Measurement based Connection Admission Control algorithm for ATM

Measurement based Connection Admission Control algorithm for ATM networks that use low level compression . Ioannis Papaefstathiou £ University of Cambridge, Computer Laboratory [email protected]

Abstract

1

Introduction

An important application of ATM is in the interconnections of private LANs, through WAN links. A typical such a network is shown in Figure 1. ATM is used as the backbone infrastructure in such networks, due to its scalability, speed and potential fault tolerance. One of the main characteristics of these networks is that the cost of the WAN links is comparatively high. By increasing the network traffic we can send over these links we can increase the effectiveness of the whole network. This can be done by compressing the data transmitted. In particular, ATM cells can be compressed and then encapsulated over newly formed standard ATM cells. In this paper we show the effectiveness (especially in this kind of interconnection networks) of this simple idea. The compression scheme we use is dynamic and understands ATM cell structures, it compresses only the payload of the cells, and adapts the payload technique depending on the type of traffic the different flows carry (i.e. whether they carry realtime, non-compressible traffic or not). In this way, each VC will rapidly acquire a compression context that suits its data. For example, traffic with a timing requirement and low compressibility (such as compressed video) will become transmitted uncompressed whereas compressible files, such as web

As it is widely supported, by compressing the cells that are transmitted over the links of an ATM network, you can significantly increase the effectiveness of the whole network. One of the main problems with networks that are using such low level compression, is that they need a special Connection Admission Control (CAC) Algorithm which will be capable of taking advantage of the increase of the effective bandwidth caused by the compression. In this paper such a CAC algorithm is presented and it is claimed that it can take almost full advantage of the compression gain. This CAC algorithm is based on Large Deviation Theory: the large deviation rate-function (entropy) of compressed (and thus more bursty) ATM traffic can be estimated from measurements of traffic activity. The entropy can be used to determine the bandwidth requirement of the traffic. Keywords: ATM, CAC, QoS, Lempel-Ziv.

Supported by a Marie Curie Research Training Grant under TMR activity 3

1

pages and wordprocessor documents will experience in other words the impact of a new call on the already admitted ones. In recent years, the notion of good compression. the effective bandwidth of a traffic source has been ATM ATM ATM BACKBONE developed to describe the effective resource requireCLOUD CLOUD WAN 000 111 ment of a source. Although different definitions of 111 000 ATM ATM ATM ATM 000 111 SWITCH WAN SWITCH SWITCH WAN SWITCH 111 1 000 0 000 111 the term exist, it has been widely used to express the LINK LINK 0 1 0 1 1 0 0111 1 000 111 000 000 111 trade-off between instantaneous bandwidth require111 000 000 111 ments and buffer space for a source (or collection of sources), given its traffic characteristics and QoS Points where compression devices constraints[7],[8], [9]. are placed The effective bandwidth of a source depends Figure 1: Typical application of our compression scheme strongly on the statistical properties of its traffic as well as its QoS requirements. It is thus essential In particular our algorithm works as follows : In to have an accurate characterization of each source the gateway of the LAN (where the WAN link is at- which contains precisely the information required to tached to) we try to collect all the cells that have the compute its effective bandwidth. Two approaches to same header, we remove all the headers and form obtaining this information are possible. The first, a long packet. Then and before we actually trans- which has been adopted to date by the majority of mit this long packet we apply our compression al- publications in the area, insists that the source should gorithm on it and then encapsulate the compressed provide the network with a full statistical characterstream into ATM cells. So for each packet we frag- ization of its behavior. Typically this is achieved by ment the compressed stream in 48 byte quantities constructing and analyzing a parametric model of the and attach to them the original header. By doing that source, based on precise or statistical descriptions of these newly formed cells are routed over the WAN its behavior. Parameters of the model may be fitnetwork just like the original ones and they are de- ted from actual measurements of the traffic, made in compressed in the gateway of the destination LAN. isolation[10],[11]. The major advantage of this approach in comparison The second approach argues that it is unreasonable with our older approach described in [2] ,[3] is that to expect traffic sources to provide these parameters. the WAN network involved can be a typical one (no Instead, its supporters argue that the network is capacompression units needed). A more detailed descrip- ble of deriving the traffic characteristics in a concise tion of the compression scheme can be found in [4]. and efficient way. This approach, based on on-line Since these newly formed cells are ordinary cells measurements of traffic activity, was pioneered by in terms of their format a standard ATM Call Admis- Courcoubetis et al.[12]. sion Control(CAC) algorithm can be used for deterBecause the behavior of the compressed traffic is mining whether or not to accept a new connection to much less predictable than the one of the original the network. data (since it depends heavily on the very difficult In the literature a lot of CAC algorithms are pro- to determine compression factor and the cross-talk posed. For all of them perhaps the most difficult as- between virtual circuits) we argue that the second pect is determining whether accepting a call will af- approach should be used in a network that uses our fect the existing traffic contracts for existing calls or compression scheme. In particular we argue that an 100Mb/s Ethernet

100Mb/s Ethernet

ATM Concentrators

ATM

Concentrators

100Mb/s Ethernet

100Mb/s Ethernet

2

Time

algorithm presented by Crosby et al.[13], and briefly described in the following sections, based on recent theoretical developments in the field of large deviations[14],[15] produces near optimal results. To briefly outline the CAC algorithm consider a network carrying a multiplex of compressed and uncompressed traffic connections, and receiving requests to set up new connections. These will have some simple user declared parameters associated with them (even if this is only the peak rate). This CAC algorithm combines estimation of the bandwidth requirement of the existing load using on-line measurement with a prediction of the bandwidth requirement of the offered traffic (after this traffic has been compressed) based on its declared parameters. If the sum of these bandwidths exceeds the line rate, the offered traffic is rejected; otherwise, it is accepted. Both the estimators and the predictors used make essential use of mathematical properties in the field of large deviations.

0

50

100

150

200

250

300

350

400

450

500

5500

Activity (cells/unit time)

5000

P’

4500

Bandwidth requirement


P

Capacity

CAC


4000

CAC

Accept()

3500

Accept()

Connect(peak= P’ )

3000

Connect(peak= P ) 0

Figure 2: Operation of the Measure CAC algorithm

taining a byte count and a history location. It is then possible for a decoder to reproduce this string exactly by copying it from the given location in its own history. If an incoming byte of data does not form part of a matching string, a one element string, containing this embedded value, is encoded and then transmitted to explicitly represent this byte. In other words it is a byte-oriented compression scheme. A decoder performs the inverse operation by first 2 Compression Algorithm parsing a compressed data stream in the two element strings (encoded string of more than one byte) and The algorithm we use for compression is a variaone element strings(encoding for just one character). tion of the LZ(Lempel Ziv) algorithm [16], which is The reasons for using this LZ algorithm are probably the most commonly used compression almainly the following : gorithm. In an encoder/decoder pair, at the two LAN gate¯ Since both the decoder and the encoder at any ways, there are two data structures which are initialgiven time maintain the same dictionary, there ized to the same known state, and they are updated in is no need for transmitting this large piece of an identical fashion. The encoder does this using the data from the encoder to the decoder. cells to be transmitted while the decoder, generates an identical stream of cells at the destination. ¯ It has an acceptable behavior for real-time data. The actual compression process consists of examining the data that are ready for transmission, to iden¯ It can be combined with a simple error recovery tify any sequences of strings of data bytes which alscheme for delivering reliable communication. ready exist in the encoder history. If an identical such In comparison with the more sophisticated LZhistory is available to a decoder, this matching string 78: can be encoded and output as a 2 element string, con3

¯ It produces much higher compression ratio fore the estimator has developed an accurate estimate of the new effective bandwidth, as shown in Figure 3, when applied to real network traces. the algorithm acts conservatively. It uses the most ¯ The decoding process is faster. recent stable estimate of the effective bandwidth of the multiplex, plus the sum of the peak rates of all ¯ It is easier to implement in hardware. subsequent calls. Thus, in Figure 3, the second call will be rejected, because the sum of the two peaks · ¼ and the first effective bandwidth estimate ex3 CAC Algorithm ceeds the link capacity. The technical specifications The CAC algorithm used, named Measure, makes of the implementation of the Measure algorithm are use of on-line measurements to develop an estimate described in [6] of the thermodynamic entropy of the traffic. Our approach is based upon the theory of Large DeviaCapacity P’ tions: the large deviations rate function, or entropy, of bursty traffic can be estimated from measurements P of traffic activity. The entropy can be used to determine the bandwidth requirement of the traffic, subject to user or network imposed Quality of Service P’ P (QoS) constraints. The key benefit of our approach is that a connection or flow need only be parameterised a straightforward way (for example by its peak rate). The behavior of Measure is shown graphically in Figure 2. Given a current multiplex of calls, with a Figure 3: Conservative CAC using the Measure alparticular (bursty) traffic activity (comprised of both gorithm compressed and uncompressed traffic), the system makes an estimate of the bandwidth requirement of the multiplex. This is obtained from the entropy estimator (depicted by a thermometer). A new call at- 4 CAC Simulation tempt declares its peak rate when requesting admission. The CAC algorithm essentially sums the This section presents the results of simulation peak rate , and the current estimate of the effective experiments using the proposed CAC algorithm. bandwidth; if the total is less than the link capac- The aim of the experiments was to evaluate the ity, then the connection can be accepted without vi- performance of our approach mainly with respect olating the QoS of any of the currently multiplexed to the number of calls our network can service calls. As soon as the new call commences, the es- without violating the QoS requirement. A mcuh timator will begin to revise its idea of the current more detailed description of the test environment resource requirement, producing in turn a new es- can be found in [5] timate of the bandwidth requirement of the sources. When the next call attempt arrives the procedure is Simulation Model. In each of our simulations we repeated, as shown. If a new call attempt arrives be- model a single output buffer and transmission link of Time (30 ms interval)

0

50

100

150

200

250

300

350

400

450

500

5500

5000

No. of cells

4500



4000

CAC

CAC

3500

REJECT()

Accept()

3000

Connect(peak=

0

4

)

Connect(peak=

)

28000

an ATM switch. The link speed used is 100 Mb/s, which corresponds to the TAXI transmission rate for the Fairisle ATM network [20] at Cambridge University. We present results for simulations using the compressed and the uncompressed http traffic1 . The compressed one was produced by the scheme described in [3]. The traces used are derived from real networks. The buffer size is 100 cells and the CLR constraint is ½¼ for all the results. The Peak Time (1/24 sec) rate for both types of traffic is 10Mb/sec whereas the Activity Level mean rate was 1.32Mb/sec for the uncompressed one and 0.8Mb/sec (uncompressed rate * compression Figure 4: Activity level (cells per interval) vs time ( sec) for ratio = 1.32/sec * 0.6) for the compressed one. Number of Cells per interval

26000 24000 22000 20000 18000 16000 14000 12000 10000

8000 6000

0

2e+08

4e+08

6e+08

8e+08

1e+09

1.2e+09

the HTTP uncompressed trace

Call Model. We study a scenario in which calls of a particular traffic type arrive according to an exponential inter-arrival time distribution, an assumption which appears to be well founded [21]. In the absence of real-world data we have used call lengths which are exponentially distributed. We present results for long calls. In both cases calls arrive at a high (Poisson) rate with mean 5 calls/s. Blocked calls are lost, but the high arrival rate means that the system is continually faced with new call attempts. We thus expect the system to remain close to maximum utilization. Calls have an exponentially distributed length with a mean of 60 seconds. Each accepted call transmits an independent trace. This trace is derived by randomly selecting a start point in the traffic trace. Calls are not correlated in any way. For the results which compare the performance of the different algorithms, the same random seed was used with each different algorithm, resulting in precisely the same call arrivals process in each case.

This allows us to compare not only the statistical properties of the algorithms, but also their dynamic behavior.

Results Before moving to the results of the Measure simulation the activity level2 of the compressed and the uncompressed traffic is displayed. The activity level is another representation of the arrival process mentioned as the basic input function of this CAC algorithm. As it can be seen in figures 4, and 4 the activity level of all the traces varies significantly with time. This is an indication that the task of the CAC algorithm is a difficult one. As the theory section describes, the effectiveness of this CAC mechanism depends mainly on how precisely it can measure the real utilisation of the network. If this measurement is accurate -and given that the algorithm acts always conservatively- the number of calls admitted will be 1 We chose to demonstrate the results of the http traces close to the theoretical maximum. throughout this paper since these trigger the worst case scenario Figure 4 shows results for cells in progress at an for our scheme (e.g. The other traces gave us better results with respect to all the issues discussed ).

2

5

The number of cells per time interval

18000

140

Number of Cells per interval

16000

120

Calls in progress

14000

100

12000

10000

8000

80

60

40 6000

20 4000

0 500

1000

1500

2000 0

2e+08

4e+08

6e+08

8e+08

1e+09

2000

2500

3000

3500

4000

Time (sec)

1.2e+09

Time (1/24 sec) Activity Level

Compressed Traffic Uncompressed Traffic Peak Allocation

Figure 5: Activity level (cells per interval) vs time ( sec) for

Compressed Traffic Optimal Limit Uncompressed Traffic Optimal Limit

Figure 6: Calls in progress for the http trace

the HTTP compressed trace

ATM network using the Measure CAC algorithm, over a period of about 4000 seconds. The graphs mainly show, for a given set of calls the number of them which could be accepted (a) using a CAC based on the call peak rate and (b) using the Measure CAC algorithm. It also demonstrates bounds on an optimal acceptance regime, which cannot be realised in practice. This bound, is the number of sources efficiently serviced if all transmitted smoothly at the mean rate of the entire trace.

histogram shows the effectiveness of our compression scheme when used together with the Measure Algorithm. As the graph shows the average number of calls in progress is approximately 54 or in other words 70/calls can be routed over the same network. Since this increase in the amount of data transmitted comes without any violation of the QoS the calls get, Finally figure 4 shows histograms of the average we argue that our scheme is extremely useful in real 3 number of connections in progress during a simu- ATM networks . lation run, for both the Measure algorithm and the Peak-Rate one. The left-most histogram in the plot was made using peak-rate admission control; this allows for an average of approximately just 10 connections in the system and it is the same for the compressed and the uncompressed data. The central histogram was made using the Measure algorithm applied to the uncompressed traffic. The advantage gained by exploiting statistical multiplexing 3 is clearly apparent as the average number of calls in Consider also that our results for the other types of traffic progress is now approximately 32. The right-most (IP,TCP etc) are even better! 6

1

[6] Andrew Moore and Raymond Russell, Evaluation of a Practical Connection Admission Control for ATM Networks Based on On-line Measurements, in preparation, 1999.

"measure.comp.cip.distn" "measure.no_comp.cip.distn" "peak.comp.cip.distn" "peak.no_comp.cip.distn" 0.8

0.6

[7] Joseph Y. Hui, Resource Allocation for Broadband Networks,IEEE JSAC, 6(9):1598–1608, December 1988.

0.4

0.2

[8] N.M. Mitrou and D.E. Pendarakis, Cell Level Statistical Multiplexing in ATM Networks: Analysis Dimensioning and Call Acceptance Control w.r.t QoS Criteria, In Broadband Technologies, New Jersey, October 1990.

0 0

10

20

30

40

50

60

70

Figure 7: Histograms of the average number of calls in progress over a simulation run.

[9] James Appleton, Modelling a Connection Acceptance Strategy for ATM Networks, In Broadband Technologies, New Jersey, October 1990.

References

[1] I. Papaefstathiou, Hardware implementation of [10] F.P. Kelly, Effective Bandwidths at Multi-Class Queues, Queueing Systems, 9:5-16, 1991. a Gb/sec network compressor, White Paper, http://www.cl.cam.ac.uk/ ip207/hard.ps. [11] R. Guerin and H. Ahmadi and Naghshineh, Equivalent Capacity and its application to [2] I. Papaefstathiou, Accelerating ATM : Onbandwidth allocation in high speed networks, line compression of ATM streams, 18th IEEE IEEE JSAC,9:968-981, 1991. IPCCC’99, Phoenix, Arizona, 10-12 February 1999.

[12] C. Courcoubetis and G. Kesidis and A. Ridder and J. Walrand and R. Weber,Admission Control and Routing in ATM Networks using Inferences from Measured Buffer Occupancy, Technical Report, University of California, Berkeley, EECS Department, UCB, California, CA94720, 1991.

[3] I. Papaefstathiou, Compressing ATM streams, IEEE Data Compression Conference 1999 (DCC’99), Utah, 29-31 March 1999 .

[4] I. Papaefstathiou, Complete Framework for low level, on-line ATM Compression, Submitted for publication, [13] Simon Crosby and Ian Leslie and John Lewis http://www.cl.cam.ac.uk/ ip207/comp.ps. and Raymond Russell and Fergal Toomey and Brian McGurk, Practical Connection Admis[5] Andrew Moore and Simon Crosby, An experision Control for ATM Networks Based on mental configuration for the evaluation of CAC On-line Measurements, Computer Communialgorithms, Performance Evaluation Review, cations, January 1998 . Vol 27:3, December 1999. 7

[14] N. G. Duffield and J. T. Lewis and Neil O’Connell and Raymond Russell and Fergal Toomey, The Entropy of an Arrivals Process: a Tool for Estimating QoS Parameters of ATM Traffic, Proceedings of the 11th UK Teletraffic Symposium, Cambridge, March 1994. [15] N.G. Duffield and J.T. Lewis and N. O’Connell and R. Russell and F. Toomey Entropy of ATM Traffic Streams, IEEE Journal on Selected Areas in Communications, Special issue on advances in the fundamentals of networking - part 1, 13(6), August 1995. [16] J. Ziv and A. Lempel, A Universal Algorithm for Sequential Data Compression, IEEE Trans. Information Theory, Vol IT-23, No 3, May 1978, pp 337-343. [17] Ecole Polytechnique Federale de Lausanne, Application REQUested IP over ATM, Technical Report, http://lrcwww.epfl.ch/arequipa/, July 1997. [18] E. Fiala and D. Greeve, Data Compression with Finite Windows, Communications of the ACM, Vol 32, No 4, April 1989, pp 490-505. [19] National Institute of Standards and Technology, NIST ATM/HFC Network Simulator: Operation and Programming Guide , March 1995. [20] I. M. Leslie and D. R. McAuley, Fairisle: An ATM Network,Computer Communications Review, 21(4):327-336, September 1991. [21] V. Paxson and S. Floyd, Wide-Area Traffic: The Failure of Poisson Modeling, Proceedings ACM SIGCOMM 94, London, UK, August 1994.

8