Reviews

A study on the management of on/off loss sensitive sources traffic in ATM networks

Description
... Iera, Salvatore Marano, Roberto Sestito DEIS Universita' della Calabria 87036 Rende (Cosenzn) ITALY phone +39 984 494.702494.703 fax +39 984 494.713 E-dl antonio@elecom.deis.unical.it ABSTRACT This paper presents ...
Categories
Published
of 5
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  A STUDY ON THE MANAGEMENT OF ON/OFF LOSS SENSITIVE SOURCES TRAFFIC IN ATM NETWORKS zyxwv oberto Beraldi, Antonio Iera, Salvatore zyxw arano, Roberto Sestito D.E.I.S. niversita' della Calabria zyxwv 7036 Rende Cosenzn) ITALY phone +39 984 494.702494.703 fax +39 984 494.713 E-dl antonio@elecom.deis.unical.it ABSTRACT This paper presents a zyxwvut tudy on the management zyxwvuts f bursty loss-sensitive sources traffic in ATM (krynduonous Transfer Mode) etworks. In particular, the study is focused on the policing mechanisms, which allow the active codons raffic to be monitored in order to avoid congestion. A new policing mechanism, Double Threshold Avalanche Tagging @TAT) is developed and its behaviour is zyxwvut ested using an ATM network model and assumhg homogeneous sources. In such a new mechanism, the zyxwvutsrqpo dvantage of adding a per-burst marking component to rraditiod percell marking based mechanisms, such as VLB (virtual Laky Bucket) and AT (Avalanche Tagging), is considered. Obtained results confirm that the intmduction of a per-burst monitoring component contrii both to inmas network performance and to improve network resource utilization. 1. INTRODUCTION In recent years, the availability of wide bandwidth optical fikr links capable of carrying hundreds of Mbps per and the growing demand for high performance integrated networks, has guided research towards wide-area b l integrated networks [ 101. ATM (Asynchronous Transfer Mode) is the standard adopted by the CCIlT for the implementation of wide-area b l integrated networks, namely the Broadband Integrated SeMces Digital Networks (€8-ISDN) [ 111 [91. The choice of ATM for transporting information in the B-ISDN produces two undamental advantages in respect to networks using the Synchronous Transfer Mode. Above all, the B-ISDN becomes a more flexible network, both in the sense of adaptability to many seMces and applications and in the capacity to transport a large amount of traffic with different characteristics and different QoS requests, thereby permitting any traffic to vary its characteristic during the call Secondly, B-ISDN, whose tdic or the most part is of the odoff type, zyxwvutsr an use its link bandwidth in a more efficient manner by adopting some special types of bandwidth allocation strategies capable of taking advantage dthe statistic nature of the traffic. One of the drawbacks of ATM is the problem of link bandwidth congestion, which produces cell loss phenomenon and thus degradation of the network performance. In order to convert a part of the cell loss problem in a cell delay problem, each node output link is supplied with a buffer. In this manner the congestion problem is transferred from the links to the buffers. However, other complex measures must be taken in order to further reduce the congestion event probability at any node Mer of a ATM network The most promising approach seems to be a preventive congestion control scheme based on a connection admission control 530 and supported by a policing function on the active connections and also supported by suitable Mer management algorithms and local reactive controls [5]. In this paper study is focused on the policing mechanism which allows the active connections traffic to be monitored in order to avoid congestion. In particular, the we f homogeneous odoff sources with high loss sensitivity is examined and the new policing mechanism, DTAT, is proposed which adds a per-burst policing component to the traditional percell policing method [2] and obtains interesting results rewng burst loss due to congestion. This s effected y incrementing the network performance in terms of a smaller retransmission time and improving effectiveness of network resource utilization by decreasing the presence of useless high priority cells in the network belonging to the damaged bursts. The paper is organized as follows. In Section 2, the problems of the policing of odoff sources with high oss sensitivity are outlined. Here the new policing technique, DTAT, is proposed. In Section 3, the model of an ATM node, its operating algorithms and assumptions are described. Furthermore, the output parameters of the system model are discussed and the simulation results, which give comparisons between the policing effectiveness of DTAT, AT and VLB, are presented. Conclusions are given in Section 4 2. THE DTAT POLICING MECHANISM. ATM odoff sources with high loss sensitivity are characterized by all their bursts arriving intact at their destination~. This means that each burst must have all its cells present when it arrives at its destination, otherwise it must be completely retransmitted. Examples of ATM odoff sources with high loss sensitivity classes of seMce are video retrieval, document retrieval and data file transfer. With a marking-based policing scheme, when a source violates the established traffic descriptors values, excess cells are marked before entering the network and are eventually discarded by the nodes in the case of congestion. Since end-to-end protocols typically operate on the basis of bunts, it is preferable that an ATM congested node reacts by discarding cells from as few bursts as possible rather than indiscriminately discarding cells from all the arriving bursts. In other words, the network must preserve the integrity of the burst as a whole and attempt to maintain intact as many bursts as possible [4]. The policing algorithm has a determinant role in decreasing the number of bursts damaged at the nodes of the network. In [l] a proposed policing algorithm, Avalanche Tagging (AT), shows an improvement over SINGAPORE ICCSf94 0-7803-2046-8/94/$4.00 1994 IEEE 1  the trahtional Virtual Leaky Bucket (VLB)[2] algorithm in as much as there is an increment in the percentage of intact bursts arriving at destination during a call. The conceptual scheme illustrated in fig.1 is used here to explain VLB technique. The fundamental component of the policer is a counter whose behaviour is regulated by the values of two parameters adequately estabilished by the CAC (Connection Admission Control) function at call set- up time: i. e. the counter decrease period Te and the counter threshold zyxwvuts . nitially the counter zyxwvut tarts from zero (which is the minimum value), is increased each time a cell comes and is decreased at each Te time zyxwvut lots. The arriving cells are marked according to the following criteria. zyxwvu I; zyxwvutsrqponm n the arrival of a cell, the value of the counter is below that of S, then the cell remains unchanged and the counter is rncremented by one unit. On the other hand, i an arriving cell finds the counter at the value of S the cell is marked and the counter is le9 unalterated. s.u.1 +[~~--)~i zyxwv   .CLL I-'t ly-: -U1 ~.(UL muumw 1 zyx ig. 1. Reference scheme of VLB policing. VLB policing was first presented coupled with a special CAC scheme called CRR ( Class Related Rule ) with the following traffic desriptors values: Bp (connection peak band), Bm (connection average band) and L connection average burst length). Obviously this punishment mechamsm needs, in order to be completed, suitable Buffer and Queue Management Algorithms , running at the network nodes, and capable of penalizing the marked cells in the zyxwvuts ase of congestion (the two best known of these are the push-out scheme and the threshold scheme (31). How the AT algorithm works is non briefly explained (the reference scheme is the same as that of VLB illustrated in fig. 1): AT digerssfi'om 4B largely because a9er marking a cell of a burst it then continues to mark d he remaining cells of the burst, independently of counter value. During this process, the counter is periodically decreased. The AT marking strategy results in a concentration of marked cells in less bursts at the policer compared to VLB. This is subsequently translated into a concentration of discarded cells in less bursts at the nodes, thereby resulting in a greater number of intact bursts arrived at destination at the end of a call. From an analysis of the per-cell marking-based strategy adopted by the AT mechanism the possibility of adding a per-burst marking component was examined. This led to a new policing algorithm, Double Threshold Avalanche Tagging (DTAT), a derivative of AT and built by simply adding a per-burst basis marking component to the per-cell marking mechanism of the policing AT. The DTAT policing algorithm differs from AT only as regards the introduction of a new threshold, S1, whereas the threshold inherited from AT, is now renamed S2 (the reference model for the DTAT algorithm is illustrated in fig.2). In addition to the AT algorithm functions , DTAT performs a further function : i the first cell of a burst is received when the counter value is above the value of SI then DTAT marks the whole burst. In particular, with reference to the conceptual scheme illustrated in fig.2 the algorithm is the following: 1. The cell is marked when it finds the counter at the value of S2 or when the initial cell of a burst finds he counter at a value above S1. Whatever, when the cell is marked all remaining cells of the burst are similarly marked (that is the same as in AT fashion). 2. The counter is incremented each time an unmarked cell is passed. 3. The counter is periodically decreased at each Te lime slots (with rate Be). vmw- 4 = 4 .nR-lWLI-. Fig. 2. Reference scheme of DTAT policing. This shows clearly how, in the policing DTAT, the threshold S1 operates differently from that of the threshold S2, that is to say, S2 controls the marking of cells while S1 controls the marking of entire bursts. Furthermore, when the value of S1 is set to the value of S2, he behaviour of the policer DTAT equals the behavior of the policer AT. In order to permit the DTAT policer to determine the beginning and end of a single burst amongst cell traffic, a distinction is made between four different types of cells: I=Initial (initial cell of a burst), C=Central (central cell of a burst), F=Final (final cell of a burst), A=Alone (burst of one cell).As this distinction is necessarily only for the policers it is therefore proposed to implement it with the CLP (Cell Loss Priority) and ECN (Explicit Congestion Notification) bits of the cell header as illustred in fig. 3. Fig. 3. The CLP and ECN bits keep this meaning from the sources to the policers and then singly assume their srcinal meanings. The DTAT algorithm can be easily implemented at the high speed at which it must zyx un SINGAPORE ICCS/'94 53 0-7803-2046-8/94/$4.00 1 994 IEEE  To zyxwvutsr etter zyxwvutsrqponm nderstand how the DTAT algorithm works, by a me l DTAT zyxwvutsr olicer fed with pndelermi~ted burstytndecq&fkaUycbarn zyxwvuts othat the mwthnsshold S1 worksinthe most significant zyxwvuts ases. Inonier to arrive at a compnrison between the behaviour ofthe two mthods, comparative results arc shown in fig. and 4b. In zyxwvut he example of fig. 4a ( policer DTAT) he values of counter parameters were fixed as fofiows: s1=6, SM c-3 (time ab) while inthe example zyxwvutsr ffig. 4b ( policm AT) correspondent values arc: W e-3 (time rlotr) fig. shows, time slot by time slot, all the zyxwvutsr ctions wrccutcd rlmbdr: zyxwvutsrqponm   =decrepmt O=cell . =mark 0.510 = slot numbers IJ~ ountlng b4bbd444444 30 P 46 51 65 51 Fig. 4r. Exampk of AT mrrking. AbAAAbb4dbd . .- 44444444444 3a 36 4) 46 66 6 Fig. 4b. Example of 2TAT marking. Note, in the case ofpolicer DTAT of fig. 4a, hat in the time slot number 36 the poh starts by decmsing the counter from 7 to6, and soh ot mark the incoming cell buse he counter value is 6 which is not greater than S14, and so nturns the counter to 7 because the cell has As will be demonstrated, the DTAT policer produces the same percentage ofbursts arriving intact as the policer ATbut has the advantage of incrementig the average length of these bursts which thmfore results in a decnase in retransmission time n the casc ofburst lass. Monover in the case of congestion, the policing DTAT reduces the presence of useless high priority cells belonging to damaged bursts in the networks, thmby producing a more efficient use of network rebowces and improving network peff0llUlce. passed unmarked. 3. OUTPUT PARAMETERS ND RESULTS A separate system apiut is adopted as a context in which to examine the performance Mered by a policing algorithm with a generic input section of an ATM based network. The reference model of this system type is illustrated in fig. 5 where N sources, each one linkad to a corresponding policer and further linked to the multiplexer containing a buffer, can be distinguished. Fig. 5. Model of the system under study. An approach based on simulation and homogeneous severalillstmw of that system, Mering only for the policing algorithm, are realized. This permits a comparison between difbmt policing algorithms by using the system The algorithms "pared arc VLB, AT and DTAT. It is assumed hat the input and output links of the multiplexer are characterizedby a maximumbandwidth of 150 Mbps and that the multiplexer Mer has a dimension of 50 cell slots (in order to guarantee maximum delay of about 141 ps, which is reasonable for the type of source considered). The distributions of the peak and silent Mods lengths are supposed to be exponential, so if T enotes the average duration of the emission periods then zyxw he transition between states occurs with a fresuency 1/T (ON to OFF) nd l/T(b- 1) (OFF o ON), where b is the burstiness, defined as the ratio between the peak rate and the mean rate. The CRR ( Clas Related Rule )2] is the CAC rule used to establish the nominal point of the model operation, probability. Used traffic descriptors are Bp (peak bandwidth), Bm (caU average bandwidth) and L (burst average ength). The study presented in this paper is made on a sample source class, named SC, escribed by the tern of parameter values Bp=lO Mbps, Bm=2 Mbps, *lo0 cells. The chosen nominal Operating point of the model is: N-33 sources, S23200, Sl=vuiable, Be4.545 MbU8 which corresponds to an assigned badwidth of W=150 Mbps and to a mean offered load dA=6 With. The b r management algorithm used is the push-out scheme [2][3]. It is more adequate coupling DTAT policing mechanism to the push-out scheme rather than o the buffer management thnsbold scheme because the push-out scheme penalizes the marked cells to a greater extent and thereby leads greater &ciency to the DTAT mechanism, as inthecaseofAT [l]. In order to examine the model behavior in the case of Bm violatiom, the following table of values is adopted. Two possibilities BZC analpcl for evaluating the DTAT policing: impact on network performance and impact on network resources utilization. In order to evaluate the first possibility two ndexes are used the % ofintact bursts arriving at destination JBAD) the % ofcells arriving in the intact bursts (CAIB). In order to evaluate the second possibility the % of unmarked cells arriving in donurged bursts at destination (UCADB) has been used as ndex. traffichasbeenadoptad. throughput as a metric. to be of 105 maxi requested cell IOSS SINGAPORE ICCW94 532 0-7803-2046-8/94/$4.00 1 994 IEEE  'These three percentage values zyxwvut re calculated in the following manner: zyxwvutsr   zyxwvutsrq oo o lntact zyxwvutsrqpo ursts anivlng at destinahon total number of transrmtted bursts % BAD = x loo o of cells arriving in the intact bursts total number of transmittal cells zyxwvutsrqpo o CAIB = n ofunmarkedcel1s mdamagedbursts ?( zyxwvutsrq oo % UCADB = total number of bansnutted cells The BAD is the most important parameter of performance in a loss sensitive environment when the retransmission protocol works on the burst basis. Reducing the number of bursts to retransmit means reducing the time of the retransmission session at the end of a call. Furthermore, if M bursts are transmitted and N<M bursts arrive damaged at the end of a call, it is more desirable that N bursts for retransmission are the sdlest ones in order to further reduce retransmission time. zyxwvut o, when evaluating the network performance in such an environment it is more zyxwv orrect o consider not only the IBAD ut also the CAIB. The parameter UCADB is important because the unmarked cells in the damaged bursts exiting from a network node, waste network resources like link bandwidth. In fact, these unmarked cells belong to bursts that must be retransmitted anyway. I ---Y.m., --t*11 *. -*,,"", ---,a ** --cur*, Fig 69. % of intact bursts arriving at destination as a tinction of the threshold S1, for different average load values. At first, the reaction of the network performance to the variation of the parameter S1, throughout the over its whole range from 0 to 200 (200 is the constant value of S2) s studied in terms of BAD (Fig. 6a) and the CAIB (Fig.6 b). Note that when S1=S2 then the DTAT is equal o AT. Each curve refers to a particular average load value. The reaction of the network resources utilization to the variation of the S1 value is studied in the same manner, that is to say, by considering UCADB. I ;,': :,,: "* -.* i I ? . a Fig 6b. % of the cells arrived in the intact bursts as a function of the threshold S1, for different average load values. Fig. 6c clearly illustrates he improvements provided by DTAT policing. In the case of the CAIB (Fig. 6b) a vast order of improvement is obtained, while in the zy ase of the IBAD (Fig. 6a) the improvements are limited. This mechanism has been realized in order to improve the network performance varying the parameter of the CAIB (Fig. 6b) compared to that of AT policing by assuring that the IBAD s not reduced relative to AT policing (this has been effected for the whole range of the main offered load). In order to comv he DTAT results with the VLB and AT results it is necessary to fix a convenient value for S1. The value of S1 which provides the best average value of the CAIB over the whole range of the average load variation was found to be S1=14. So, all the following results referring to the policing DTAT are found by using S1=14 (Fig 6.b). -I l.,.r...??~i31CIII Fig 6c. % of the unmarked cells arrived in damaged bursts as a function of the threshold S1, for different average load values. In Fig. 6d, 6e and 6f the comparison is made between DTAT (with S1=14), AT and VLB policings with respective reference to IBAD, the CAB and UCADB as functions of the average load, when L=IOO nominal). Note as in Fig. 6d he IBAD provided by DTAT is substantially the same as that provided by AT, but, in Fig. 6e the CAIB is much greater in the case of DTAT demonstrating a better network performance. Fig. 6f clearly shows the increase in the UCADB given by DTAT as compared with the other two types of policing. This proves that the DTAT policing method provides a much more efficient utilization of the network resources. The impact of DTAT policing on the network performance and the efficiency of the network resources utilization was also tested in the critical case in which the sources violate z 5 SINGAPORE zyxwvutsrqpo CCS/ 94 0-7803-2046-8/94/$4.00 1 994 IEEE  not only zyxwvutsr he declared value of burstiness, but also the value of the average dcclamd zyxwvutsr urst length. In particuliu, the behaviour ofthe zyxwvuts ystem zyxwvutsrqpo as zyxwvut en zyxwvuts bsemd zyxwvu y maintaining constant the average burst length Lrso cells, which is half of the declared valw,and also for Lr400 cells, which is four times the declared value. Significant improvements are zyxwvutsr btahed with DTAT policing in all the three observed parameters. Even if this is a very stringent condition, the DTAT continues to respond positively. In order to investigate the reasons for which DTAT policing produces the observed advantages, cell loss phenomenon was investigated. The total loss of cells produced in the zyxwvut  se of DTAT policing is precisely the same as in the c ses of AT and VLB policing. But DTAT policing penalizes the marked succcss obtaind by DTAT is largely due to the particular method adopted in distrii markings to a dishonest traflic. In facf it distributes markings in such a manner that the congested nodes of the path am amunigd to concentrate the same number of cell elimination in the same number of bursts (sometimes minor); but smaller bursts, do respect at AT and, above all VLB cells fpr mre with respect to the unmarked cells. The 4. CONCLUSIONS In this paper a new policing mechanism named Double Threshold Avalanche Tagging @TAT) is proposed. The advantage of using the DTAT mechanism is shown in the c se of odoff and high loss sensitivity sounxs and also when the ntransrm ssion protocOr is of end-to-end type and works on a bursty basis The DTAT uses a new per-burst marking component which is better disposed to the needs of that type of soucct which results in a minor retransmission time in the case of burst loss due to network congestion. This esult is reached by incrementing the average length of the bursts passing intact across a congested n0de.h the case of congestion DTAT policing redm the presence of useless high priority cells in the network belonging to damaged bunts thereby facilitating more efIicient use of the network resources and contributing to the improvement of the network performance. The robustness of DTAT policing was tested both in the c se of burstness violation and in the c se of a simultaneous violation of burstness and average burst length. The implementation complexity of the new algorithm is kept low, as with Virtual Leaky Bucket , so that it can be simply implemented with existing technology at B-ISDN transmission zyxwvutsr peed. BIBLIOGRAPHY [ ] E. Di Nitto, A. Iera, S.Marano, Avalanche Tagging: a new proposal for policing hurction in ATM envinmmcnt, ICCC 1992. [2] zyxwvutsrqponm   Gelaui et al. ATM Badwidth Asigmmmt end Bandwidth Enfoment Policies , GLOBECOM 1989. [3] B. A. Malrrucki, A study of source MI anagement &d baer allocation in ATM nehvorks., ITC 12,1990 [4] J.S. Tm, hhagiq bandwidth in ATM ne.tworks with bursty traflic , IEEE Nehwrks, September 1992. [SI 2. Haas, J. H. Winters, Congestion control by adaptive admission , EEE INFOCOM 1991 [6] E. P. Rathgeb, Policing Mechanisms for ATM networks. Modeling end performance comparison. , lTC 12,1990. [q U kina et al. Bandwidth assignmart and virtual dl blocking in ATM networks , NFOCOM 1990. 181 E. P. web, Policing s or ATM networks. Modeliag d erfamuutce Camperison , TC 12,1990. [9] M. ina, V. Trccordi, TratEc gcmmt and congestion control for ATM networks , IEEE Network, septrmba 1992. for EISDN Savices , EEE Network September 1992, pg. 10. [Ill Steven E. Mi , % I SDN and Asynchronous Transfer Mode (ATM) , I Communications Magaune, [IO] M. Wanilr, 0 bOul-Magd, H. Oilbat, Tratltic - September 1989, ~g 17-57. n n ................................. ~ ..... ........... ..: ,: n ............. : ............. ....................... ..._............... .......................... 'i 1 I* In cmma omnsox.u,.w. Fig. 6d % of intact bursts arrived at destination as a function of the average 104 in the case of VLB, AT, DTAT (S1=14), with average bursts length IrlOO (nominal). ............ ..-A I. .......... :\ a i Cn.,s,.YI ............ .............. .............. : ............................ ............. 4 . I* 1 IWBYD 0m m mu I Flg. 6e. % of the cells anived in intact bursts as a function of the avg. load, in the case of VLB, AT, DTAT (S1=14), when average bursts length is LAO0 (nominal). I* I N. * mUOm.RlnUI.I* Fig. 61. 90 of he unmarked cells anived in damaged bursts as a function of the average load, in the case of VLB, AT, DTAT (S1=14), with IilO0 verage. bursts length. SINGAPORE ICCS1 94 534 0-7803-2046-8/94/$4.00 1 994 IEEE 1
Search
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks