- Email: [email protected]
Wavelength and time domain exploitation for QoS management in optical packet switches
Computer Networks 44 (2004) 569–582 www.elsevier.com/locate/comnet
Wavelength and time domain exploitation for QoS management in optical packet switches q F. Callegati *, W. Cerroni, C. Raffaelli, P. Zaffoni DEIS, University of Bologna, Viale Risorgimento 2, 40136 Bologna, Italy
Abstract This paper addresses the problem of congestion resolution and quality of service differentiation in optical packet switching. The paper shows that by designing congestion resolution algorithms that combine the use of the wavelength and the time domain it is possible to significantly reduce information loss phenomena and also to guarantee quality of service differentiation among traffic classes. In particular this is achieved by means of QoS algorithms specifically designed to exploit the characteristics of optical technology. The results are different from the QoS techniques typically implemented in electronic networks. Performance evaluation obtained by simulation shows the influence of the main system parameters on packet loss probability and delay for two service classes. 2003 Elsevier B.V. All rights reserved. Keywords: Optical networks; Wavelength multiplexing; Optical packet switching; Congestion resolution; Quality of service
1. Introduction The combination of all-Optical Packet Switching (OPS) with the very high capacity of DWDM optical fibers promises to be the most powerful and flexible networking technology developed so far. OPS is still a medium-to-long term networking solution because of the still immature state of
q This work is partially funded by the Commission of the European Community, Project IST-1999-11742 ‘‘DAVID–– Data And Voice Integration over DWDM’’ and by the Italian Ministry of Scientific Research, project ‘‘INTREPIDO––Endto-end Traffic Engineering and Protection for IP over DWDM Optical Networks’’. * Corresponding author. E-mail addresses: [email protected] (F. Callegati), [email protected] (W. Cerroni), craff[email protected] (C. Raffaelli), pzaff[email protected] (P. Zaffoni).
technology and of the current difficulties in the telecommunication market, but it represents an important research topic for the scientific community because of its huge potential. Several research projects, such as the UE funded KEOPS and DAVID [1–3], demonstrated the feasibility of switching matrixes able to switch alloptical packet payloads. In these cases synchronous switching operation is considered in order to simplify the switching matrix design. The drawbacks of this approach are the need of all-optical synchronization, implying a substantial increase in hardware complexity, and the inefficient interworking with network protocols based on variable-length packets, such as IP. The adoption of asynchronous and variable-length optical packets overcomes these drawbacks at the cost of a more complex switching matrix design both from the physical and from the logical point of view [4]. In
1389-1286/$ - see front matter 2003 Elsevier B.V. All rights reserved. doi:10.1016/j.comnet.2003.12.010
570
F. Callegati et al. / Computer Networks 44 (2004) 569–582
the last few years, several important studies have analyzed this scenario and investigated the related switching control problem and optical queuing performance [5–8]. In this context congestion resolution becomes one of the most critical issues. Optical memories are not available and congestion resolution in the time domain can usually be implemented by means of Fiber Delay Lines (FDLs), emulating a normal queue by means of lengthy fiber coils used to delay the signal. FDL buffers are not very effective in this scenario, both because of the very limited capacity and because of the mismatch between the variable packet sizes and the fixed length of the FDLs. In spite of these limitations, this paper shows how congestion can be effectively limited to tolerable values by designing suitable congestion resolution algorithms that combine the use of the time and the wavelength domain, and how such algorithms can also provide support for quality of service differentiation. Some heuristic congestion resolution algorithms proposed in previous works for connectionless operation [9] and for connection-oriented operation are reviewed here [10]. Then we propose extensions to these congestion resolution algorithms in order to manage the quality of service in an all-optical packet-switched network scenario, according to the Differentiated Service model [11]. In particular, it has to be taken into account that in an optical network scenario QoS management cannot be performed by means of queuing priorities or fair queuing scheduling algorithms [12], because of the very limited queuing space and, most of all, because queuing by delay lines does not allow a new incoming packet to overtake other packets already queued, a function that is necessary to implement conventional queuing priorities. Under these considerations we aim at showing that QoS differentiation among different traffic classes can be provided with good flexibility and very limited queuing resources, by means of resource reservation both in the time and in the wavelength domain. The paper reviews the overall problem of congestion resolution in OPS with variable length packets in Section 2, providing related results in Section 3. Then in Section 4 the problem of QoS management is addressed and results are presented
in Section 5. In Section 6 some conclusions are drawn.
2. OPS and the congestion resolution problem It is well known that in conventional routing the network layer functions of the routers can be separated into two basic components: control and forwarding. The former element has the task to build up and manage the forwarding table, whereas the latter makes the forwarding decision using the information carried in the packet header. Furthermore, feasible optical packet-switched networks can be realized by introducing an additional partition of the forwarding component into forwarding algorithm and switching. The former allows establishing the correct next-hop destination by exploring the routing tables, while the latter performs the physical transfer of a packet to the proper output interface. In particular, the execution of the forwarding function is carried out in the electronic domain after optical-to-electrical header conversion, whereas the payload is dealt with in the optical domain without any conversion. Acting in this way it is possible to limit electro-optical conversions and achieve better interfacing with optical WDM systems. Taking into account these considerations, a generic architecture of an all-optical packet switch is illustrated in Fig. 1. It is worth outlining that the figure represents a general diagram and will be only used to show the main logical building blocks of the switch architecture. The switch has N inputs and N outputs, each connected to a WDM fiber carrying W wavelengths. The wavelengths of each
W 1
…
N
FDL W
SS
IU
OU
+
IU
1
… OU
N
SCL Fig. 1. Generic architecture of a WDM all-optical packet switch with FDL buffer.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
input fiber are de-multiplexed by an Input Unit (IU), making available the single wavelengths of the bundles to the switching matrix, and are remultiplexed together by an Output Unit (OU) for each output fiber. A non-blocking Space Switch (SS) and a set of fiber delay lines (FDL) are located within an intermediate stage and have the task of performing space switching and queuing respectively. Finally the Switch Control Logic (SCL) makes all the decisions and configures the hardware in order to realize the proper switching actions. We wish to stress that we do not focus on implementation issues, the analysis of the logical behavior of the general switch architecture being the main object of this study. In particular, the functions performed by the SCL for congestion resolution are mainly considered with the aim of exploiting the DWDM domain. The forwarding algorithm represents the basic function that permits delivery of the optical packet to the proper output fiber according to the routing information included in the packet header. After having determined the output fiber, the SCL has to select one of the wavelengths in that fiber in order to properly control the OU. It is important to observe that this is a routing-independent action, since the choice of the output fiber is enough to determine the network path. The selected wavelength may not be immediately available and in the worst case all the wavelengths of a given output fiber may be busy in transmitting other packets at a given instant in time. When this happens the SCL decides whether the packet may be delayed by using the fiber delay lines or if it has to be dropped, because the required delay is not available. Consequently, it is possible to say that the forwarding algorithm is independent of the wavelength and delay selection, while these latter actions are somewhat correlated, the need to delay a packet being related to the availability of the wavelengths. This is what we call the Wavelength and Delay Selection (WDS) problem. For the generic packet i arriving at time ti , after the output fiber is chosen by the forwarding component, the WDS algorithm must choose the output wavelength w 2 ½1 : W and consequently the time tiw when the packet will start being transmitted on it. If wavelength w is currently busy
571
the packet has to be delayed using the FDLs and the choice of the time tiw is not free because the number of delays available is discrete. In the case of B FDLs, only B delays Dj with j 2 ½1 : B are available. Therefore, if packet i needs to be delayed the only possible choices are tiw ¼ ti þ Dj , j 2 ½1 : B. In case tiw should be greater than ti þ DB , the packet cannot be accommodated in the buffer and is lost. For the sake of simplicity, we assume that the available delays are linearly increasing with a fixed rate called D, that is the time unit in which the delays are measured (also called delay granularity). Therefore Dj ¼ kj D, with kj an integer number. Moreover, we also assume a degenerate buffer [13] with kj ¼ j 1 but the ideas here presented are valid also for non-degenerate buffers with different arrangements of kj . Under these assumptions, it is clear that the availability of a discrete number of delays usually do not allow the delayed packet to be immediately transmitted after the end of the previous one and this behavior creates time gaps (also called voids) between queued packets. The effect of these gaps can be considered equivalent to an increase of the packet service time, meaning an artificial increase of the traffic load (excess load) [8], and therefore is very detrimental for the overall system performance. In conclusion, in an undifferentiated networking environment the WDS algorithm is expected to be able to minimize the gaps and maximize the wavelength utilization, in order to maximize the system performance. 2.1. Connectionless operation WDS algorithms for connectionless network operation require a trade-off between performance and computational complexity. In fact, both the forwarding decision and the WDS are applied on a per-packet basis and should be kept very simple in order to avoid bottleneck problems at the network nodes, otherwise causing an unwelcome depletion on the system throughput. Therefore the less the computational complexity, the better the WDS algorithm. A very simple resource allocation scheme for the wavelength selection problem, but designed for
572
F. Callegati et al. / Computer Networks 44 (2004) 569–582
electronic buffers, is called Horizon Channel Scheduling [5]. Our proposal aims at designing effective WDS algorithms, suitable for all-optical buffers, as simple as the Horizon Channel Scheduling. The basic idea is that the algorithm must work simply by employing an array of floating point variables Tw , that is the time when the last bit of the last packet scheduled to wavelength w will leave the node. This information allows the SCL to keep track of when wavelength w will be free for transmitting the next incoming packet arriving at time t0 . Obviously, if Tw t0 6 0, wavelength w is not busy and the packet can be transmitted immediately. Under this idea the whole family of algorithms we present works according to the same basic steps described as follows: 1. Search for the set of wavelengths F such that, for each w 2 F , Tw t0 6 DB ; in case no such wavelength exists all the queues are full and the packet is lost. 2. Select wavelength w 2 F to transmit the packet on. 3. Select delay Dj such that t0 þ Dj P Tw and then select the FDL j 2 ½1 : B to send the packet to. The different behavior of the various algorithms is characterized by the logic used at step 2, which has a strong influence on performance. The three algorithms analyzed work as follows: • Random choice among Neither Empty Nor Full queues (RNENF): Search first for a wavelength with Tw t0 6 0. If no free wavelength is found, choose w randomly excluding full queues (i.e. such that Tw t0 > DB ). In case all wavelengths are full, no choice is made and the packet is lost. • MINimum Length queue (MINL): When a new packet arrives the shortest queue is chosen, that is wavelength w such that Tw 6 Ti for all i 6¼ w. In case two or more queues have the same minimum length, the one that introduces the minimum gap is chosen. This algorithm fully exploits the WDM dimension, but the SCL is more complex than for RNENF case, because it has to compare the values of delay and gap among all queues.
• MINimum Gap queue (MING): Here the choice of the queue introducing the smallest gap between the new packet and the last buffered one is performed by spanning the whole set of non-full wavelengths F . Only in the rare case when choosing two or more wavelengths would give the same value of the gap, the shortest one is chosen. Again, WDM resources are fully exploited and the complexity is similar to the MINL case. In order to better understand how the presented algorithms work, we refer to the situation depicted in Fig. 2, which shows the occupancy of the wavelengths inside the same output fiber for the case W ¼ 4 and B ¼ 5. The RNENF algorithm will choose at random among wavelength k1 , k3 and k4 . The MINL algorithm will choose wavelength k4 (the shortest queue), while the MING algorithm will choose wavelength k1 (which introduces the minimum gap). 2.2. Connection-oriented operation The connection-oriented scenario investigated in this paper is represented by an MPLS-like connection-oriented network [14]. MPLS is a connection-oriented protocol setting up unidirectional connections called Label Switched Paths (LSPs) along which packets are forwarded according to a fixed-length label added to the IP header. In fact, packets are classified into a finite
λ1 λ2 λ3 λ4 t0
t0+D t0+2D t0+3D t0+4D
Fig. 2. A graphical example of the wavelength allocation problem for the case W ¼ 4 and B ¼ 5: a different delay is necessary for any of the wavelengths; how should the choice be made?
F. Callegati et al. / Computer Networks 44 (2004) 569–582
number of subsets, called Forwarding Equivalent Classes (FECs) and packets belonging to the same FEC are forwarded through the same LSP, being indistinguishable from the forwarding point of view. Performing switching operations through simple label processing rather than whole inspection of the IP header satisfies the requirements of a limited processing time. Proposals such as MPkS realize end-to-end wavelength-switched paths mapping LSPs into wavelengths, thus extending these concepts to an optical wavelength-routed network scenario. To achieve maximum flexibility in terms of bandwidth allocation and capacity sharing, the integration of the OPS capabilities in a MPLSover-DWDM network scenario is a critical target. In this environment the WDS algorithm needs to take into account that packets belong to connections (LSP) and that the forwarding table links LSPs to wavelengths on the related output fibers. A possibility is a static LSP-to-wavelength assignment performed at call set-up, which is maintained over the whole connection lifetime, regardless of the congestion state of the wavelengths. On the other hand, the connectionless WDS algorithms working on a per-packet basis may be applied at connection level to design dynamic LSP-to-wavelength assignment techniques. In this case the WDS algorithms are not executed packet by packet, but only when congestion arises, aiming at realizing a trade-off between control complexity and performance. In order to clarify how the nature of the connection-oriented traffic may influence the performance, it is useful to analyze what happens at the input stage of a switching matrix. In Fig. 3 an example with two wavelengths per fiber is depicted. On wavelength k2 three LSPs are active labeled L1, L2 and L3, while only LSP labeled L4 is active on k1 . By observing the figure it is trivial to understand that, whatever the output destination, packets from L1, L2 and L3 will never overlap, while they may overlap with packets from L4. Since congestion in the output queues is due to overlapping arrivals, only packets from LSPs incoming from different input wavelengths will cause congestion. As a consequence, if L1, L2 and L3 are the only LSPs forwarded to the same output wavelength, con-
573
INPUT FIBER
λ1
L4
λ2
L1 L2 L3 t0
t1 t2
t3 t4
Fig. 3. Packets incoming on different LSPs belonging to the same input wavelength do not overlap in time because of the serial nature of the transmission line.
gestion will never arise on that wavelength. This situation will be called optimal allocation. The example shows that, in a connection-oriented environment, performance depends on the forwarding table that affects the behavior of the output queues and it may vary in time following the changes in the forwarding table itself. In order to thoroughly analyze this typical behavior, in [10] an analytical method to quantify the amount of LSPs grouped on the input and on the output was defined. In this paper three different heuristic algorithms are considered to achieve dynamic wavelength allocation, aiming at maximizing the performance. The algorithms can be described as follows: • Round-Robin Wavelength Selection (RRWS): This algorithm performs a very simple roundrobin search for a non-congested wavelength (on the same fiber) each time an incoming packet (following an LSP) has to be delivered to a congested wavelength. If such a wavelength is present, the packet is sent to it and the forwarding table is updated accordingly in order to guarantee that all the subsequent packets on that LSP will follow the same route. The weakness of this approach is the non-optimal allocation strategy based on a round-robin search. Nevertheless, it has been demonstrated in [10] that RRWS is able to significantly improve the performance with respect to a pure static wavelength allocation.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
• Minimum Queue length Wavelength Selection (MQWS): The choice of the output wavelength is made similarly to the MINL scheme adopted for connectionless operation. Therefore, when the wavelength to which an incoming packet has to be sent is congested, MQWS selects among all the non-congested wavelengths the one that is characterized by the shortest waiting queue. Then all packets on the same LSP are rerouted to this wavelength until a new congestion arises. This approach is smarter than RRWS because the search is made to optimize the queuing resources utilization. • Empty Queue Wavelength Selection (EQWS): This algorithm is designed to take into account the nature of the connection-oriented traffic in the sense that in presence of congestion it aims at exploiting also the available queuing space that may be not used by an optimal allocation set-up. In fact, as explained previously, if nonoverlapping incoming packets are delivered to an optimally allocated wavelength the related queuing resource is always empty and it may be exploited to solve congestion phenomenon arising in other FDLs queues. Therefore, when a packet following an LSP arrives and has to be delivered to a wavelength that is congested, EQWS searches for a wavelength with an empty queue. The LSP is switched to that wavelength as long as needed, that is until congestion arises in the new wavelength or congestion disappears from the previous one. In this case the LSP is switched back to the original wavelength.
3. Performance evaluation for undifferentiated services The parameters used in the following are: • • • • •
N , the number of input and output fibers; W , the number of wavelengths per fiber; B, the number of buffer delay lines; D, the delay granularity; q, the wavelength utilization.
The performance of the WDS algorithms for connectionless operation is analyzed first. A com-
parison of the behavior of the different algorithms adopted is made in terms of packet loss probability, showing also the reference case of purely random wavelength selection. Fig. 4 shows the curves of the packet loss rate for all the presented WDS algorithms as a function of the delay granularity with 8 FDLs. As expected, RNENF being aware of the congestion state of the output queues, differently from a pure random choice, allows performance improvement of about two orders of magnitude. By means of even more intelligent algorithms, such as MINL and MING, a further improvement has been obtained. In particular, MING provides better performance than MINL. In spite of what intuition may suggest, the limitation of unused resources due to the minimization of the gaps between packets is more important than the minimization of the queue length. The typical behavior of the curves of the packet loss probability is due to the strong influence of the fiber delay granularity D on the buffer capacity, while the position of the minimum depends on the adopted wavelength selection strategy. In particular, when D 1 the buffering space becomes negligible, the system is almost buffer-less and the four ways of selecting the wavelength are equivalent. When D 1 the gaps inside the buffer are so large that the buffer itself is useless and the system acts as if there was no buffer at all, leading to the 1
Packet Loss Probability
574
10
-1
10
-2
10
-3
10 -4
10
10
RANDOM RNENF MINL MING
-5
-6
0
0.5
1
1.5
2
2.5
3
3.5
4
D
Fig. 4. Packet loss probability as a function of D normalized to the average packet length for different connectionless wavelength allocation algorithms, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
Packet Loss Probability
1
10
B=2 B=4 B=8
-1
10 -2
10
-3
10
-4
10
-5
10 -6
0
10
20
30
40
50
60
W
Fig. 5. Packet loss probability as a function of the number of wavelengths per input/output fiber for the MING algorithm, q ¼ 0:8, N ¼ 4, D ¼ 1 and B ¼ 2, 4, 8.
1.00 RNENF MINL MING
0.90 0.80 0.70
Distribution
same situation as above. The same asymptote for both D ! 0 and D ! 1 is a result of the fact that all the presented algorithms drop a packet only when all the queues are full. Under these assumptions it is also comprehensible that the poor performance provided by RNENF is due to the fact that it does not optimize either the buffer space or the gaps between queued packets. The primary role of the number of wavelengths per input/output fiber with respect to contention resolution is proved by Fig. 5, plotting the packet loss probability given by MING as a function of W for different numbers of buffer delay lines. The figure clarifies how, by combining buffer and wavelength resources, it is possible to solve potential packet contentions, exploiting both the time and wavelength dimensions in order to achieve the required loss rate. As an example, to lose no more than a packet per million an optical buffer with B ¼ 8 delay lines requires at least W ¼ 15 wavelengths per output fiber, while shorter buffers with B ¼ 4 and B ¼ 2 need W ¼ 25 and W ¼ 55 respectively. Another important topic that has to be considered in the design of WDS algorithms is related to the amount of delay that they introduce during switching operations. In effect, it has to be taken into account that some applications are very sensitive to the network latency, whose minimization represents the most important issue to be considered. In Fig. 6 the probability distribution of the
575
0.60 0.50 0.40 0.30 0.20 0.10 0.00 0
1
2
3
4
5
6
7
Delay
Fig. 6. Delay distribution for RNENF, MINL and MING wavelength allocation, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 8 and D optimal.
assigned delay evaluated by simulation is presented for RNENF, MINL and MING for the respective optimal values of D. It is possible to observe that zero is the most frequent delay, which demonstrates a preference for empty queues. The delay distribution clearly shows that MINL and MING tend to keep the queues as short as possible and, although MINL seems to keep the delay lower than MING, the average delay is actually higher, a consequence of the lower packet loss probability of MING. In any case, taking into account the very limited length of the optical buffer, the queuing delay is very small, in particular when compared to propagation delays in a large geographical network. Simulations for connection-oriented operation assume that each input wavelength carries three different LSPs randomly addressed to the output fibers, for a total of 192 incoming LSPs. An important issue in the simulation of a connectionoriented environment is the design of the initial forwarding table. Here the configuration for each input fiber assumes one wavelength out of 16 carrying 3 LSPs (triple) which are directed to the same output, 9 wavelengths with only 2 LSPs (couple) directed to the same output (the third LSP being directed to a different one), while the remaining 6 wavelengths with the 3 LSPs directed to three different outputs. As described in [10], such a distribution of the outputs is the average
576
F. Callegati et al. / Computer Networks 44 (2004) 569–582 1
10
-1
10
-2
Packet Loss Probability
Packet Loss Probability
Static Connectionless RRWS
10-3
10
-1
10
-2
10
-3
10
-4
B=8 B = 16 B = 32
10-5
10 -4 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
D
Fig. 7. Comparison of the RRWS with the static and connectionless case, for N ¼ 4, W ¼ 16, B ¼ 16, q ¼ 0:8.
Fig. 9. Packet loss probability for RRWS as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and varying the number B of delay lines in the buffer.
Packet Loss Probability
1
10
-1
10
-2
10
-3
10
-4
10
-5
10
RRWS MQWS EQWS
the performance by several orders of magnitude with respect to a pure static allocation strategy, approaching performance of the connectionless random case. Fig. 8 shows that EQWS outperforms both RRWS and MQWS, confirming that, when possible, the exploitation of both the unused queuing space and the grouping phenomena are fundamental issues. In the range of optimal values of D the packet loss probability improves significantly with the number B of delay lines, as shown in Fig. 9 for
-6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
Fig. 8. Packet loss probability as a function of the delay unit D for different connection-oriented wavelength allocation algorithms, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8. Packet Loss Probability
distribution in case of uniform traffic, i.e., when no hot spots are considered and all the outputs have the same probability to be addressed. This initial forwarding table will then be changed by the dynamic behavior of the algorithms analyzed, according to the evolution of the arising congestion phenomena. In Fig. 7 a comparison among a static, a random connectionless and a simple dynamic connection-oriented WDS algorithm is shown in terms of packet loss probability as a function of D. It is possible to note that RRWS is able to improve
1
10
-1
10
-2
10
-3
10
-4
10
-5
10 -6
RRWS MQWS EQWS MINL MING
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16
B
Fig. 10. Packet loss probability as a function of the number B of delay lines, for N ¼ 4, W ¼ 16, q ¼ 0:8 and D optimal.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
RRWS and in Fig. 10 for both connectionless and connection-oriented WDS algorithms.
4. QoS management in OPS Techniques to support different levels of QoS in both the connectionless and connection-oriented environments can be defined based on the WDS algorithms presented in Section 2 to provide the best performance. A simple scenario is considered with only two classes of service, which defines QoS differentiation in terms of packet loss probability. In order to meet the target of service differentiation, the algorithm adopted in this paper for connectionless operation is MINL, whereas MQWS is taken as a reference approach for the connectionoriented network scenario. This choice has been made in order to have a benchmark to compare performance of the presented algorithms within both network scenarios. The relative difference between the various algorithms can be deduced from the results for the undifferentiated case that we have presented in the previous section. It is important to outline that QoS management techniques in optical packet switches must be kept very simple due to the delay-oriented characteristics of sequential FDL buffers. In particular, it is not possible to change the order of packets already fed to the delay lines, thus making pre-emption based techniques not applicable. Therefore, mechanisms based on a-priori access control of packets to the optical buffers are necessary [15]. To this end, we intend to exploit the above-mentioned WDS algorithm to differentiate the quality of service by differentiating the amount of choices given to the algorithm. We aim at applying some form of reservation to the resources managed through the WDS algorithm, which are the available wavelengths and delays, in order to privilege one traffic class over the other. We have investigated two alternatives: • Threshold-based technique: The resource reservation is applied to the delay units, and a delay threshold Tlow lower than the maximum delay DB is defined. The WDS algorithm does not allow an incoming low-priority packet to be
577
accepted on a wavelength if the current buffer occupancy is such that the only available delays on that wavelength are greater than or equal to the threshold, i.e. Dj P Tlow , while high-priority packets obviously see the whole buffer capacity. It is expected that the two different classes of traffic suffer different loss rates. In Fig. 11, an example of the QoS management with this technique in presence of four wavelengths per fiber and an optical buffer with 5 FDLs is depicted. The delay threshold Tlow is set equal to 3D. • Wavelength-based technique: The resource reservation is applied to the wavelength domain. Fig. 12 illustrates how the WDS algorithm exploits this procedure to differentiate the performance of low- and high-priority traffic. High-priority packets can be sent to all the wavelengths on a fiber, while low-priority packets are allowed to use only a subset of Wlow wavelengths and, in any case, they share them with high-priority packets. Therefore, low-priority packets are
Low High
Tlow
λ1 λ2 λ3 λ4 t0
t0+D
t0+2D
t0+3D
t0+4D
Fig. 11. A graphical example of the threshold-based technique for the case W ¼ 4, B ¼ 5 and delay threshold Tlow ¼ 3D.
Low High
λ1 λ2 λ3 λ4 Wlow
t0
t0+D
t0+2D t0+3D t0+4D
Fig. 12. A graphical example of the wavelength-based technique for the case W ¼ 4, B ¼ 5 and wavelength subset Wlow ¼ 2.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
expected to suffer higher congestion and typically to experience higher losses and delays than high-priority ones. In the example of Fig. 12 Wlow is set to 2. These two concepts have been applied to the previously explained WDS algorithms in order to achieve service differentiation, leading to the algorithms named MINL-D and MQWS-D, in the case of the threshold-based technique, and MINLLIM and MQWS-LIM, with reference to the wavelength-based technique.
1
Packet Loss Probability
578
10
10
10
10
10
-1
-2
-3
-4
Tlow = 2D
-5
Tlow = 3D 10
-6
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
D
Numerical results are presented here in order to evaluate how the proposed techniques for QoS management are able to provide service differentiation between the two different traffic classes. The performance figures are again the packet loss probability and the packet delay. Results have been obtained by using the same simulation set-up defined in Section 3 to analyze the undifferentiated network scenario. In particular, in the connectionoriented environment the two classes of service are assumed to be associated with LSPs at simulation start up. 5.1. Connectionless operation
Fig. 13. Packet loss probability for MINL-D for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 4, varying the value of the threshold Tlow (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
1.00 0.90
Low
0.80
High
0.70
Distribution
5. Performance evaluation with QoS
0.60 0.50 0.40 0.30 0.20 0.10
Fig. 13 plots the packet loss probability for the MINL-D algorithm for B ¼ 4 and different values of Tlow . The restricted exploitation of the buffer capacity for low-priority packets leads to a high penalty in terms of packet loss probability for the low-priority traffic, while a good behavior is shown for the high-priority traffic. Moreover, performance of each class of service can be straightforwardly tuned by varying the value of the threshold. Fig. 14 shows the delay distribution for low- and high-priority classes for B ¼ 4 and for Tlow ¼ 2D. The high usage of the zero-delay line by both classes is evident. On the other hand, the highest delays are exploited only by the high-priority traffic as a consequence of the algorithm and, in any case, with an extremely limited probability, that is below 106 . In fact, it is very important to
0.00
0
1
2
3
Delay
Fig. 14. Delay distribution for MINL-D for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 4, D ¼ 1:5 and a value of the threshold Tlow ¼ 2D.
note that, for the case of B ¼ 4 considered here, the optical buffer is very short and the queuing delay is always very small, in particular when compared to propagation delays. So, it is possible to state that the application of the threshold-based scheme turns out to be an efficient technique to achieve QoS differentiation, with good values of performance and with a reasonably small processing effort.
F. Callegati et al. / Computer Networks 44 (2004) 569–582 1
579
1.00 0.90
Low
0.80
High
0.70 10 -2
Distribution
Packet Loss Probability
10 -1
10 -3
10
-4
10
-5
0.60 0.50 0.40 0.30 0.20
10 -6
w low = 8 w low = 9 wlow = 10 w low = 12
0.10 0.00 0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
1
2
3
Delay
2
D
Fig. 15. Packet loss probability for MINL–LIM for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 4, varying the number of wavelengths available to low class (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
Fig. 16. Delay distribution for MINL-LIM for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 4, D ¼ 1:3 and a value of the threshold Wlow ¼ 8.
1
10 -1
Packet Loss Probability
In Fig. 15 the packet loss probability for the MING-LIM algorithm with different values of Wlow and B ¼ 4 is plotted. By exploiting the wavelength dimension a sensible differentiation in terms of packet loss can be achieved, allowing a very effective tuning of the order of magnitude of packet loss probability of both classes of service and also providing QoS differentiation with good flexibility. In comparison with Fig. 14, it is interesting to observe the delay distribution plotted in Fig. 16 for Wlow ¼ 8 and B ¼ 4. It can be observed that differentiation is also achieved in terms of delay in the sense that high-priority packets are mostly served with the minimum delay, while lowpriority packets typically suffer more congestion due to the limited wavelength resources available.
10 -2
10 -3
10 -4
10 -5
10 -6 0
Tlow = 5D Tlow = 6D Tlow = 7D 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
Fig. 17. Packet loss probability for MQWS-D for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8, varying the value of the threshold Tlow (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
5.2. Connection-oriented operation The packet loss probability for the MQWS-D algorithm is plotted in Fig. 17, for B ¼ 8 and for different values of the threshold Tlow . As expected, the threshold-based technique is able to provide service differentiation also in a connection-oriented scenario, with a good benefit in terms of packet loss rate for high-priority traffic, achievable by varying the threshold position. Fig. 18 plots the delay dis-
tribution for low- and high-priority classes for B ¼ 8 and for Tlow ¼ 5D. Both classes of service experience high usage of the lowest value of delay, while it is easy to observe the different exploitation of the buffer, in terms of delay distribution, by the two classes of service, as a consequence of the threshold value. On one side the delay for lowpriority packets is limited by the threshold value, while over this value the delay probability for
580
F. Callegati et al. / Computer Networks 44 (2004) 569–582 1.00
1.00
0.90
low
0.90
low
0.80
high
0.80
high
0.70
Distribution
Distribution
0.70 0.60 0.50 0.40 0.30
0.60 0.50 0.40 0.30
0.20
0.20
0.10
0.10
0.00
0.00 0
1
2
3
4
5
6
7
0
Delay
high-priority packets decreases with the delay value due to the lower usage of the buffer positions reserved to high-priority traffic only. In Fig. 19 the performance of MQWS-LIM is presented for B ¼ 8, varying the value of Wlow . As in the connectionless case, the figure shows that in the wavelength domain a sensible service dif-
Packet Loss Probability
1
-1
10
-2
10
-3
10
-4
10
-5
10
-6
0
w low = 4 w low = 6 w low = 8 w low = 10 0.1
0.2
0.3
0.4
0.5
2
3
4
5
6
7
Delay
Fig. 18. Delay distribution for MQWS-D for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 8, D ¼ 0:4 and a value of the threshold Tlow ¼ 5D.
10
1
0.6
0.7
0.8
0.9
Fig. 20. Delay distribution for MQWS-LIM for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 8, D ¼ 0:4 and a value of the threshold Wlow ¼ 8.
ferentiation level can be achieved in terms of packet loss by effectively tuning the performance of both classes of service through the parameter Wlow . Fig. 20 shows the delay distribution for lowand high-priority classes for Wlow ¼ 8 and B ¼ 8. The delay distribution behaves in the same way for both classes, exhibiting the highest probability value for delay equal to zero. This is due to the fact that the two classes are characterized by the same load that is split almost equally on the two subsets of 8 wavelengths each. When queues are significantly not full, the algorithm tends to isolate the two subsets of wavelengths, devoting one to the low-priority packets plus a small fraction of high-priority packets. This causes this subset to be loaded by a little more than half of the total traffic and leads to a bit greater delay probability with respect to high-priority traffic. In fact, highpriority LSPs find the wavelengths used by lowpriority LSPs less convenient in terms of queue occupancy and tend to use only their exclusive wavelengths.
1
D
Fig. 19. Packet loss probability for MQWS-LIM for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8, varying the number of wavelengths available to low class (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
6. Conclusions Congestion resolution algorithms able to support quality of service in DWDM optical packetswitched networks have been considered and analyzed in both connectionless and connection-
F. Callegati et al. / Computer Networks 44 (2004) 569–582
oriented contexts. Good differentiation and QoS tuning capabilities with two classes of service have been shown, although performance of the two classes is correlated. QoS differentiation is achieved by means of two techniques, the former reserving resources to the high priority traffic in the delay queue (threshold-based) the latter reserving resources in the wavelength domain (wavelengthbased). The threshold-based technique achieves good discrimination between the two classes in terms of packet loss while naturally limiting the delay of the low-priority class that, on the other hand, suffers the highest loss. The wavelengthbased technique achieves similar performance and is also more effective for finer QoS tuning. Some effects of the algorithms on delay distribution have also been shown and discussed.
581
[10] F. Callegati, W. Cerroni, C. Raffaelli, P. Zaffoni, Dynamic wavelength assignment in MPLS optical packet switches, Optical Networks Magazine 4 (5) (2003) 41– 51. [11] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, Z.W. Weiss, An architecture for differentiated services, IETF RFC 2475, December 1998. [12] D.C. Stephens, J.C.R. Bennett, H. Zhang, Implementing scheduling algorithms in high-speed networks, IEEE Journal on Selected Areas in Communications 17 (6) (1999) 1145–1158. [13] L. Tancevski, L.S. Tamil, F. Callegati, Non-degenerate buffers: a paradigm for building large optical memories, IEEE Photonic Technology Letters 11 (8) (1999) 1072– 1074. [14] E. Rosen, A. Viswanathan, R. Callon, Multiprotocol label switching architecture, IETF RFC 3031, January 2001. [15] F. Callegati, G. Corazza, C. Raffaelli, Exploitation of DWDM for optical packet switching with quality of service guarantees, IEEE Journal on Selected Areas in Communications 20 (1) (2002) 190–201.
References [1] P. Gambini et al., Transparent optical packet switching: network architecture and demonstrators in the KEOPS project, IEEE Journal on Selected Areas in Communications 16 (7) (1998) 1245–1259. [2] C. Guillemot et al., Transparent optical packet switching: The European ACTS KEOPS project approach, IEEE/ OSA Journal of Lightwave Technology 16 (12) (1998) 2117–2134. [3] L. Dittmann et al., The European IST project DAVID: A viable approach towards optical packet switching, IEEE Journal on Selected Areas in Communications 21 (7) (2003) 1026–1040. [4] D. Chiaroni et al., First demonstration of an asynchronous optical packet switching matrix prototype for multiterabitclass routers/switches, in: Proceedings of 27th European Conference on Optical Communication (ECOC 2001), Amsterdam, The Netherlands, October 2001. [5] J. Turner, Terabit burst switching, Journal of High Speed Networks 8 (1) (1999) 3–16. [6] F. Callegati, H.C. Cankaya, Y. Xiong, M. Vandenhoute, Design issues for optical IP routers, IEEE Communications Magazine 37 (12) (1999) 124–128. [7] L. Tancevski, S. Yegnanarayanan, G. Casta~ non, L. Tamil, F. Masetti, T. McDermott, Optical routing of asynchronous, variable length packets, IEEE Journal on Selected Areas in Communications 18 (10) (2000) 2084– 2093. [8] F. Callegati, Optical buffers for variable length packets, IEEE Communications Letters 4 (9) (2000) 292– 294. [9] F. Callegati, W. Cerroni, G. Corazza, Optimization of wavelength allocation in WDM optical buffers, Optical Networks Magazine 2 (6) (2001) 66–72.
F. Callegati received his Master and Ph.D. in Electrical Engineering in 1989 and 1992 from the University of Bologna, Italy. In 1993 he was research scientist at the Teletraffic Research Centre of the University of Adelaide, Australia and in 1994 at Fondazione U. Bordoni, Italy. He has been with DEIS at the University of Bologna since 1995, where he is now serving as Associate Professor. His research interests are in the field of traffic modeling and performance evaluation of telecommunication networks. He is currently working in the field of optical packet switching and optical networking for Internet traffic. He has been active in UE funded research projects on Optical Packet Switching such ACTS KEOPS and IST DAVID.
W. Cerroni received the Master degree in Telecommunication Engineering and the Ph.D. in Electrical and Computer Engineering both from the University of Bologna, Italy, in 1999 and 2003 respectively. He is currently holding a post-doc position at the University of Bologna. During 1999 he was a visitor researcher at the University of Texas at Dallas (UTD) working in cooperation with Alcatel Corporate Research Center, Richardson (Texas), USA. His research interests cover traffic performance evaluation of communication networks and architectures for very high-speed optical packet switching. He contributed to some of the activities within the EU-funded IST DAVID (Data And Voice Integration over DWDM) project as well as to national research projects on optical packet switching and grid computing infrastructures.
582
F. Callegati et al. / Computer Networks 44 (2004) 569–582
C. Raffaelli was born in Bologna, Italy, in 1960. She received the Dr. Eng. Degree in Electronic Engineering from the University of Bologna in 1985 and the Ph.D. Degree in Electronic Engineering and Computer Science in 1990. Since 1985 she has been with the Department of Electronics, Computer Science and Systems of the University of Bologna, where she became a Research Associate in 1990. Since then her research activity was concerned with broadband communications, and in particular with performance and implementation of ATM switching architectures. She has been involved in many National and International research projects on high-speed networks among which the KEOPS project of the European Union on photonic transport of information. She is currently involved in National projects on high-speed TCP/IP transport and in the DAVID IST project of the European Union (Data And Voice Integration over DWDM). She is
Associate Professor in Switching Systems and Telecommunication Networks at the University of Bologna. P. Zaffoni received the Master degree in Telecommunication Engineering from the University of Bologna, Italy, in 2001. He is currently pursuing the Ph.D. degree in Electrical and Computer Engineering at the University of Bologna. He is working in the field of traffic performance and architectures for broadband communications networks. In particular his research topics include optical packet switching, optical networking and TCP performance in heterogeneous networks scenarios. He is currently involved in National projects on optical network architectures.
Wavelength and time domain exploitation for QoS management in optical packet switches q F. Callegati *, W. Cerroni, C. Raffaelli, P. Zaffoni DEIS, University of Bologna, Viale Risorgimento 2, 40136 Bologna, Italy
Abstract This paper addresses the problem of congestion resolution and quality of service differentiation in optical packet switching. The paper shows that by designing congestion resolution algorithms that combine the use of the wavelength and the time domain it is possible to significantly reduce information loss phenomena and also to guarantee quality of service differentiation among traffic classes. In particular this is achieved by means of QoS algorithms specifically designed to exploit the characteristics of optical technology. The results are different from the QoS techniques typically implemented in electronic networks. Performance evaluation obtained by simulation shows the influence of the main system parameters on packet loss probability and delay for two service classes. 2003 Elsevier B.V. All rights reserved. Keywords: Optical networks; Wavelength multiplexing; Optical packet switching; Congestion resolution; Quality of service
1. Introduction The combination of all-Optical Packet Switching (OPS) with the very high capacity of DWDM optical fibers promises to be the most powerful and flexible networking technology developed so far. OPS is still a medium-to-long term networking solution because of the still immature state of
q This work is partially funded by the Commission of the European Community, Project IST-1999-11742 ‘‘DAVID–– Data And Voice Integration over DWDM’’ and by the Italian Ministry of Scientific Research, project ‘‘INTREPIDO––Endto-end Traffic Engineering and Protection for IP over DWDM Optical Networks’’. * Corresponding author. E-mail addresses: [email protected] (F. Callegati), [email protected] (W. Cerroni), craff[email protected] (C. Raffaelli), pzaff[email protected] (P. Zaffoni).
technology and of the current difficulties in the telecommunication market, but it represents an important research topic for the scientific community because of its huge potential. Several research projects, such as the UE funded KEOPS and DAVID [1–3], demonstrated the feasibility of switching matrixes able to switch alloptical packet payloads. In these cases synchronous switching operation is considered in order to simplify the switching matrix design. The drawbacks of this approach are the need of all-optical synchronization, implying a substantial increase in hardware complexity, and the inefficient interworking with network protocols based on variable-length packets, such as IP. The adoption of asynchronous and variable-length optical packets overcomes these drawbacks at the cost of a more complex switching matrix design both from the physical and from the logical point of view [4]. In
1389-1286/$ - see front matter 2003 Elsevier B.V. All rights reserved. doi:10.1016/j.comnet.2003.12.010
570
F. Callegati et al. / Computer Networks 44 (2004) 569–582
the last few years, several important studies have analyzed this scenario and investigated the related switching control problem and optical queuing performance [5–8]. In this context congestion resolution becomes one of the most critical issues. Optical memories are not available and congestion resolution in the time domain can usually be implemented by means of Fiber Delay Lines (FDLs), emulating a normal queue by means of lengthy fiber coils used to delay the signal. FDL buffers are not very effective in this scenario, both because of the very limited capacity and because of the mismatch between the variable packet sizes and the fixed length of the FDLs. In spite of these limitations, this paper shows how congestion can be effectively limited to tolerable values by designing suitable congestion resolution algorithms that combine the use of the time and the wavelength domain, and how such algorithms can also provide support for quality of service differentiation. Some heuristic congestion resolution algorithms proposed in previous works for connectionless operation [9] and for connection-oriented operation are reviewed here [10]. Then we propose extensions to these congestion resolution algorithms in order to manage the quality of service in an all-optical packet-switched network scenario, according to the Differentiated Service model [11]. In particular, it has to be taken into account that in an optical network scenario QoS management cannot be performed by means of queuing priorities or fair queuing scheduling algorithms [12], because of the very limited queuing space and, most of all, because queuing by delay lines does not allow a new incoming packet to overtake other packets already queued, a function that is necessary to implement conventional queuing priorities. Under these considerations we aim at showing that QoS differentiation among different traffic classes can be provided with good flexibility and very limited queuing resources, by means of resource reservation both in the time and in the wavelength domain. The paper reviews the overall problem of congestion resolution in OPS with variable length packets in Section 2, providing related results in Section 3. Then in Section 4 the problem of QoS management is addressed and results are presented
in Section 5. In Section 6 some conclusions are drawn.
2. OPS and the congestion resolution problem It is well known that in conventional routing the network layer functions of the routers can be separated into two basic components: control and forwarding. The former element has the task to build up and manage the forwarding table, whereas the latter makes the forwarding decision using the information carried in the packet header. Furthermore, feasible optical packet-switched networks can be realized by introducing an additional partition of the forwarding component into forwarding algorithm and switching. The former allows establishing the correct next-hop destination by exploring the routing tables, while the latter performs the physical transfer of a packet to the proper output interface. In particular, the execution of the forwarding function is carried out in the electronic domain after optical-to-electrical header conversion, whereas the payload is dealt with in the optical domain without any conversion. Acting in this way it is possible to limit electro-optical conversions and achieve better interfacing with optical WDM systems. Taking into account these considerations, a generic architecture of an all-optical packet switch is illustrated in Fig. 1. It is worth outlining that the figure represents a general diagram and will be only used to show the main logical building blocks of the switch architecture. The switch has N inputs and N outputs, each connected to a WDM fiber carrying W wavelengths. The wavelengths of each
W 1
…
N
FDL W
SS
IU
OU
+
IU
1
… OU
N
SCL Fig. 1. Generic architecture of a WDM all-optical packet switch with FDL buffer.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
input fiber are de-multiplexed by an Input Unit (IU), making available the single wavelengths of the bundles to the switching matrix, and are remultiplexed together by an Output Unit (OU) for each output fiber. A non-blocking Space Switch (SS) and a set of fiber delay lines (FDL) are located within an intermediate stage and have the task of performing space switching and queuing respectively. Finally the Switch Control Logic (SCL) makes all the decisions and configures the hardware in order to realize the proper switching actions. We wish to stress that we do not focus on implementation issues, the analysis of the logical behavior of the general switch architecture being the main object of this study. In particular, the functions performed by the SCL for congestion resolution are mainly considered with the aim of exploiting the DWDM domain. The forwarding algorithm represents the basic function that permits delivery of the optical packet to the proper output fiber according to the routing information included in the packet header. After having determined the output fiber, the SCL has to select one of the wavelengths in that fiber in order to properly control the OU. It is important to observe that this is a routing-independent action, since the choice of the output fiber is enough to determine the network path. The selected wavelength may not be immediately available and in the worst case all the wavelengths of a given output fiber may be busy in transmitting other packets at a given instant in time. When this happens the SCL decides whether the packet may be delayed by using the fiber delay lines or if it has to be dropped, because the required delay is not available. Consequently, it is possible to say that the forwarding algorithm is independent of the wavelength and delay selection, while these latter actions are somewhat correlated, the need to delay a packet being related to the availability of the wavelengths. This is what we call the Wavelength and Delay Selection (WDS) problem. For the generic packet i arriving at time ti , after the output fiber is chosen by the forwarding component, the WDS algorithm must choose the output wavelength w 2 ½1 : W and consequently the time tiw when the packet will start being transmitted on it. If wavelength w is currently busy
571
the packet has to be delayed using the FDLs and the choice of the time tiw is not free because the number of delays available is discrete. In the case of B FDLs, only B delays Dj with j 2 ½1 : B are available. Therefore, if packet i needs to be delayed the only possible choices are tiw ¼ ti þ Dj , j 2 ½1 : B. In case tiw should be greater than ti þ DB , the packet cannot be accommodated in the buffer and is lost. For the sake of simplicity, we assume that the available delays are linearly increasing with a fixed rate called D, that is the time unit in which the delays are measured (also called delay granularity). Therefore Dj ¼ kj D, with kj an integer number. Moreover, we also assume a degenerate buffer [13] with kj ¼ j 1 but the ideas here presented are valid also for non-degenerate buffers with different arrangements of kj . Under these assumptions, it is clear that the availability of a discrete number of delays usually do not allow the delayed packet to be immediately transmitted after the end of the previous one and this behavior creates time gaps (also called voids) between queued packets. The effect of these gaps can be considered equivalent to an increase of the packet service time, meaning an artificial increase of the traffic load (excess load) [8], and therefore is very detrimental for the overall system performance. In conclusion, in an undifferentiated networking environment the WDS algorithm is expected to be able to minimize the gaps and maximize the wavelength utilization, in order to maximize the system performance. 2.1. Connectionless operation WDS algorithms for connectionless network operation require a trade-off between performance and computational complexity. In fact, both the forwarding decision and the WDS are applied on a per-packet basis and should be kept very simple in order to avoid bottleneck problems at the network nodes, otherwise causing an unwelcome depletion on the system throughput. Therefore the less the computational complexity, the better the WDS algorithm. A very simple resource allocation scheme for the wavelength selection problem, but designed for
572
F. Callegati et al. / Computer Networks 44 (2004) 569–582
electronic buffers, is called Horizon Channel Scheduling [5]. Our proposal aims at designing effective WDS algorithms, suitable for all-optical buffers, as simple as the Horizon Channel Scheduling. The basic idea is that the algorithm must work simply by employing an array of floating point variables Tw , that is the time when the last bit of the last packet scheduled to wavelength w will leave the node. This information allows the SCL to keep track of when wavelength w will be free for transmitting the next incoming packet arriving at time t0 . Obviously, if Tw t0 6 0, wavelength w is not busy and the packet can be transmitted immediately. Under this idea the whole family of algorithms we present works according to the same basic steps described as follows: 1. Search for the set of wavelengths F such that, for each w 2 F , Tw t0 6 DB ; in case no such wavelength exists all the queues are full and the packet is lost. 2. Select wavelength w 2 F to transmit the packet on. 3. Select delay Dj such that t0 þ Dj P Tw and then select the FDL j 2 ½1 : B to send the packet to. The different behavior of the various algorithms is characterized by the logic used at step 2, which has a strong influence on performance. The three algorithms analyzed work as follows: • Random choice among Neither Empty Nor Full queues (RNENF): Search first for a wavelength with Tw t0 6 0. If no free wavelength is found, choose w randomly excluding full queues (i.e. such that Tw t0 > DB ). In case all wavelengths are full, no choice is made and the packet is lost. • MINimum Length queue (MINL): When a new packet arrives the shortest queue is chosen, that is wavelength w such that Tw 6 Ti for all i 6¼ w. In case two or more queues have the same minimum length, the one that introduces the minimum gap is chosen. This algorithm fully exploits the WDM dimension, but the SCL is more complex than for RNENF case, because it has to compare the values of delay and gap among all queues.
• MINimum Gap queue (MING): Here the choice of the queue introducing the smallest gap between the new packet and the last buffered one is performed by spanning the whole set of non-full wavelengths F . Only in the rare case when choosing two or more wavelengths would give the same value of the gap, the shortest one is chosen. Again, WDM resources are fully exploited and the complexity is similar to the MINL case. In order to better understand how the presented algorithms work, we refer to the situation depicted in Fig. 2, which shows the occupancy of the wavelengths inside the same output fiber for the case W ¼ 4 and B ¼ 5. The RNENF algorithm will choose at random among wavelength k1 , k3 and k4 . The MINL algorithm will choose wavelength k4 (the shortest queue), while the MING algorithm will choose wavelength k1 (which introduces the minimum gap). 2.2. Connection-oriented operation The connection-oriented scenario investigated in this paper is represented by an MPLS-like connection-oriented network [14]. MPLS is a connection-oriented protocol setting up unidirectional connections called Label Switched Paths (LSPs) along which packets are forwarded according to a fixed-length label added to the IP header. In fact, packets are classified into a finite
λ1 λ2 λ3 λ4 t0
t0+D t0+2D t0+3D t0+4D
Fig. 2. A graphical example of the wavelength allocation problem for the case W ¼ 4 and B ¼ 5: a different delay is necessary for any of the wavelengths; how should the choice be made?
F. Callegati et al. / Computer Networks 44 (2004) 569–582
number of subsets, called Forwarding Equivalent Classes (FECs) and packets belonging to the same FEC are forwarded through the same LSP, being indistinguishable from the forwarding point of view. Performing switching operations through simple label processing rather than whole inspection of the IP header satisfies the requirements of a limited processing time. Proposals such as MPkS realize end-to-end wavelength-switched paths mapping LSPs into wavelengths, thus extending these concepts to an optical wavelength-routed network scenario. To achieve maximum flexibility in terms of bandwidth allocation and capacity sharing, the integration of the OPS capabilities in a MPLSover-DWDM network scenario is a critical target. In this environment the WDS algorithm needs to take into account that packets belong to connections (LSP) and that the forwarding table links LSPs to wavelengths on the related output fibers. A possibility is a static LSP-to-wavelength assignment performed at call set-up, which is maintained over the whole connection lifetime, regardless of the congestion state of the wavelengths. On the other hand, the connectionless WDS algorithms working on a per-packet basis may be applied at connection level to design dynamic LSP-to-wavelength assignment techniques. In this case the WDS algorithms are not executed packet by packet, but only when congestion arises, aiming at realizing a trade-off between control complexity and performance. In order to clarify how the nature of the connection-oriented traffic may influence the performance, it is useful to analyze what happens at the input stage of a switching matrix. In Fig. 3 an example with two wavelengths per fiber is depicted. On wavelength k2 three LSPs are active labeled L1, L2 and L3, while only LSP labeled L4 is active on k1 . By observing the figure it is trivial to understand that, whatever the output destination, packets from L1, L2 and L3 will never overlap, while they may overlap with packets from L4. Since congestion in the output queues is due to overlapping arrivals, only packets from LSPs incoming from different input wavelengths will cause congestion. As a consequence, if L1, L2 and L3 are the only LSPs forwarded to the same output wavelength, con-
573
INPUT FIBER
λ1
L4
λ2
L1 L2 L3 t0
t1 t2
t3 t4
Fig. 3. Packets incoming on different LSPs belonging to the same input wavelength do not overlap in time because of the serial nature of the transmission line.
gestion will never arise on that wavelength. This situation will be called optimal allocation. The example shows that, in a connection-oriented environment, performance depends on the forwarding table that affects the behavior of the output queues and it may vary in time following the changes in the forwarding table itself. In order to thoroughly analyze this typical behavior, in [10] an analytical method to quantify the amount of LSPs grouped on the input and on the output was defined. In this paper three different heuristic algorithms are considered to achieve dynamic wavelength allocation, aiming at maximizing the performance. The algorithms can be described as follows: • Round-Robin Wavelength Selection (RRWS): This algorithm performs a very simple roundrobin search for a non-congested wavelength (on the same fiber) each time an incoming packet (following an LSP) has to be delivered to a congested wavelength. If such a wavelength is present, the packet is sent to it and the forwarding table is updated accordingly in order to guarantee that all the subsequent packets on that LSP will follow the same route. The weakness of this approach is the non-optimal allocation strategy based on a round-robin search. Nevertheless, it has been demonstrated in [10] that RRWS is able to significantly improve the performance with respect to a pure static wavelength allocation.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
• Minimum Queue length Wavelength Selection (MQWS): The choice of the output wavelength is made similarly to the MINL scheme adopted for connectionless operation. Therefore, when the wavelength to which an incoming packet has to be sent is congested, MQWS selects among all the non-congested wavelengths the one that is characterized by the shortest waiting queue. Then all packets on the same LSP are rerouted to this wavelength until a new congestion arises. This approach is smarter than RRWS because the search is made to optimize the queuing resources utilization. • Empty Queue Wavelength Selection (EQWS): This algorithm is designed to take into account the nature of the connection-oriented traffic in the sense that in presence of congestion it aims at exploiting also the available queuing space that may be not used by an optimal allocation set-up. In fact, as explained previously, if nonoverlapping incoming packets are delivered to an optimally allocated wavelength the related queuing resource is always empty and it may be exploited to solve congestion phenomenon arising in other FDLs queues. Therefore, when a packet following an LSP arrives and has to be delivered to a wavelength that is congested, EQWS searches for a wavelength with an empty queue. The LSP is switched to that wavelength as long as needed, that is until congestion arises in the new wavelength or congestion disappears from the previous one. In this case the LSP is switched back to the original wavelength.
3. Performance evaluation for undifferentiated services The parameters used in the following are: • • • • •
N , the number of input and output fibers; W , the number of wavelengths per fiber; B, the number of buffer delay lines; D, the delay granularity; q, the wavelength utilization.
The performance of the WDS algorithms for connectionless operation is analyzed first. A com-
parison of the behavior of the different algorithms adopted is made in terms of packet loss probability, showing also the reference case of purely random wavelength selection. Fig. 4 shows the curves of the packet loss rate for all the presented WDS algorithms as a function of the delay granularity with 8 FDLs. As expected, RNENF being aware of the congestion state of the output queues, differently from a pure random choice, allows performance improvement of about two orders of magnitude. By means of even more intelligent algorithms, such as MINL and MING, a further improvement has been obtained. In particular, MING provides better performance than MINL. In spite of what intuition may suggest, the limitation of unused resources due to the minimization of the gaps between packets is more important than the minimization of the queue length. The typical behavior of the curves of the packet loss probability is due to the strong influence of the fiber delay granularity D on the buffer capacity, while the position of the minimum depends on the adopted wavelength selection strategy. In particular, when D 1 the buffering space becomes negligible, the system is almost buffer-less and the four ways of selecting the wavelength are equivalent. When D 1 the gaps inside the buffer are so large that the buffer itself is useless and the system acts as if there was no buffer at all, leading to the 1
Packet Loss Probability
574
10
-1
10
-2
10
-3
10 -4
10
10
RANDOM RNENF MINL MING
-5
-6
0
0.5
1
1.5
2
2.5
3
3.5
4
D
Fig. 4. Packet loss probability as a function of D normalized to the average packet length for different connectionless wavelength allocation algorithms, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
Packet Loss Probability
1
10
B=2 B=4 B=8
-1
10 -2
10
-3
10
-4
10
-5
10 -6
0
10
20
30
40
50
60
W
Fig. 5. Packet loss probability as a function of the number of wavelengths per input/output fiber for the MING algorithm, q ¼ 0:8, N ¼ 4, D ¼ 1 and B ¼ 2, 4, 8.
1.00 RNENF MINL MING
0.90 0.80 0.70
Distribution
same situation as above. The same asymptote for both D ! 0 and D ! 1 is a result of the fact that all the presented algorithms drop a packet only when all the queues are full. Under these assumptions it is also comprehensible that the poor performance provided by RNENF is due to the fact that it does not optimize either the buffer space or the gaps between queued packets. The primary role of the number of wavelengths per input/output fiber with respect to contention resolution is proved by Fig. 5, plotting the packet loss probability given by MING as a function of W for different numbers of buffer delay lines. The figure clarifies how, by combining buffer and wavelength resources, it is possible to solve potential packet contentions, exploiting both the time and wavelength dimensions in order to achieve the required loss rate. As an example, to lose no more than a packet per million an optical buffer with B ¼ 8 delay lines requires at least W ¼ 15 wavelengths per output fiber, while shorter buffers with B ¼ 4 and B ¼ 2 need W ¼ 25 and W ¼ 55 respectively. Another important topic that has to be considered in the design of WDS algorithms is related to the amount of delay that they introduce during switching operations. In effect, it has to be taken into account that some applications are very sensitive to the network latency, whose minimization represents the most important issue to be considered. In Fig. 6 the probability distribution of the
575
0.60 0.50 0.40 0.30 0.20 0.10 0.00 0
1
2
3
4
5
6
7
Delay
Fig. 6. Delay distribution for RNENF, MINL and MING wavelength allocation, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 8 and D optimal.
assigned delay evaluated by simulation is presented for RNENF, MINL and MING for the respective optimal values of D. It is possible to observe that zero is the most frequent delay, which demonstrates a preference for empty queues. The delay distribution clearly shows that MINL and MING tend to keep the queues as short as possible and, although MINL seems to keep the delay lower than MING, the average delay is actually higher, a consequence of the lower packet loss probability of MING. In any case, taking into account the very limited length of the optical buffer, the queuing delay is very small, in particular when compared to propagation delays in a large geographical network. Simulations for connection-oriented operation assume that each input wavelength carries three different LSPs randomly addressed to the output fibers, for a total of 192 incoming LSPs. An important issue in the simulation of a connectionoriented environment is the design of the initial forwarding table. Here the configuration for each input fiber assumes one wavelength out of 16 carrying 3 LSPs (triple) which are directed to the same output, 9 wavelengths with only 2 LSPs (couple) directed to the same output (the third LSP being directed to a different one), while the remaining 6 wavelengths with the 3 LSPs directed to three different outputs. As described in [10], such a distribution of the outputs is the average
576
F. Callegati et al. / Computer Networks 44 (2004) 569–582 1
10
-1
10
-2
Packet Loss Probability
Packet Loss Probability
Static Connectionless RRWS
10-3
10
-1
10
-2
10
-3
10
-4
B=8 B = 16 B = 32
10-5
10 -4 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
D
Fig. 7. Comparison of the RRWS with the static and connectionless case, for N ¼ 4, W ¼ 16, B ¼ 16, q ¼ 0:8.
Fig. 9. Packet loss probability for RRWS as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and varying the number B of delay lines in the buffer.
Packet Loss Probability
1
10
-1
10
-2
10
-3
10
-4
10
-5
10
RRWS MQWS EQWS
the performance by several orders of magnitude with respect to a pure static allocation strategy, approaching performance of the connectionless random case. Fig. 8 shows that EQWS outperforms both RRWS and MQWS, confirming that, when possible, the exploitation of both the unused queuing space and the grouping phenomena are fundamental issues. In the range of optimal values of D the packet loss probability improves significantly with the number B of delay lines, as shown in Fig. 9 for
-6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
Fig. 8. Packet loss probability as a function of the delay unit D for different connection-oriented wavelength allocation algorithms, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8. Packet Loss Probability
distribution in case of uniform traffic, i.e., when no hot spots are considered and all the outputs have the same probability to be addressed. This initial forwarding table will then be changed by the dynamic behavior of the algorithms analyzed, according to the evolution of the arising congestion phenomena. In Fig. 7 a comparison among a static, a random connectionless and a simple dynamic connection-oriented WDS algorithm is shown in terms of packet loss probability as a function of D. It is possible to note that RRWS is able to improve
1
10
-1
10
-2
10
-3
10
-4
10
-5
10 -6
RRWS MQWS EQWS MINL MING
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16
B
Fig. 10. Packet loss probability as a function of the number B of delay lines, for N ¼ 4, W ¼ 16, q ¼ 0:8 and D optimal.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
RRWS and in Fig. 10 for both connectionless and connection-oriented WDS algorithms.
4. QoS management in OPS Techniques to support different levels of QoS in both the connectionless and connection-oriented environments can be defined based on the WDS algorithms presented in Section 2 to provide the best performance. A simple scenario is considered with only two classes of service, which defines QoS differentiation in terms of packet loss probability. In order to meet the target of service differentiation, the algorithm adopted in this paper for connectionless operation is MINL, whereas MQWS is taken as a reference approach for the connectionoriented network scenario. This choice has been made in order to have a benchmark to compare performance of the presented algorithms within both network scenarios. The relative difference between the various algorithms can be deduced from the results for the undifferentiated case that we have presented in the previous section. It is important to outline that QoS management techniques in optical packet switches must be kept very simple due to the delay-oriented characteristics of sequential FDL buffers. In particular, it is not possible to change the order of packets already fed to the delay lines, thus making pre-emption based techniques not applicable. Therefore, mechanisms based on a-priori access control of packets to the optical buffers are necessary [15]. To this end, we intend to exploit the above-mentioned WDS algorithm to differentiate the quality of service by differentiating the amount of choices given to the algorithm. We aim at applying some form of reservation to the resources managed through the WDS algorithm, which are the available wavelengths and delays, in order to privilege one traffic class over the other. We have investigated two alternatives: • Threshold-based technique: The resource reservation is applied to the delay units, and a delay threshold Tlow lower than the maximum delay DB is defined. The WDS algorithm does not allow an incoming low-priority packet to be
577
accepted on a wavelength if the current buffer occupancy is such that the only available delays on that wavelength are greater than or equal to the threshold, i.e. Dj P Tlow , while high-priority packets obviously see the whole buffer capacity. It is expected that the two different classes of traffic suffer different loss rates. In Fig. 11, an example of the QoS management with this technique in presence of four wavelengths per fiber and an optical buffer with 5 FDLs is depicted. The delay threshold Tlow is set equal to 3D. • Wavelength-based technique: The resource reservation is applied to the wavelength domain. Fig. 12 illustrates how the WDS algorithm exploits this procedure to differentiate the performance of low- and high-priority traffic. High-priority packets can be sent to all the wavelengths on a fiber, while low-priority packets are allowed to use only a subset of Wlow wavelengths and, in any case, they share them with high-priority packets. Therefore, low-priority packets are
Low High
Tlow
λ1 λ2 λ3 λ4 t0
t0+D
t0+2D
t0+3D
t0+4D
Fig. 11. A graphical example of the threshold-based technique for the case W ¼ 4, B ¼ 5 and delay threshold Tlow ¼ 3D.
Low High
λ1 λ2 λ3 λ4 Wlow
t0
t0+D
t0+2D t0+3D t0+4D
Fig. 12. A graphical example of the wavelength-based technique for the case W ¼ 4, B ¼ 5 and wavelength subset Wlow ¼ 2.
F. Callegati et al. / Computer Networks 44 (2004) 569–582
expected to suffer higher congestion and typically to experience higher losses and delays than high-priority ones. In the example of Fig. 12 Wlow is set to 2. These two concepts have been applied to the previously explained WDS algorithms in order to achieve service differentiation, leading to the algorithms named MINL-D and MQWS-D, in the case of the threshold-based technique, and MINLLIM and MQWS-LIM, with reference to the wavelength-based technique.
1
Packet Loss Probability
578
10
10
10
10
10
-1
-2
-3
-4
Tlow = 2D
-5
Tlow = 3D 10
-6
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
D
Numerical results are presented here in order to evaluate how the proposed techniques for QoS management are able to provide service differentiation between the two different traffic classes. The performance figures are again the packet loss probability and the packet delay. Results have been obtained by using the same simulation set-up defined in Section 3 to analyze the undifferentiated network scenario. In particular, in the connectionoriented environment the two classes of service are assumed to be associated with LSPs at simulation start up. 5.1. Connectionless operation
Fig. 13. Packet loss probability for MINL-D for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 4, varying the value of the threshold Tlow (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
1.00 0.90
Low
0.80
High
0.70
Distribution
5. Performance evaluation with QoS
0.60 0.50 0.40 0.30 0.20 0.10
Fig. 13 plots the packet loss probability for the MINL-D algorithm for B ¼ 4 and different values of Tlow . The restricted exploitation of the buffer capacity for low-priority packets leads to a high penalty in terms of packet loss probability for the low-priority traffic, while a good behavior is shown for the high-priority traffic. Moreover, performance of each class of service can be straightforwardly tuned by varying the value of the threshold. Fig. 14 shows the delay distribution for low- and high-priority classes for B ¼ 4 and for Tlow ¼ 2D. The high usage of the zero-delay line by both classes is evident. On the other hand, the highest delays are exploited only by the high-priority traffic as a consequence of the algorithm and, in any case, with an extremely limited probability, that is below 106 . In fact, it is very important to
0.00
0
1
2
3
Delay
Fig. 14. Delay distribution for MINL-D for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 4, D ¼ 1:5 and a value of the threshold Tlow ¼ 2D.
note that, for the case of B ¼ 4 considered here, the optical buffer is very short and the queuing delay is always very small, in particular when compared to propagation delays. So, it is possible to state that the application of the threshold-based scheme turns out to be an efficient technique to achieve QoS differentiation, with good values of performance and with a reasonably small processing effort.
F. Callegati et al. / Computer Networks 44 (2004) 569–582 1
579
1.00 0.90
Low
0.80
High
0.70 10 -2
Distribution
Packet Loss Probability
10 -1
10 -3
10
-4
10
-5
0.60 0.50 0.40 0.30 0.20
10 -6
w low = 8 w low = 9 wlow = 10 w low = 12
0.10 0.00 0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
1
2
3
Delay
2
D
Fig. 15. Packet loss probability for MINL–LIM for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 4, varying the number of wavelengths available to low class (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
Fig. 16. Delay distribution for MINL-LIM for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 4, D ¼ 1:3 and a value of the threshold Wlow ¼ 8.
1
10 -1
Packet Loss Probability
In Fig. 15 the packet loss probability for the MING-LIM algorithm with different values of Wlow and B ¼ 4 is plotted. By exploiting the wavelength dimension a sensible differentiation in terms of packet loss can be achieved, allowing a very effective tuning of the order of magnitude of packet loss probability of both classes of service and also providing QoS differentiation with good flexibility. In comparison with Fig. 14, it is interesting to observe the delay distribution plotted in Fig. 16 for Wlow ¼ 8 and B ¼ 4. It can be observed that differentiation is also achieved in terms of delay in the sense that high-priority packets are mostly served with the minimum delay, while lowpriority packets typically suffer more congestion due to the limited wavelength resources available.
10 -2
10 -3
10 -4
10 -5
10 -6 0
Tlow = 5D Tlow = 6D Tlow = 7D 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
Fig. 17. Packet loss probability for MQWS-D for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8, varying the value of the threshold Tlow (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
5.2. Connection-oriented operation The packet loss probability for the MQWS-D algorithm is plotted in Fig. 17, for B ¼ 8 and for different values of the threshold Tlow . As expected, the threshold-based technique is able to provide service differentiation also in a connection-oriented scenario, with a good benefit in terms of packet loss rate for high-priority traffic, achievable by varying the threshold position. Fig. 18 plots the delay dis-
tribution for low- and high-priority classes for B ¼ 8 and for Tlow ¼ 5D. Both classes of service experience high usage of the lowest value of delay, while it is easy to observe the different exploitation of the buffer, in terms of delay distribution, by the two classes of service, as a consequence of the threshold value. On one side the delay for lowpriority packets is limited by the threshold value, while over this value the delay probability for
580
F. Callegati et al. / Computer Networks 44 (2004) 569–582 1.00
1.00
0.90
low
0.90
low
0.80
high
0.80
high
0.70
Distribution
Distribution
0.70 0.60 0.50 0.40 0.30
0.60 0.50 0.40 0.30
0.20
0.20
0.10
0.10
0.00
0.00 0
1
2
3
4
5
6
7
0
Delay
high-priority packets decreases with the delay value due to the lower usage of the buffer positions reserved to high-priority traffic only. In Fig. 19 the performance of MQWS-LIM is presented for B ¼ 8, varying the value of Wlow . As in the connectionless case, the figure shows that in the wavelength domain a sensible service dif-
Packet Loss Probability
1
-1
10
-2
10
-3
10
-4
10
-5
10
-6
0
w low = 4 w low = 6 w low = 8 w low = 10 0.1
0.2
0.3
0.4
0.5
2
3
4
5
6
7
Delay
Fig. 18. Delay distribution for MQWS-D for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 8, D ¼ 0:4 and a value of the threshold Tlow ¼ 5D.
10
1
0.6
0.7
0.8
0.9
Fig. 20. Delay distribution for MQWS-LIM for both classes of services, for N ¼ 4, W ¼ 16, q ¼ 0:8, B ¼ 8, D ¼ 0:4 and a value of the threshold Wlow ¼ 8.
ferentiation level can be achieved in terms of packet loss by effectively tuning the performance of both classes of service through the parameter Wlow . Fig. 20 shows the delay distribution for lowand high-priority classes for Wlow ¼ 8 and B ¼ 8. The delay distribution behaves in the same way for both classes, exhibiting the highest probability value for delay equal to zero. This is due to the fact that the two classes are characterized by the same load that is split almost equally on the two subsets of 8 wavelengths each. When queues are significantly not full, the algorithm tends to isolate the two subsets of wavelengths, devoting one to the low-priority packets plus a small fraction of high-priority packets. This causes this subset to be loaded by a little more than half of the total traffic and leads to a bit greater delay probability with respect to high-priority traffic. In fact, highpriority LSPs find the wavelengths used by lowpriority LSPs less convenient in terms of queue occupancy and tend to use only their exclusive wavelengths.
1
D
Fig. 19. Packet loss probability for MQWS-LIM for both classes of services as a function of the delay unit D, for N ¼ 4, W ¼ 16, q ¼ 0:8 and B ¼ 8, varying the number of wavelengths available to low class (the upper set of curves refers to the low priority class, while the lower one refers to the high priority class).
6. Conclusions Congestion resolution algorithms able to support quality of service in DWDM optical packetswitched networks have been considered and analyzed in both connectionless and connection-
F. Callegati et al. / Computer Networks 44 (2004) 569–582
oriented contexts. Good differentiation and QoS tuning capabilities with two classes of service have been shown, although performance of the two classes is correlated. QoS differentiation is achieved by means of two techniques, the former reserving resources to the high priority traffic in the delay queue (threshold-based) the latter reserving resources in the wavelength domain (wavelengthbased). The threshold-based technique achieves good discrimination between the two classes in terms of packet loss while naturally limiting the delay of the low-priority class that, on the other hand, suffers the highest loss. The wavelengthbased technique achieves similar performance and is also more effective for finer QoS tuning. Some effects of the algorithms on delay distribution have also been shown and discussed.
581
[10] F. Callegati, W. Cerroni, C. Raffaelli, P. Zaffoni, Dynamic wavelength assignment in MPLS optical packet switches, Optical Networks Magazine 4 (5) (2003) 41– 51. [11] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, Z.W. Weiss, An architecture for differentiated services, IETF RFC 2475, December 1998. [12] D.C. Stephens, J.C.R. Bennett, H. Zhang, Implementing scheduling algorithms in high-speed networks, IEEE Journal on Selected Areas in Communications 17 (6) (1999) 1145–1158. [13] L. Tancevski, L.S. Tamil, F. Callegati, Non-degenerate buffers: a paradigm for building large optical memories, IEEE Photonic Technology Letters 11 (8) (1999) 1072– 1074. [14] E. Rosen, A. Viswanathan, R. Callon, Multiprotocol label switching architecture, IETF RFC 3031, January 2001. [15] F. Callegati, G. Corazza, C. Raffaelli, Exploitation of DWDM for optical packet switching with quality of service guarantees, IEEE Journal on Selected Areas in Communications 20 (1) (2002) 190–201.
References [1] P. Gambini et al., Transparent optical packet switching: network architecture and demonstrators in the KEOPS project, IEEE Journal on Selected Areas in Communications 16 (7) (1998) 1245–1259. [2] C. Guillemot et al., Transparent optical packet switching: The European ACTS KEOPS project approach, IEEE/ OSA Journal of Lightwave Technology 16 (12) (1998) 2117–2134. [3] L. Dittmann et al., The European IST project DAVID: A viable approach towards optical packet switching, IEEE Journal on Selected Areas in Communications 21 (7) (2003) 1026–1040. [4] D. Chiaroni et al., First demonstration of an asynchronous optical packet switching matrix prototype for multiterabitclass routers/switches, in: Proceedings of 27th European Conference on Optical Communication (ECOC 2001), Amsterdam, The Netherlands, October 2001. [5] J. Turner, Terabit burst switching, Journal of High Speed Networks 8 (1) (1999) 3–16. [6] F. Callegati, H.C. Cankaya, Y. Xiong, M. Vandenhoute, Design issues for optical IP routers, IEEE Communications Magazine 37 (12) (1999) 124–128. [7] L. Tancevski, S. Yegnanarayanan, G. Casta~ non, L. Tamil, F. Masetti, T. McDermott, Optical routing of asynchronous, variable length packets, IEEE Journal on Selected Areas in Communications 18 (10) (2000) 2084– 2093. [8] F. Callegati, Optical buffers for variable length packets, IEEE Communications Letters 4 (9) (2000) 292– 294. [9] F. Callegati, W. Cerroni, G. Corazza, Optimization of wavelength allocation in WDM optical buffers, Optical Networks Magazine 2 (6) (2001) 66–72.
F. Callegati received his Master and Ph.D. in Electrical Engineering in 1989 and 1992 from the University of Bologna, Italy. In 1993 he was research scientist at the Teletraffic Research Centre of the University of Adelaide, Australia and in 1994 at Fondazione U. Bordoni, Italy. He has been with DEIS at the University of Bologna since 1995, where he is now serving as Associate Professor. His research interests are in the field of traffic modeling and performance evaluation of telecommunication networks. He is currently working in the field of optical packet switching and optical networking for Internet traffic. He has been active in UE funded research projects on Optical Packet Switching such ACTS KEOPS and IST DAVID.
W. Cerroni received the Master degree in Telecommunication Engineering and the Ph.D. in Electrical and Computer Engineering both from the University of Bologna, Italy, in 1999 and 2003 respectively. He is currently holding a post-doc position at the University of Bologna. During 1999 he was a visitor researcher at the University of Texas at Dallas (UTD) working in cooperation with Alcatel Corporate Research Center, Richardson (Texas), USA. His research interests cover traffic performance evaluation of communication networks and architectures for very high-speed optical packet switching. He contributed to some of the activities within the EU-funded IST DAVID (Data And Voice Integration over DWDM) project as well as to national research projects on optical packet switching and grid computing infrastructures.
582
F. Callegati et al. / Computer Networks 44 (2004) 569–582
C. Raffaelli was born in Bologna, Italy, in 1960. She received the Dr. Eng. Degree in Electronic Engineering from the University of Bologna in 1985 and the Ph.D. Degree in Electronic Engineering and Computer Science in 1990. Since 1985 she has been with the Department of Electronics, Computer Science and Systems of the University of Bologna, where she became a Research Associate in 1990. Since then her research activity was concerned with broadband communications, and in particular with performance and implementation of ATM switching architectures. She has been involved in many National and International research projects on high-speed networks among which the KEOPS project of the European Union on photonic transport of information. She is currently involved in National projects on high-speed TCP/IP transport and in the DAVID IST project of the European Union (Data And Voice Integration over DWDM). She is
Associate Professor in Switching Systems and Telecommunication Networks at the University of Bologna. P. Zaffoni received the Master degree in Telecommunication Engineering from the University of Bologna, Italy, in 2001. He is currently pursuing the Ph.D. degree in Electrical and Computer Engineering at the University of Bologna. He is working in the field of traffic performance and architectures for broadband communications networks. In particular his research topics include optical packet switching, optical networking and TCP performance in heterogeneous networks scenarios. He is currently involved in National projects on optical network architectures.