ), A 1 =3, B 1 =25, a 1 = 9, A 2 =1, B 2 =2, a 2 =85, A 3 =2, B 3 =3, a 3 =95 For negative expansions (x < ) c (x ) = - f -1 i (B i abs(x ) ai ) h (I +1 ) = f i -1 H i I +1, H i = 4 (I > ) for i = 1, 2, 3 Shortages (I < ) are not allowed; g t (y j, ) = f -1 i G i y j,, G i =1 for y j, >, i < j Conversions in opposite direction (y i j,, < ) are not allowed; For all positions and for all quality levels it is the sae value f i = 9 All cost values are the sae no atter of the link position on the path In this test-exaple expansion liits L are satisfied Optial usage of the capacity (expansion sequence): y 1,3,1 = 1, x 1,2 = 1, x 2,2 = 1, x 1,4 = 2, y 1,3,4 = 1, x 2,4 = -1, x 1,5 =-1, x 1,6 = 1, x 2,6 = 1 Minial cost: capacity points The nuber of possible CES is wellknown, so it is the easure of the coplexity for the CEPproble Also, for each test-exaple we can see the nuber of acceptable sub-probles, satisfying basic and additional properties of optial flow; see exaple fro table 1 For all nuerical test-exaples the best possible result (near-optial expansion sequence) can be obtained with iproved algorith (denoted with Basic_A), sae as with referent algorith (without reduction of unacceptable expansion solutions) For N=3 and M=6 algorith coplexity savings in percents are on average ore that 4 % that is proportionally reflected on coputation tie savings; see Figure 6 The nuber of all possible CES values depends on the total nuber of capacity points So CEP requires the coputation effort of O(NMN d ) with linear influence of N In real application we norally apply definite granularity of capacity values through discrete values (integer) of traffic deands R i It reduces the nuber of the capacity points significantly Because of that the inial step of capacity change (step_i i ) has strong influence on the algorith coplexity In real situation we can introduce soe liitations on the capacity state value, talking about heuristic algorith options: b) Total su of the link capacity values (for all quality levels) is positive A_H (Acceptable Heuristic option); c) Total su is positive but only one value can be negative Such option is denoted with R_H (Real Heuristic option); d) Algorith option that allows only non-negative capacity state values is denoted with P_H (Positive Heuristic option); e) Only null capacity values are allowed A trivial heuristic option (denoted with T_H) allows only zero values in capacity point (only one capacity point) a) Only one negative capacity value in the capacity point Such option is denoted with M_H (Minial-shortage Heuristic option); Figure 7 The coplexity savings increase with value M Copyright 29 SciRes Int J Counications, Network and Syste Sciences

10 LOAD CONTROL FOR OVERLOADED MPLS/DIFFSERV NETWORKS DURING SLA NEGOTIATION 431 We copared the efficiency of algorith in above entioned options In Figure 6 we can see the average values of results for N=3 and M=6 Only for few test-exaples (see table 1) we can find the best expansion sequence, providing the inial cost, no atter of algorith option we use For the ost exaples algorith option M_H can obtain the best result with average saving ore than 6 % For other algorith options the significant reduction of coplexity is obvious but deterioration of result appears In the ost cases the trivial algorith option (T_H) shows the significant deterioration of final result A good fact for all algorith options is that efficiency rises with increase of value M; see Figure 7 6 Conclusions Inappropriate bandwidth reservation or wrong traffic load could result in congestion possibilities In this paper we propose an efficient algorith for congestion control and load balancing purpose during SLA negotiation process We can check congestion probabilities on the path with algorith of very low coplexity first (eg P_H algorith option) It eans that only if congestion appears we need optiization with ore coplex algorith (eg A_H) With the ost coplex algorith option (Basic_A) we can get the best possible result, so we can be sure if congestion on the path could appear or not In the case of congestion appearance new SLA cannot be accepted or adding capacity arrangeent should be done It eans that SLA re-negotiation has to be done and custoer has to change the service paraeters: eg bandwidth (data speed), period of service utilization etc The proposed algorith for load control (with different options) can be efficiently incorporated in SLA negotiation process It ay iprove end-to-end provisioning, especially for overloaded and poorly connected networks where over-provisioning is not acceptable 7 References [1] F L Faucheur, et al, Multi-protocol label switching (MPLS) support of differentiated services, Technical Report RFC 327, IETF, 22 [2] V Sarangan and C Jyh-Cheng, Coparative study of protocols for dynaic service negotiation in the next-generation Internet, IEEE Counication Magazine, Vol 44, No 3, pp , 26 [3] N Degrande, G V Hoey, P L V Poussin, and S Busch, Inter-area traffic engineering in a differentiated services network, Journal of Network and Systes Manageent (JNSM), Vol 11, No 4, 23 [4] M D Arienzo, A Pescape, and G Ventre, Dynaic service anageent in heterogeneous networks, Journal of Network and Systes Manageent (JNSM), Vol 12, No 3, pp , 24 [5] K Haddadou, S G Doudane, et al, Designing scalable on-deand policy-based resource allocation in IP networks, IEEE Counications Magazine, Vol 44, No 3, pp , 26 [6] R Boutaba, W Szeto, and Y Iraq DORA: Efficient routing for MPLS traffic engineering, Journal of Network and Systes Manageent (JNSM), Vol 1, No 3, pp , 22 [7] Y Cheng, R Farha, A Tizghada, et al, Virtual network approach to scalable IP service deployent and efficient resource anageent, IEEE Counication Magazines, Vol 43, No 1, pp 76-84, 25 [8] S Lia, P Carvalho, and V Freitas, Distributed adission control for QoS and SLS anageent, Journal of Network and Systes Manageent (JNSM), Vol 12, No 3, pp , 24 [9] J Guichard, F L Faucheur, and J P Vasseur, Definitive MPLS Designs, Cisco Press, pp , 25 [1] O Younis and S Fahy, Constraint-based routing in the internet: Basic principles and recent research, IEEE Counications Surveys & Tutorials, Vol 5, No 1, pp 2-13, 3rd quarter 23 [11] K H Ho, P H Michael, N Wang, G Pavlou, and S Georgoulas, Inter-autonoous syste provisioning for end-to-end bandwidth guarantees, Coputer Counications, Vol 3, No 18, pp , 27 [12] S Bhatnagar, S Ganguly, and B Nath, Creating ultipoint-to-point LSPs for traffic engineering, IEEE Counications Magazines, Vol 43, No 1, pp 95-1, 25 [13] M Morrow and A Sayeed, MPLS and next-generation networks: Foundations for NGN and enterprise virtualization, Cisco Press, 26 [14] S Bakiras and L Victor, A scalable architecture for end-to-end QoS provisioning, Coputer Counications, Vol 27, No 13, pp , 24 [15] S Giordano, S Salsano, and G Ventre, Advanced QoS provisioning in IP networks: The European preiu IP projects, IEEE Counication Magazines, Vol 41, No 1, pp 3-36, 23 [16] D Kagklis, C Tsakiris, and N Liapotis, Quality of service: A echanis for explicit activation of IP services based on RSVP, Journal of Electrical Engineering, Vol 54, No 9-1, Bratislava, pp , 23 [17] H Luss, A heuristic for capacity expansion planning with ultiple facility types, Naval Res Log Quart, Vol 33, No 4, pp , 1986 [18] W I Zangwill, Miniu concavecost flows in certain networks, Mgt Sc Vol 14, pp , 1968 Copyright 29 SciRes Int J Counications, Network and Syste Sciences

11 432 S KRILE ET AL [19] S Krile and D Kuzuilovic, The application of bandwidth optiization technique in SLA negotiation process, Proceedings of 11th CAMAD 6, International Workshop on Coputer-Aided Modeling, Analysis and Design of Counication Links and Net Network, Trento, pp , 26 [2] S Dasgupta, J C D Oliveira, and J P Vasseur, A new distributed dynaic bandwidth reservation echanis to iprove resource utilization: Siulation and analysis on real network and traffic scenarios, Proceedings of 25th IEEE International Conference on Coputer Counications INFOCOM, Barcelona, pp 1-12, 26 [21] S Dasgupta, J C D Oliveira, and J P Vasseur, Dynaic traffic engineering for ixed traffic on international networks, Coputer Networks, Vol 11, pp , 28 Copyright 29 SciRes Int J Counications, Network and Syste Sciences