real-time bandwidth renegotiation through learnt behaviour

*Correspondence to: J. M. Gri$ths, Department of Electronic Engineering, Queen Mary and West"eld College,University of London, London E1 4NS, U.K.

�E-mail: [email protected]

Published online 1 June 2001 Received October 1999Copyright � 2001 John Wiley & Sons, Ltd. Accepted September 2000

INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMSInt. J. Commun. Syst. 2001; 14:549}560 (DOI: 10.1002/dac.492)

Real-time bandwidth renegotiation through learnt behaviour

J. M. Gri$ths*�� and L. G. Cuthbert

Department of Electronic Engineering, Queen Mary and Westxeld College, University of London, London E1 4NS, U.K.

SUMMARY

In many cases tra$c is aggregated from several sources, or simply has no well-understood characteristics,and so it is di$cult for the user to decide the basis on which bandwidth should be demanded. A strategy isdescribed to determine the optimumway to decide when to renegotiate the tra$c capacity provided in ATMand similar systems. The performance of the strategy is evaluated using simulations on tra$c withtheoretical distributions and also on recorded real tra$c. The strategy appears to work and the results giveinsights into the very limited bene"ts obtained by increasing the granularity of the rates provided. Copyright� 2001 John Wiley & Sons, Ltd.

KEY WORDS: ATM; IP; charging; negotiation; renegotiation; call acceptance control (CAC)

1. INTRODUCTION

Various bases have been proposed for charging for access to IP, ATM and similar networks.Most of them are characterized by the initial negotiation of some sort of bandwidth capacity, andcorresponding price, by the user. This capacity may be in the form of some e!ective bandwidths[1] which will determine the manner in which the user is charged or, at the other extreme, mightbe a simple peak-rate-limited scheme. In either case user needs may vary as the connectionprogresses and often explicit provision is made to renegotiate the initial parameters so that theymatch the changed needs of the customer. Capacity will either be negotiated up, to meet increasedneeds, or down, to reduce charges when lesser capacity is required. In this study, no assumption ismade as to the basis of the charges*they may be cost or market based.If no charge is made for renegotiation, then the decision when to renegotiate will be straight-

forward, but if a charge is made (or any other overhead is incurred) then it is more di$cult assavings from renegotiation will be o!set by the renegotiation charge. If the tra$c is understood indetail then a simple economic decision can be made on the known characteristics of the tra$c. If,on the other hand, little is known of the tra$c then the decision is more di$cult; an example iswhere tra$c is aggregated from several sources, such as IP users. Some previous work has

reported on this [2}10] but here we set out a more general strategy which is easily implementedand requires no knowledge of the underlying structure. The results we derive consider a case with"xed bandwidth channels, a linear charge for the bandwidth and a "xed renegotiation charge.However, the strategy is easily modi"ed to meet other situations. Throughout the paper we havenormalized the bandwidth to the peak capacity. In a "xed peak bandwidth environment, thismight be the 150 or 600 Mbit/s of an ATM system, a 2 Mbit/s DSL connection, or whateverbearer rate has been provided.

2. PRINCIPLE OF OPERATION

It is assumed that N equally spaced levels of bandwidth are available and that the cost ofproviding a particular bandwidth is proportional to the bandwidth itself; thus, if the bandwidthused is n, (1)n(N), then the cost per interval of time is also n/N. The assumption of equalspacing is not essential to the theory and will be considered later; however, it will be seen thatunequal spacing shows little advantage. The cost, C, of renegotiating the bandwidth is alsoexpressed in terms of the bandwidth available; thus if C"1 then the cost of renegotiation is thesame as the cost of transmitting the maximum bandwidth for an interval of time. The intervals oftime are not de"ned; the assumption is that the intervals are su$ciently long so that there is littlecorrelation between the bandwidth needs of one interval and those of the next; we assumed forour simulation that there is no correlation. However, real-tra$c information used clearly hassigni"cant correlation between intervals and the algorithm appears to continue to work satisfac-torily. The main constraint on the choice of interval length is the time required for therenegotiation to take place; intervals must necessarily be longer than the time required torenegotiate.The simplest renegotiation algorithm operates on the basis that the bandwidth selected will be

the bandwidth required. As already discussed, this is not optimum because in the case ofa decrease in the bandwidth required, it may be that the savings made may be less than the costincurred for renegotiation. To work out the bene"t (or otherwise) of renegotiating down toa lower rate, it is necessary to work out the probable saving. Let us consider the situation whenwehave bandwidth n

�and we will require for the next period, the lesser bandwidth n

�. The saving in

transmission cost per unit time interval will be (n�!n

�)/N; the cost per unit time interval of the

renegotiation will depend on how long we expect to stay at rate n�. We de"ne this &mean sojourn

time' at n�as S(n

�) and the mean cost per unit time is C/S(n

�). If (n

�!n

�)/N'C/S(n

�), it is worth

renegotiating the bandwidth down to the new level. However it may be more attractive torenegotiate the bandwidth down to some intermediate rate, n

�, (n

�'n

�'n

�), where the sojourn is

greater and the spreading of the renegotiation costs over a longer period maymore than o!set thelesser saving in bandwidth charges.A similar consideration may arise where an increase in bandwidth is required. In this case it

may be that the increase in rate selected should be beyond that demanded, to a rate which is morecommonly required and where the mean sojourn is thus longer. Once again the longer sojournwill mean that the cost of the renegotiation can be spread over a longer period and the lower costwill o!set the greater bandwidth charge. In the cases considered in this paper, the probability ofthis e!ect coming into play is not signi"cant and has not been taken into account, but it may wellbe that with certain types of tra$c this should be considered. We have simply assumed that whenan increase in bandwidth is required then that rate is negotiated.

550 J. M. GRIFFITHS AND L. G. CUTHBERT

Copyright � 2001 John Wiley & Sons, Ltd. Int. J. Commun. Syst. 2001; 14:549}560

The obvious question is how to "nd the mean sojourn time for each level. In the appendixa situation is considered where the probability distribution of the required bandwidths is knownand the sojourns may then be derived with relative case provided that there is no correlationbetween the intervals. However, although this gives signi"cant insight into the mechanism of theprocess, it is not particularly practical as a real-time mechanism. A more straightforwardapproach is possible as it is not actually necessary to know the basic bandwidth demandprobabilities; the sojourn times for each level are derived directly.This strategy can be implemented as follows: initially one assumes that the sojourn time is

in"nite at each level and hence it will always be worth changing the rate to follow the demands ofthe user. A record of the mean sojourn times at each level is kept and is updated continuously;decisions on whether to change bandwidth are always based on the updated sojourn table.Increases in the demands of the user will, as discussed above, have to be met.Simulations have been run to establish the viability of the strategy. Two main features of the

technique have been identi"ed:

1. There is no obvious &best way' of calculating the mean sojourn time. In our simulations wecounted the number of times that the level was moved to, X

�, and the total number of

intervals in which the system remained at that level, ¹�. The mean sojourn time was hence

S(n)"¹�/X

�. By starting with all ¹

�"1 and X

�"0, it automatically "xed the desired

starting conditions of in"nite sojourns. As the simulation progressed it became apparentthat it was more appropriate to take more account of recent events, so when X

�reached

some value X��, X

�was no longer incremented but instead ¹

�was multiplied by

X��/(X

��#1) every time a renegotiation to that level took place. While the allocated

bandwidth remained at level n,¹�was incremented in the usual way. For our simulations we

used a value of X��

of 100.2. It may happen that at an early stage in the operation of the algorithm, some low bandwidthlevel may have an untypically short mean sojourn entirely due to the random nature of therequests. The result will be that a transition to that level will never appear to be economi-cally justi"ed, transitions to that level will be avoided and hence it will never be discoveredthat the mean sojourn time at that level is in fact longer and hence, that level should be used.We overcame this e!ect by making a random decision that infrequently a downwardtransition should drop to the bandwidth required rather than the bandwidth which gives thegreatest economic return; we found that doing this on 1 per cent of occasions removes thisaberration with little impact on overall cost.

A major advantage of this technique is that the behaviour of the system automatically adjusts tothe changing statistics of tra$c. Implementation is possible in two ways; when a lower rate isrequested:

� either the system calculates the bene"t of shifting to all lower rates between the present rateand the required rate, "nding which (if any) rate o!ers the best return,

� or the system, as a background task, compiles a table of acceptable transitions. This could bedone somewhat in arrears resulting in less-than-optimal decisions in an environment oftra$c with time varying statistics, but reducing considerably the computational load.

As our simulations were not in real time we used the "rst of these techniques.

551REAL-TIME BANDWIDTH RENEGOTIATION


3. RESULTS

3.1. Theoretical distributions

The need to renegotiate will most likely arise in situations where the application requiresa relatively low level of bandwidth most of the time but occasionally requires more to accommod-ate a sudden demand. In previous papers [7}10] two distributions of demand probability werechosen where the probability of the level was inversely proportional to level and the level squared;unfortunately these distributions meant that it was not possible to compare directly the e!ect ofchanging the number of levels as any such change resulted in a change in the overall distributionof demand. To overcome this problem we have chosen for this paper an alternative. We haveassumed that the tra$c demand lies between 0 and 1 and the pdf is of the form

p(<)J(1!<)�

where � may be varied to change the nature of the probability distribution; for �"0 any level isequally probable and as � is increased, the probability of a high demand becomes ever lower. Theprobability that a particular actual level, n, will be required is given by the probability that theactual demand is between n/N and (n!1)/N.That is to say

p(n)"��

��p(< ) d<

For our simulations we used values of � of 1, 2 and 3. The simulations were run to investigate:

(a) The impact of the renegotiation charge on

(i) the overall cost of the connection,(ii) the overall level of the tra$c.

(b) The impact of the number of discrete levels of tra$c.

Figure 1 shows the results when tra$c with the amplitude probability distributions discussedabove are carried with renegotiation charges. In deriving each point 100 000 intervals wereconsidered.

� The top graph relates to tra$c which is fairly evenly spread over its possible amplitude rangebut with a tendency to use the lower levels (�"1); the bottom graph relates to tra$c inwhich the higher amplitudes are only very rarely used (�"3).

� In each case the lower set of curves illustrates how the mean tra$c rate rises as therenegotiation charge increases owing to the fact that the increased charge is a deterrent toreducing the rate negotiated. The upper set of curves shows how the charge goes up to carrythe tra$c due to the two e!ects, of the higher level of tra$c carrying capacity retained and ofthe costs of the renegotiations themselves. Both curves will be asymptotic to 1 as, ultimately,if renegotiation charges become large enough then it will never be worth seeking a lower rate!Note that the lower (tra$c level) curves are less smooth, owing to the variability of thesimulation, than the upper (cost level) curves; this is a topic which will be returned to later.



Figure 1. E!ect of renegotiation charge on mean cost and mean level.

Figure 2. Transient behaviour renegotiation cost"0.5.

� Each group of curves consists of the tra$c relating to 8 (top), 32, 128 and 512 (bottom) levels.The fact that the curves are so closely spaced indicates that the bene"ts of having a largenumber of levels is clearly quite limited.

Figure 2 gives an example of the transient response of the algorithm. The erratic line showshow during the course of the simulation, the cost is accumulating. The strobe marks are spaced at100 intervals. Initially the tra$c distribution has a value of �"3 and a renegotiation cost of 0.5.



Figure 3. ISABEL video tra$c.

A straight line has been drawn indicating the asymptote of the curve; the area A indicates that atstart-up the algorithm is adopting the strategy to minimize the cost. Instantaneously the tra$cdistribution is changed to the one with �"1 and the area B indicates the loss while the algorithmis readjusting its parameters. This (and more detailed study) has shown that the algorithm reachesa stable-cost behaviour in a few hundred intervals.

3.2. Real trazc

However realistic, it can be argued, the above theoretical tra$c might be, the real test is to applythe technique to real tra$c. From the project ISABEL [11] equipment, operating on the ACTSproject EXPERT [12] test-bed, a record of over an hour of real video tra$c was obtained. Itwould be normal for this sort of tra$c to be carried via a VBR service, but in this instance weinvestigated how this could be carried using renegotiated CBR. The record contained theinter-cell gaps of the video tra$c on a 155 Mbit/s link. This was analysed so that the mean tra$cover a range of intervals could be used as a basis for the algorithm. Intervals used ranged between30 ms and 3 s.Figure 3 shows the results of the simulation. The 155 Mbit/s stream was divided into 2048

levels. The actual mean rate required by the video stream was about 1.2 Mbit/s corresponding toa normalized rate of about 0.008 ("��

�) and embracing the lowest 30 levels of the 2048 available.

The interval lengths are set out in milliseconds. The left-hand graph shows how the mean levelchanged as the renegotiation cost changed; it can be said that no clear behaviour pattern is visibleand this will be discussed later. The right-hand graph shows how the cost increases as therenegotiation charge increases. In this case there is a clear trend with increasing renegotiationcost which is very similar, in form, to that displayed in the simulations using the theoreticaldistributions.We can also see that the longer the interval the less the impact of renegotiation charges on the

overall cost. This can be traced to the fact that with longer intervals, the #uctuations of the meanrate is less, hence the need for renegotiation is less and the renegotiation costs involved will alsobe less. In practice perhaps we can assume that renegotiation periods of much less than 1 s will beunreasonable, based on the premise that renegotiation will take a similar time to implement asrequired for initial call set-up, at present set at 0.3 s for ATM.

3.3. Discussion

One recurring theme in the last two sections has been the fact that the mean level of demandedbandwidth has been uncertain in the presence of a renegotiation charge and this has been the



Figure 4. E!ect of micro-changes in interval lengths for ISABEL tra$c.

subject of a qualitative investigation. What has been discovered during this study is that there isa very wide range of near optimal ways of accommodating the renegotiation charge. For exampleone strategy may be to renegotiate frequently by minimizing the tra$c charge; alternatively onemay renegotiate rarely and incur a higher tra$c charge. With a particular type of o!ered tra$c,both strategies may give a similar total cost result. Similarly it may be that the use of a particularset of (apparently preferred) levels may give a "nancial optimum; with the same tra$c, another setof levels may give a similar "nancial optimum. At one point we suspected that it was possible forthe algorithm to reach di!erent, stable operating conditions given the same tra$c statistics. Thequalitative reasoning behind this is that a particular level might fortuitously get a long-meansojourn period and so there would be frequent transitions to this level. As a result of this, otherlevels would tend to get shorter sojourn periods and so be less popular as levels to which to shift.Although this argument is persuasive, we must say that we cannot con"rm that the algorithm cansettle into particular stable modes; we have found that

� given a long-enough simulation the system will always settle to the single mode that can becalculated as shown in the appendix, although this may be approached by various routes;

� the mean cost of this "nal state will be negligibly better than the states leading up to it; inother words the optimum is very #at and the algorithm very rapidly tends to a "nanciallyattractive state but may take considerable time to reach the theoretical "nal state.

This then explains the results in Figure 1 where the mean level results show considerable variationwhile the cost curves are quite smooth. Similarly, rather chaotic mean level results of Figure 3 canbe reconciled with the well-behaved cost results. Some understanding of this may be derived fromthe result highlighted in Figure 4. This shows the algorithm applied to the ISABEL video tra$cwith a renegotiation cost of 0.01. The video tra$c is broken into intervals varying between 40 and50 ms. It can be seen that the mean cost does not #uctuate too much but there is a distinctdiscontinuity in the mean level of the tra$c between 45 and 46 ms indicating that the detailed beststrategy for achieving minimum cost changes at this point. It might be due to some correlationwithin the tra$c but this is not obviously related to, for example, the frame rates of video services.We should also discuss the evidence that there is little requirement for closely spaced levels. The

advantage of closely spaced levels is obviously the ability to match the users requirement and thebandwidth o!ered; however, even with only 10 levels, the bene"t would only be of the order of5 per cent and it is easy for this to be o!set by the renegotiation charges. One must also remember



Figure 5. Renegotiation with logarithmically spaced levels.

the overhead introduced by a large number of levels in calculating the sojourn times andcalculating the optimum transitions; with fewer levels it is easier to maintain on-line calculations,as suggested in Section 3, with the obvious bene"ts of more rapid response.

4. UNEQUAL SPACING OF LEVELS

The work above assumed an equal spacing of levels. It could be argued that as the tra$c requiresa low capacity most of the time, there would be greater economy if during these periods thebandwidth could more nearly be matched to the needs. This implies a closer spacing of levels atlow rates. We have, therefore, examined the case where the N levels available are logarithmicallydistributed between 1 and some low level �

��. Thus the probability that level �

�is required is

p (��)�

��

��

p(< ) d<

Figure 5 summarizes the results obtained with 32 logarithmically spaced levels at levels from��

"0.01 to 1. Comparison between these results and those shown in Figure 1 show noimprovement in performance.

5. CONCLUSIONS

The algorithm described above provides a practical and e!ective way of minimizing the costwhere there is an option of renegotiating channel rates in an ATM (and similar) environments. Itshould be noted that minimizing the cost does not imply minimizing the transmission rate andhence may result in a wide range of renegotiation and bandwidth utilization strategies, and theimpact on the tra$c rate required may vary.



Use of the theoretical distributions resulted in a doubling of the total cost if the renegotiationcharge is set at the average transmission cost of one interval of tra$c. In the case of the chosenreal tra$c, the impact was not as great; however, this is clearly less variable. The studies show thatit is not bene"cial to o!er excessively "ne granularity in the tra$c rates o!ered; a "gure ofbetween 8 and 30 appears adequate.

APPENDIX: THEORETICAL DERIVATION OF MEAN SOJOURNS

Let the probability that bandwidth n is required be p(n). Let us consider the situation where weare transmitting at the lowest rate 1. At the end of the unit interval, the probability u(1) thata move up to a higher rate will be required is simply

u (1)" ��

p(n)

and, the probability that a lower rate is required is zero, the mean sojourn at this level,S(1)"1/u(1).Considering the situation where bandwidth is being used is at the next rate up (i.e. 2), the

probability u(2) that a move to a higher rate will be needed in the next interval is once again givenby

u(2)" ��

p(n)

In this case there is also a probability d(2) that there will be a move down, to a lower rate at theend of the interval. However, in this case, it is not the simple probability p(1) because such a movewill take place only if it makes economic sense to do so. Thus a move from level 2 to 1 will savea cost of 1/N during each unit interval; on the other hand the renegotiation charge of C will haveto be written o! over the mean sojourn time S(1) and only if 1/N'C/S(1) will it be worth movingfrom level 2 to 1. Thus d(2) is either p(1) or 0 according to the cost of renegotiation, C, and thevalue of already calculated mean level 1 sojourn S(1). Thus the sojourn S(2)"1/(u(2)#d(2)).In general to "nd the mean sojourn at level m, S(m),

u(m)" ��

p(n)

d(m)" ��

p (n)z

where

z"

igjgk

1 if (m!x)/N'C/S(x) for any

value of x where m'x*n

0 otherwise

S(m)"1/(u(m)#d(m))



Table A1. Results for renegotiation cost of 0.4.

Worth droppingfrom level to

Probability MeanLevel �"3 sojourn Max. Min. Range

1 0.12 1.1 A2 0.11 1.33 0.098 1.54 0.088 1.7

5 0.079 2.0 B6 0.071 2.37 0.063 2.78 0.056 3.29 0.049 3.710 0.043 4.511 0.038 5.4

12 0.033 1.3 7 5 C13 0.028 1.1 10 5

14 0.024 1.1 11 5 D15 0.020 1.1 11 516 0.017 1.1 11 517 0.014 1.2 11 518 0.012 1.2 11 519 0.009 1.2 11 520 0.007 1.2 11 521 0.006 1.2 11 5

22 0.004 1.2 12 5 E23 0.003 1.2 12 524 0.002 1.2 12 525 0.0016 1.1 13 526 0.0010 1.1 14 527 0.0006 1.1 15 528 0.0003 1.1 16 529 0.00016 1.0 18 530 0.00006 1.0 19 531 0.000012 1.0 20 532 0.0000005 1.0 21 5

Clearly S(m) has to be evaluated in increasing values ofm, as the evaluation requires knowledgeof S(m) for the lower values of m. The term &z' is included to allow only those downward dropswhich make economic sense. Note that it is not assumed that a drop to the demanded value of thebandwidth is appropriate; it may make economic sense to drop to some higher value than thatdemanded, where the sojourn is longer and hence the renegotiation cost can be written o! overa longer period. In other words, if there is some level below the starting level, m, and above (andincluding) the "nishing level, n, to which it makes economic sense to drop, then the probabilityhas to be included that the demand for level n will result in a move.



Having calculated S(m) it is possible to calculate the range of changes in bandwidth which canbe economically justi"ed. Given a starting bandwidth level, m, then the highest level it is worthdropping to, n

��, is the highest value of n where (m!n)/N'C/S(n). The lowest level it is worth

dropping to, n��, is the value of n for which (m!n)/N!C/S(n) has the greatest positive value;

a drop to a lesser value of nmay still give a return but it will be smaller than that o!ered at n��. If

there are no values of nwhich satisfy these conditions then there is no justi"ed drop in bandwidth.Table A1 shows a typical form of the results obtained. The table shows the behaviour with 32

levels, a demand probability with �"3, and a renegotiation cost of 0.4. It can be seen that there isonly a signi"cant sojourn in bandwidths up to level 11; furthermore it is never "nancially worthdropping to levels 1}4 (range A) because the minimum level worth dropping to is never below 5;although sojourns can be calculated for layers 1}4, these are of no signi"cance. Thus most timewill be spent in levels 5}11 (range B). Range C is levels 12 and 13 where it is worth dropping onlyto the lower popular levels 5}7 and 5}10, respectively. In range D it is worth dropping to all theseven popular levels in range B. In the upper levels (range E) it is worth dropping to even shortsojourn levels just to save on bandwidth.The "gure of 0.4 and the value of � of 3 for the renegotiation cost are chosen as the results

demonstrate the whole range of behaviours possible. For lower "gures of renegotiation cost theranges D and A disappear and it is much easier to justify any decrease in bandwidth. For higher"gures of renegotiation cost, range E disappears as it is never economic to drop to a short sojournlevel just to save bandwidth.

ACKNOWLEDGEMENTS

This work was funded by the European Community ACTS Project AC014 CanCan*Contract Negotiationand Charging in ATM Networks. The authors would like to acknowledge useful ideas from Dee Denteneerof Philips Research Laboratories, Eindhoven, and the use of the tra$c information from the ISABELequipment on the EXPERT test bed.

REFERENCES

1. Kelly FP. Tari!s, policing and admission control for multiservice networks. 10th;K IEE ¹eletra.c Symposium, BTLaboratories, Martlesham Heath, April 1993.

2. Gerla M, Tai T-CC, Gallassi G. Internetting LANs and MANs to B-ISDNs for connectionless tra$c support. IEEEJournal on Selected Areas in Communication 1993; 1145}1159.

3. Grossglauser M, Keshav S, Tse D. RCBR: a simple and e$cient service for multiple time-scale tra$c. Proceedings ofthe ACM SIGCOMM +95, Boston, MA, August 1995; 219}230.

4. Reininger DJ, Raychaudhuri D, Hui JY. Bandwidth renegotiation for VBR video over ATM networks. IEEE Journalon Selected Areas in Communication 1996; 1076}1086.

5. Zhang H, Knightly EW. RED}VBR: a new approach to support VBR video in packet switching networks.Proceedings of the IEEE =orkshop on Network and Operating System Support for Digital Audio and <ideo (NOS-SDA<+95), 275}286.

6. Knightly EW, Zhang H. Connection admission control for RED-VBR, a renegotiation-based service. Proceedings ofthe IFIP I=QOS+96.

7. Gri$ths JM, Cuthbert LG. Impact of the resource needed for renegotiating ATM rates. IFIP 5th =orkshop onPerformance Modelling and Evaluation of A¹M Networks, 21}24 July 1997.

8. Pronk V, Denteneer D, Gri$ths JM, Cuthbert LG. Renegotiation strategies.CanCan deliverable 11 Source Models forA¹M Charging, Section 6.3, November 1997, 83}95.

9. Gri$ths JM, Cuthbert LG. Real time renegotiation of ATM bandwidth. IFIP 6th =orkshop on PerformanceModelling and Evaluation of A¹M Networks, 20}22 July, 1998.

10. Denteneer D, Pronk V, Gri$ths J, Cuthbert L. Impact of the resource needed for renegotiating ATM rates. ¹heInternational Journal of Computer Networks and ¹elecommunications Networking 2000; 34(1):211}225.



11. Quemada J et al. ISABEL: a CSCW application for the distribution of events. COS¹ 237=orkshop on MultimediaNetworks and Systems, Barcelona, November 1996.

12. EXPERT. http://elec.qmw.ac.uk/expert/intro.html.

AUTHORS' BIOGRAPHIES

John Gri7ths received a BSc at the University of Manchester in 1966. He thenworked at BT Laboratories on digital transmission in the trunk and access networks,leading to ISDN and Broadband Networks technology and performance. He is nowa Professorial Fellow at Queen Mary University of London. John is a Fellow of theIEE and a Senior Member of the IEEE.

Laurie Cuthbert is Head of the Department of Electronic Engineering at QueenMaryUniversity of London where he also leads the Telecoms Research Group. The Groupwas started in 1988 and is interested in IP, ATM, multimedia, mobile, charging,applications and intelligent control. Laurie is a member of the IEE professionalGroup on Telecommunications Networks and Services and takes an active interest inthe Institution.



real-time bandwidth renegotiation through learnt behaviour

Documents