ture resources to support the service levels guaranteed in the SLA. A service level is used to define the expected performance behavior of a deployed EF service ...
Dimensioning Network Resources in DiffServ over MPLS based Expedited Forwarding Service Subclasses Hamada ALshaer and Eric Horlait Networks and Performance Analysis University Pierre et Marie Curie Paris,France {hamada.alshaer,eric.horlait}@lip6.fr Abstract We use the highest differentiated service code point (DSCP ) values in DiffServ architecture, and employ the traffic engineering possibilities through MPLS techniques in order to introduce new premium services by splitting the expedited forwarding service class into three premium service subclasses in MPLS enabled DiffServ network. Consequently, Internet Service Providers (ISP s) will be able to announce and provide their customers different premium services based on Service Level agreements (SLAs) accomplished between them. Keywords DiffServ , MPLS , Network Dimensioning , SLA , e2e QoS , Traffic Engineering
1.
Introduction
The IP SLAs typically contain many specifications, among these we will focus on the expected performance level of Expedited Forwarding (EF ) service class in DiffServ network, which includes two major aspects: reliability and responsiveness. The first includes availability requirements; when the service is available, and what the bounds are on service outages that may be expected. Responsiveness includes how soon the service is performed in the normal course of operations. Which require provisions sufficient infrastructure resources to support the service levels guaranteed in the SLA. A service level is used to define the expected performance behavior of a deployed EF service, where the performance metrics are end to end (e2e) delay bound, jitter and loss. During deployment of EF service, the resources of an underlying EF service container can be reconfigured to provide customers one of the different EF service levels according to their SLA.
2.
Network Model and Problem Formulation
We suppose that the network shown in in Fig. 1 represents the Internet network installed where the Olympic games have taken place in Athens. Over this network, different virtual private production networks are configured for establishing connections between outside broadcasting (OB) vans and their main studios in the different countries mentioned in the figure.
Germany Site Main Studio
Greece Site Main Studio
AF Dest 3 EF Dest 3
Italy Site Outside Broadcast
AF Dest 4
BE Dest 3 BE Dest 4
Edge 7
EF Dest 4
AF Source 2 EF Source 2 BE Source 2
Edge 9
Edge 5 EF source 1
Edge 1
France Site Outside Broadcast
Core A
Core C
Core2
Core3
Core F EF Dest1
France Site Main Studio
Core 5
AF source 1
Edge 2
Core 4
Core 1
AF Dest 1
Edge 4
BE source 1
Edge 3
BE Dest 1 Core B
Core D
Core G
Default EF Service Enhanced EF Service Top EF Service
Edge 8 Edge 10
Edge 6
BE Source 3
Priority, Drop Precedence
EF Subclass_1
EF Subclass_2
EF Subclass_3
EF Dest 2 BE Dest 2 EF Source 4
BE Source 4 AF Source 4
AF Dest 2
Italy Site Main Studio
Greece Site Outside Broadcast
AF Source 3 EF Source 3
Germany Site Outside Broadcast
Low , Low
101 010
110 010
Medium, Medium
101 100
110 100
111 100
High , High
101 110
110 110
111 110
111 010
Figure 1: Production Networks Communicate Figure 2: Recommended EF service Through DiffServ Over MPLS network. subclasses code point value.
In [1], the authors propose to use the DiffServ network and MPLS to transport the production networks traffic, so they introduce a traffic engineering and control system for handling efficiently the different service level specifications (SLSs) while optimizing the use of network resources. In [2], the authors formulate a multi-criterion optimization problem on a predefined traffic matrix. Furthermore, they design in [3] VPNs such that a customer traffic is optimally routed over his VPN, which contributes in increasing the number of accepted connections or the traffic intensity in the network causing increase to the network revenue. However, non of these references [1, 2, 3] explain how an ISP can provide his customers the premium service, namely the expedited forwarding service class through different performance levels with different prices. So an ISP can transport real time applications; such as, video and voice traffic, where they receive its restricted e2e QoS, while there is a possibility to transport at the same time non real time applications through other premium service classes. Thus, in this paper, we propose that the highest three DiffServ code point (DSCP ) values, which are already defined in the DiffServ standards to be used by the EF subclasses as shown in Fig. 2.Then, we introduce a mechanism based on the impact of network resources, mainly buffer and link capacity on the QoS offered through EF service class. This impact represents the different performance levels of the different EF subclasses. The network resources are optimally used through traffic engineering (T E) algorithms such as dynamic routing and resource management(DRtM, DRsM ), which are installed at the edge and core routers respectively by developing and employing a Class of Service Based LSP Selection (CoSBLS) algorithm introduced in [4]. Consequently, a customer will have a choice among different levels of EF service, so he will subscribe to the EF service subclass according to its e2e QoS based on the current results of this paper and corresponding price which we will do in the future work.
140 EF2 e2e delay EF1 e2e delay EF3 e2e delay EF4 e2e delay
105
130 100
120 95
EF e2e delay (ms)
EF e2e Delay (ms)
110
90
EF1 e2e Delay (ms) EF2 e2e Delay (ms) EF3 e2e Delay (ms) EF4 e2e Delay (ms)
85
100
90
80
80 75
70 70
60 65
0
50
100
150
200
250
300
350
400
450
500
QueueSize (# Of Packets)
Figure 3: EF1,2,3,4 e2e Delay Bounds Versus Buffer Size Increase.
3.
1 0.45
2 0.9
3 1.35
4 5 1.8 2.25 Bottlenecks Capacity ( Mbps) Assigned BW portion to EF
6 2.7
7 3.15
8 3.6
Figure 4: EF1,2,3,4 e2e Delay Bounds versus the Bottlenecks Capacity.
Buffer Sizing and Capacity Sharing In DiffServ Network
We have conducted some preliminary experiments on the network shown in Fig. 1 using a modified version of UC Berekeley ns-2 simulator to derive simple dimensioning rules to be integrated in the CoSBLS algorithm, in order to satisfy a performance objective per aggregate bandwidth and buffer size. As a result, figures 3 and 4 show us that the buffer size and link capacity at the output interfaces of DiffServ routers have a positive and negative impact on DiffServ network performance. This in turn gives us deep insight in the design of different buffers length and links bandwidth to support the different EF subclasses shown in Fig. 2 along the LSPs which are managed by CoSBLS algorithm. This results in solving the unfairness problem of EF service class, where the customers contracted for the same EF service class, but they do not receive the same service performance as we can conclude from Fig. 4. However, after employing the CosBLS algorithm they almost receive the same performance as we can conclude from Fig 6.
4.
Developing and Simulation CoSBLS Algorithm Results
Although the CoSBLS algorithm and CBLS proposed in [4] are based on the same idea, their purposes are not exactly the same. In [4], the authors applied CBLS algorithm for mapping general traffic classes,however we could reemploy this algorithm after having developed it in a determined DiffServ-aware MPLS TE network topology shown in Fig. 1. The developments are represented by adding two modules to the CBLS algorithm: one to implement the dynamic Dijkstra algorithm in order to generate a minimum hop-count path that can accommodate the required e2e QoS of the different EF service subclasses to coming EF traffic. The second module is to do mapping for the different EF traffic subclasses into their corresponding sub-FECs, then into their convenient LSPs as indicated in the algorithm flowchart shown in Fig. 5. Among the different scenarios that have been realized, we focus on only one scenario due to space limitation. In this scenario, we evaluated the performance of EF service subclasses by serving traffic of EF source1 through EF subclass3 , traffic of EF source2 through EF subclass1 ,traffic of EF source3 through EF subclass2 , and traffic of EF source4 through EF subclass1 . Also, we replaced the traffic sources in Greek site by those of Germany site, while keeping their
LER entery A Packet Received
A packet belongs to AF or BE Traffic
Yes
Mapping AF and BE Flows onto LSPs Perform L3 Switching
EF Packet received
No 100
The Packet belongs to new flow
Yes FEC exists
Yes
Sub-FEC exists 90
80
No
Input Input Output Output Port Label Port Label
No LIBptr
Yes 70
LIBptr EF _default_LIBptr
EFT FEC
Yes PHB
LIBptr
Top_EF_CoSBLSptr
CIB Sub-LSP
LSP with Top_ EF mapping
No Sub-FEC
CoSBLS-LIBptr
EF e2e delay (ms)
LIB
60
50
40
30
20
Yes LIB : Label Information Base EFT : Enhanced Forwarding Table
Enhanced EF_CoSBLSptr
LSP with Enhanced _EF_mapping
CIB : CoSBLS Information Base
0
ER_LIBptr No
Figure 5: CoSBLS Data Base Structure and Forwarding Procedure.
LSP EF1 LSP EF2 LSP EF3 LSP EF4
10
0
1
2
3
4
5
6
LSRs Between LERs
Figure 6: EF1,2,3,4 e2e delay.
destinations connected to their LERs in order to check Whether the algorithm will choose a shortest path or continue to route the EF traffic over the old LSPs. However, we noticed that the EF traffic of sources1,2 was routed through different LSPs traversing the same number of hops. But, EF traffic of source3,4 was routed through new shortest paths. Consequently, Fig. 6 indicates that the e2e delay of EF traffic of source1 is improved remarkably, and the e2e delay of EF traffic of source3 is improved relatively in comparison to the e2e delay of EF traffic of source2 which is not changed as we can conclude from Fig. 4. However, the e2e delay of EF traffic of source4 increases, because its LSP became longer than it was.
5.
Conclusion
We have introduced new premium service classes in MPLS enabled DiffServ network. Hence, this will open new avenues for ISPs to introduce new IP SLAs, which it will contribute highly in increasing their revenues.
References [1] P.Trimintzios,T.Baug,G.Pavlou,P.Flegkas,R.Egan, “Quality of service provisioning through traffic engineering with applicability to IP-based production networks,” Computer Communications 26(8): 845-860 (2003). [2] D.Mitra,K.G.Ramakrishnan, “A case study of multiservice, multipriority traffic engineering design for data networks,” Proc. of IEEE GLOBCOM, Rio de Janeiro, December 1999, PP.1087-1093. [3] D.Mitra,J.A.Morrisom,K.G.Ramakrishnan, “Virtual private networks: joint resource allocation and routing design,” Proc. of IEEE INFOCOM,New York, USA, March 1999. [4] P.Kumar,N. Dhanakoti,S.Gopalan,V.Sridhar, “CoS Based LSP Selection in MPLS Networks,”HSNMC 2004,LCNS 3079,pp.314-323,2004.