TO meet the increasing bandwidth demands and reduce

Transcription

1 Assembling TCP/IP Packets in Optical Burst Switched Networks Xiaojun Cao Jikai Li Yang Chen Chunming Qiao Department o Computer Science and Engineering State University o New York at Bualo Bualo, NY - Abstract Optical Burst Switching (OBS) is a promising paradigm or the next-generation Internet inrastructure. In this paper, we study the perormance o TCP traic in OBS networks and in particular, the eect o assembly algorithms on TCP traic. We describe three assembly algorithms in this paper and compare them using the same TCP traic input. The results show that the perormance o the proposed Adaptive-Assembly-Period (AAP) algorithm is better than that o the Min-BurstLength-Max- Assembly-Period (MBMAP) algorithm and the Fixed-Assembly- Period (FAP) algorithm in terms o goodput and data loss rate. The results also indicate that burst assembly mechanisms aect the behavior o TCP in that the assembled TCP traic becomes smoother in the short term, and more suitable or transmission in optical networks. I. INTRODUCTION TO meet the increasing bandwidth demands and reduce costs, several optical network paradigms have been under intensive research. O all these paradigms, optical circuit switching (e.g. wavelength routing) is relatively easy to implement but lacks lexibility to cope with the luctuating traic and the changing link state; Optical Packet Switching (OPS) is conceptually ideal, but the required optical technologies such as optical buer and optical logic are too immature or it to happen anytime soon. A new approach called Optical Burst Switching (OBS) that combines the best o optical circuit switching and optical packet switching was proposed by researchers [1] [], and has received increasing amount o attention rom both academia and industry worldwide [3 7]. Within an OBS network, an ingress OBS node assembles IP packets into bursts and sends out a corresponding control packet or each data burst. This control packet is delivered out-o-band and leads the data burst by an oset time. The control packet reserves necessary resources all the way rom the ingress node to the egress node where the data burst will be disassembled. While TCP is the dominant transport protocol today and likely to be adopted in uture optical networks, all existing work evaluating the perormance o OBS networks assumes that the input traic is generated by a random generator or at the best a sampled trace ile o IP packets, neither o which can be used to evaluate the eects o the congestion control and retransmission by TCP [,9]. In this paper, we are interested in inding out how the assembly algorithms perorm under TCP congestion control mechanisms and what the assembled TCP traic would be. To this end, we use the TCP traic as input to study perormance o OBS networks and compare dierent assembly algorithms using NS [1]. One o the insights obtained rom our study is that TCP perormance is more sensitive to the assembly period (AP) than the burst length, and hence ideally, the assembly period should adapt to the TCP s sending window size. The remainder o the paper is organized as ollows. Section II briely introduces the main idea o OBS and describes an existing burst assembly algorithm, then we propose two new assembly algorithms. We present simulation results and analysis in Section III. Section IV concludes our work. II. BURST ASSEMBLY ALGORITHM IN OPTICAL BURST SWITCHED NETWORKS Fig. 1 shows the basic procedure o sending one burst rom an ingress node to an egress node in an OBS network. At Burst Assembly O/E/O O/E/O O/E/O Fig. 1. Control packet Data burst Control Channel Data Channel Optical Burst Switched Networks O/E/O the ingress nodes o an OBS network, all TCP/IP packets are assembled into bursts. The ingress node sends out a control (or setup) packet beore sending out the data burst. There is an oset time between the control packet and the data burst to give the intermediate OBS nodes enough time to conigure their switching abrics and reserve channel or the ollowing data burst. The control packets are sent out on one or more dedicated control channels (e.g. wavelengths) and go through O/E/O conversion at each intermediate node to provide inormation about the coming burst. However, the data burst will go through each intermediate node in the optical domain without any O/E/O conversion. There are many interesting topics in OBS, such as burst scheduling, burst assembly, oset time setting, and contention resolution []. The ocus o this work is the assembly mechanisms and their interactions with TCP/IP. Currently, how to eiciently assemble IP packets to bursts in an OBS network is still an open issue. In this section, we will describe three burst assembly algorithms. We assume that the ingress node has one dedicated burst assembly queue or each egress node. All incoming packets will be orwarded to the corresponding queue according to their destinations. When the queue size reaches a threshold or the waiting time o the packets in the queue reaches a threshold, the packets in this queue are sent out as a burst. Fig. illustrates the structure o such a burst assembler.

2 IP traic rom traditional IP Routers Fig.. Burst Assembly d1 dn di where, min(rt O RT T ) represents the minimum value o (RT O RT T ) over all TCP lows rom s to d. Since changes with the value o AP s, it is diicult to obtain the accurate lower bound o AP s. One possible approach, however, is to iner rom the traic history. More speciically, we may use the ollowing Equation (3), which is similar to the equation or RTT calculation [, 9, 11]. A. The Fixed-Assembly-Period (FAP) Algorithm A simple and intuitive assembly algorithm is the Fixed- Assembly-Period (FAP) algorithm. The edge OBS node assembles IP packets with the same destination that arrive during a ixed Assembly Period (AP) into a burst and generates a corresponding control packet. The ollowing pseudo code provides a more detailed description o this FAP algorithm, where q sd denotes the assembly queue at ingress node s or destination d and AP s is the assembly period used by node s. d=destination o the new IP packet; i (AssembleTimer (q sd ) is not running) then Start a new assembly timer or q sd, AP qsd =initial value; Update the burst length inormation; Fill in and send out a control packet on a control channel; Schedule the data burst to be sent out on a data channel ater an oset time; Stop the assembling timer or q sd ; Assume that the network uses a static routing algorithm. For the link between two node i and node j, we deine a set Q ij, which contains all q sd i and only i the bursts assembled rom q sd use the link rom node i to node j. Then, or any Q ij,we have the ollowing inequality based on the link capacity. Channel BandWidth (1) AP s q sd Q ij where denotes the average burst length o queue q sd. Channel is the number o wavelengths on this link and BandWidth is the bandwidth o one channel. Note that or a speciic TCP low, in order to prevent TCP rom unnecessarily (or prematurely) retransmitting packets in the burst, the assembly period should not be greater than retransmit timeout value (RTO) minus the round trip time value (RTT) associated with low [, 9, 11]. From inequality (1) and above discussion, we have the ollowing constraints on assembly period (Assume that every ingress node uses the same AP.). q sd Q ij Channel BandWidth AP s min(rt O RT T ) () AvgBL q = ξ AvgBL q + η SAvgBL q (3) where SAvgBL q is the most recently sampled average burst length, AvgBL q is the smoothed average burst length, and ξ and η are two positive weights (ξ + η =1). Note that there can be many variants o the FAP algorithm. One extreme is to use the same ixed AP value or all ingress nodes (i.e. AP s = AP s ), and another extreme is to use dierent but ixed AP values or dierent assembly queues at each ingress node (i.e. AP qsd AP qsd ). In the ollowing simulation study, we assume that all the ingress nodes use the same AP value. B. The Adaptive-Assembly-Period (AAP) Algorithm As we can see, algorithm FAP is too rigid in that it does not take current traic situation into account so as to adapt the burst assembly accordingly. This is expected to adversely impact the network perormance. Thereore, we propose the Adaptive- Assembly-Period (AAP) algorithm. This algorithm is similar to FAP algorithm, with the dierence being that AAP can dynamically change the value o AP o any queue at every ingress node (e.g. AP qsd ) according to the length o burst recently sent. Given a queue q sd, i we assume that the network uses a single path to route each low, in the extreme case, the assembled burst traic rom q sd occupies all the output bandwidth. Hence, similar to the rationale behind equation (), we have ollowing constraints. BandWidth Channel AP q sd min(rt O RT T ) () where AP qsd is assembly period o queue q sd. We now use the ollowing equation to determine the value o assembly period, AP qsd = α () BandWidth Channel where α 1 is called the assembly actor. Note that rom equation (1) and (), we get, α = BandWidth Channel Q ij () AP qsd where Q ij is the size o Q ij. In other words, Equation () indicates that α is the average number o the assembly queues that are sharing the link. As ar as determining in equation () is concerned, we propose to use Equation (3) by setting ξ = 1 and η = 3 in our simulation. The reason or putting more weight on the most recently sampled burst (i.e. η > 1 ) is to allow this assembly algorithm to synchronize with TCP as much as

3 possible. More speciically, ater a long burst is sent out, it is very likely that TCP will still send out more (sometimes twice as many) packets and thus it is better to increase AP. On the other hand, i TCP is done or entering a slow start stage (e.g. a short burst is sent out), it is desirable to reduce the assembly period signiicantly. In addition, to prevent the burst length rom increasing or decreasing by too much, we use lower and upper thresholds (i.e. β>1 and <γ<1) to keep the assembly period within a reasonable range. More details about the AAP algorithm are shown in the ollowing pseudo code. d=destination o the new IP packet; i (AssembleTimer (q sd ) is not running) then Start a new assembly timer or q sd,(ap qsd =Current (or initial) value); Update the burst length inormation; Fill in and send out the control packet on control channel; Schedule the data burst to be sent out on a data channel ater an oset time; Update the statistical average burst length (see equation (3)); Update the value o Assembly Period or queue q sd (see equation ()); i (AP qsd >β current assembly period) then AP qsd = β current assembly period; else i (AP qsd <γ current assembly period) then AP qsd = γ current assembly period; Stop the assembling timer or q sd ; Note that while AAP is much more lexible than FAP, it also involves more complexity (in adjusting the AP) at each ingress node. A compromising approach, called Semi- Adaptive-Assembly-Period (SAP) algorithm is a possible alternative, where all the queues in Fig. at each given node use the same assembly period value, which dynamically changes based on the current traic. Due to the space limitation, we will not discuss SAP urther in this paper. C. The Min-BurstLength-Max-Assembly-Period (MBMAP) Algorithm While we do not want to send out data bursts that are too small in order to reduce the overhead, we should send out a burst as soon as possible, and certainly beore the irst packet in the burst misses its deadline. So we propose the third algorithm, which generates a control packet when a burst exceeds a minimum burst length (MBL) or when the assembly period times out, whichever comes irst. These two parameters (maximum assembly period (MAP) and MBL) can be set such that the minimum burst length is smaller than the average burst length (obtained using equation (3)), and the maximum assembly period is approximately min (RT O RT T ) (as in equation ()). The pseudo code or the MBMAP algorithm is as ollows. d=destination o the new IP packet; i (AssembleTimer (q sd ) is not running) then Start a new assembly timer or q sd, MAP=initial value; Update the burst length inormation; i (AssembedLength MBL) then Generate burst control packet or this burst; Fill in and send out the control packet on a control channel; Schedule the data burst to be sent out on a data channel ater an oset time; Stop the assembling timer or q sd ; Generate a burst control packet or this burst; Fill in and send out a control packet on a control channel; Schedule the data burst to be sent out on a data channel ater an oset time; Stop the assembling timer or q sd ; Note that when we disable the constraint on the minimum burst length 1, this algorithm degenerates into FAP. III. NUMERICAL RESULTS &PERFORMANCE ANALYSIS In this section, we study the perormance o previously proposed assembly algorithms using TCP traic as input to an OBS network, as shown in Fig. 3. The italic number on the link in the igure represent the propagation delay (or distance) between two nodes. According to our experiments, there is no signiicant dierence in the perormance among dierent TCP implementations (such as Reno, NewReno or Sack ). Due to the space limitation, we report only the results obtained using the TCP NewReno in this section. We make the ollowing assumptions: Fig NSFNet Every OBS node has ull wavelength conversion capability. Each link has bi-directional ibers, each iber has wavelengths. 1 In ollowing simulation, we denote this by setting MBL=.

4 There is no buer or FDL in the OBS network. The TCP setup and ACK packets go through the control channel. FTP connections are used to generate TCP traic, whose destination is randomly distributed among all the nodes. A. Comparison Between OBS and OPS Conceptually, OPS (Optical Packet Switching) is a special case o OBS i the assembly schemes are disabled in OBS (i.e when AP=). Burst number 1 1 OBS:FAP(AP=.s) Time(s)[TCPNum=3] traic rate(mb/s) 1 OBS:FAP(AP=.s) Time(s)[TCPNum=3] Fig.. Burst number Time Fig. 9. Data rate Time as well as the data rate. The smoothing eect is veriied by the R/S and variance plot [] [13] [1] [1] in Fig. 1 and 11, which show the statistical properties o OPS traic (no assembly) and OBS traic (ater assembly). Fig.. Goodput TcpNum Fig.. Loss Rate TcpNum Fig. and Fig. are the perormance comparisons o OBS (using FAP) and OPS in terms o goodput (i.e. the eective data that the destinations receive.) and loss rate as a unction o the number o TCP connections. The igures indicate that with an appropriate assembly algorithm and assembly parameters, the OBS networks have better perormance than that o OPS networks in that the goodput and data loss rate o OBS with FAP are about 1% better than OPS in most cases. Also note that, as the number o TCP connection increases, the network gets saturated, and the goodput and loss rate tend to be a constant. Packet number Fig Time(s)[TCPNum=3] Packet number Time traic rate(mb/s) Time(s)[TCPNum=3] Fig. 7. Data rate Time In order to explain the above results, we examine the traic proiles. Fig. Fig. 9 show the dierence between original TCP traic and assembled TCP traic at a randomly picked node. Fig. and Fig. 7 are the packet number arrival rate and data arrival rate o the original TCP traic. The number o simultaneously arrived packets may be about in OPS networks. It is very likely that they will cause contentions at the output ports. Similarly, Fig. and Fig. 9 are the results or assembled traic using FAP algorithm, which shows that burst assembly reduces the variance in the number o bursts/packets Log1(R/S) Ater Assembly Log1(n) Fig. 1. R/S plot Log1(Variances) Ater Assembly Log1(n) Fig. 11. Variance plot The slopes or no-assembly TCP traic and ater-assembly traic are almost the same, which indicates that the selsimilarity degrees o both traic are not much dierent. In other words, burst assembly algorithm does not change the traic s long-range dependence property. Fig. 11 exhibits that the variance o assembled traic is lower than that o the unassembled traic, indicating that burst assembly indeed changes the traic s short-term properties. Such a smoothing eect in the short-term makes the assembled TCP traic more suitable or transmission (i.e. results in a higher goodput and lower loss rate as shown in Fig. and ). B. Comparison Between Assembly Algorithms 1) FAP: Fig. shows that the perormance o FAP is aected by the value o assembly period. More speciically, when the selected assembly period is not appropriate (e.g. >.3 when TCPNum=), the goodput o OBS is worse than that o Optical Packet Switching (i.e. AP=). This is because when the assembly period is too large, it takes more time to send packets and receive the corresponding acknowledgement packets, and hence the increase in the TCP sending window size is slowed, which results in a poor perormance. This Figure also indicates that the optimal value o assembly period increases with the number o TCP connections. For example, it is about.1.1 and. when the number o TCP connections increases rom, to 3 respectively. ) MBMAP: There are two parameters in Min- BurstLength-Max-Assembly-Period algorithm: minimum burst length (MBL) and maximum assembly period (MAP). Fig. 13 and Fig. 1 are the goodput or MBMAP by varying

5 3) AAP: Fig. 1 and Fig. 1 show the simulation results or the AAP algorithm. Here, we set α =1as the assembly actor (based primarily on the topology and the shortest path routing or every node pair) and set β =3,γ =1/β. The curves representing FAP in these igures are the best perormance we can obtain rom FAP algorithm by selecting the best (ixed) value o AP. Still, the perormance o AAP is better than FAP, especially or a large number o TCP connections. The reason is that AAP adapts the assembly period properly. The ideal situation is that the variation in assembly period can synchronize with TCP s congestion control mechanism, and thus enhances the perormance. Fig. 13. Goodput Minimum Burst Length Fig.. Goodput Assemble Period Fig. 1. Goodput Maximum Assembly Period one o these parameters at a time. More speciically, Fig. 13 indicates how the value o MBL aects the system s goodput, where MBL= actually represents the FAP algorithm. Fig. 1 shows how the value o maximum assembly period aects the system s goodput. Apparently, the network perormance is more sensitive to MAP than MBL because the TCP congestion control and retransmission mechanisms are basically controlled by timers. As can be seen rom the igures, or the same assembly period, MBMAP perorms as well as FAP in most cases. Fig. 1. Goodput TCPNum Fig. 1. Loss Rate TCPNum IV. CONCLUSION In this paper, we have proposed and compared three burst assembly algorithms, which are Fixed-Assembly-Period (FAP) algorithm, Adaptive-Assembly-Period (AAP) algorithm and Min-BurstLength-Max-Assembly-Period (MBMAP) algorithm. Based on the properties o TCP low and bandwidths provisioning in the networks, we have proposed approaches to determine the parameters such as assembly period and the minimum burst length o above algorithms. The simulation results show that AAP is the best, because it adapts assembly periods to match with the TCP congestion control mechanisms. We have used the R/S plot and variance plot to provide valuable insights into the eect o OBS assembly mechanisms on the burstiness and sel-similarity degree o the TCP traic. Our analysis and simulation results have indicated that burst assembly can reduce the simultaneous contention in the core network and make the TCP traic smoother in the short time scale, and hence improve the perormance in terms o goodput and loss rate. REFERENCES [1] M. Yoo, M. Jeong, and C. Qiao, A high speed protocol or bursty traic in optical networks, in SPIE s All-Optical Communication Systems: Architecture, Control and Protocol Issues, Vol. 33, pp. 79-9, Nov [] C. Qiao and M. Yoo, Optical burst switching (OBS)-A new paradigm or an optical internet, J. High Speed Networks, Vol., pp. 9-, [3] J. Turner, Terabit burst switching, Journal o High Speed Networks, Vol. 9 No.1, pp. 9-, [] Y. Xiong, M. Vandenhoute and H. Cankaya, Control Architecture in Optical Burst-Switched WDM Networks, IEEE Journal on Selected Areas in Communications, Vol. 1 No. 1, October, PP [] L. Xu, H. Perros, and G. Rouskas, Techniques or Optical Packet Switching and Optical Burst Switching, IEEE Comm. Mag., Vol. 39, No. 1, pp. 13-, Jan. 1. [] M. Dser and P. Bayvel, Analysis o wavelength-routed optical burstswitched network perormance, Optical Comm., ECOC 1, pp. -7. [7] K. Dolzer, C. Gauger, J. Spth and S. Bodamer, Evaluation o Reservation Mechanisms or Optical Burst Switching, Journal o Electronics, Vol., No. 1, pp. - 9, 1. [] V. Jacobson Congestion Avoidance and Control, SIGCOMM Symposium on Communications Architectures and Protocols, pages 31-39, 19. [9] V. Jacobson, Modiied TCP congestion avoidance algorithm, April 3, 199, endend-interest mailing list. [1] The Network Simulator ns-, http: // [11] D.D. Clark and J. Hoe, Start-up Dynamics o TCP s Congestion Control and Avoidance Schemes, Technical report, Jun.199. Presentation to the Internet End-to-End Research Group. [] W. Leland, M. Taqqu, W. Willinger, and D. Wilson, On the Sel-Similar Nature o Ethernet Traic (Extended Version), IEEE/ACM Trans. on Networking, Vol., No.1, PP. 1-1, February 199. [13] F. Avram and M. Taqqu, Robustness o the R/S statistic or ractional stable noises, Statistical Inerence or Stochastic Processes, pp.9-3,. [1] B. B. Mandelbrot, Limit theorems on the sel-normalized range or weakly and strongly dependent processes, Zeitschrit ür Wahrscheinlichkeitstheorie und verwandte Gebiete, 31:71-, 197. [1] B. B. Mandelbrot and J. R. Wallis, Computer experiments with ractional Gaussian noises, Parts 1,,3 Water Resources Research, :-7, 199.