TCP over ABR in ATM Networks with Variable ... - Semantic Scholar

5 downloads 0 Views 156KB Size Report
in the literature with the name `parking lot'. ..... forced to close by the backo algorithm, as indicated by the black diamond markers on the top side of the frame.
TCP over ABR in ATM Networks with Variable Topology and Background Trac  M.Ajmone Marsan, A.Bianco, R.Lo Cigno, M.Munafo Dipartimento di Elettronica, Politecnico di Torino { Italy E-mail: fajmone,bianco,locigno,[email protected] Abstract

The performance of TCP connections exploiting the ATM EFCI ABR transfer capability is investigated through simulation of two simple ATM network topologies, considering variable types of background trac. The rst network topology comprises only two ATM switches, connected by one link, that is the system bottleneck. The second topology comprises four nodes arranged in the layout generally known in the literature with the name `parking lot'. The background trac is taken to be either moderately bursty or extremely bursty, with ON-OFF characteristics. TCP connections are assumed to perform long le transfers, as in ftp application, during the whole simulation experiment, thus operating in overload conditions. ABR parameters such as RIF and RDF are varied in the simulation study in order to examine their in uence on the performance of TCP connections.

1 Introduction

The essence of the ABR (Available Bit Rate [1]) ATM transfer capability lies in the periodic insertion of resource management (RM) cells within the ow of data cells along the connection; RM cells travel from the source to the destination (along this forward path they are named `forward' RM cells), and they are then returned from the destination to the source (`backward' RM cells). The ATM switches along the connection can use RM cells to control the rate at which the source injects cells into the network, in order to both eciently exploit the available bandwidth and prevent congestion. The algorithm to be implemented within ATM switches for the control of ABR connections is not speci ed by standardization bodies. However, the recommended format of the RM cells hints at two possible techniques that can be chosen in order to convey the feedback information to the source: 1) the node can convey a very simple information through the Congestion Indication (CI) and No Increase (NI) bits, whose values can be combined in a three-state feedback corresponding to the indications \increase rate", \keep rate", or \decrease rate"; this technique is generally known as EFCI; 2) the node can use a speci c eld in the RM cell where the maximum rate at which the source is allowed to transmit is explicitly stated; this technique is named ER. Furthermore, the node can operate on the forward ow of RM cells, or on the backward ow, or on both ows. The aim of this work is the study of a simple EFCI ABR implementation running on networks with variable operational conditions. The impact of parameters such as the Rate Increase Factor (RIF) and the Rate Decrease Factor (RDF) on the performance of TCP connections is studied in detail, and a very simple algorithm is proposed to improve the network fairness while maintaining good performance. In order to study the bene ts provided by ABR, we implemented a simpli ed version of ABR in CLASS [2], an ATM network simulator developed at Politecnico di Torino under contract with CSELT. We concentrate in this paper on the ABR version based on the CI and NI bits, that are set in backward RM cells. The modi cation of the source transmission rate is obtained through an adaptive GCRA (Generic Cell Rate Algorithm [3]) shaping device, whose parameters can be dynamically adjusted on the basis of the informations carried in the CI and NI bits of the RM cells.  This work was supported in part by a research contract between Politecnico di Torino and CSELT, in part by the EC through the Copernicus project 1463 ATMIN, and in part by the Italian Ministry for University and Research.

1

In real networks, at the time of connection setup, ABR users negotiate the values for their shaping parameters and for their modi cation; in our implementation these parameters are selected in the simulation experiment description. These parameter comprise: RIF, RDF, the Peak Cell Rate (PCR), the Minimum Cell Rate (MCR), and the Initial Cell Rate (ICR). PCR and MCR de ne the maximum and minimum values of the ABR source rate that are never surpassed, regardless of the messages carried in the backward RM cells; RIF and RDF de ne the dynamics of the transmission speed modi cation; ICR in uences the behavior of the network when ABR connections initially start transmission, or re-start after a longer silence period. The rate of the transmission rate increase is linear (PCR/RIF), while both an exponential decrement (ACR/RDF) and a linear decrement (PCR/RDF) are considered.

The Congestion Control Algorithm

The control algorithm is local, without coordination among nodes, and is based on the node bu er occupancy. Two thresholds are associated with each bu er. Congestion is detected when the bu er occupancy surpasses the high threshold, while underutilization of resources is assumed when the bu er occupancy is below the low threshold.

Fairness Improvement

Since fairness among connections can be a critical issue in ABR, we also considered an algorithm that tries to achieve fairness among ABR connections by keeping track of the current rates of all the connections traversing the switch; when congestion is detected, the bandwidth allocated to connections is balanced using the following fairness metrics: ACR(i) ? MCR(i) (1) f (i) = PCR(i)

where i is the connection index, MCR(i) is the minimum cell rate negotiated by connection i, ACR(i) is the cell rate presently used by connection i and PCR(i) is the negotiated peak cell rate. The node also computes the average value of the fairness function over all connections (say N ) of the considered link: f =

X

1 N f (i) N i=1

If the bu er occupancy is greater than the high threshold, the CI bit is set only in the RM cells belonging to the connections for which f (i)  f , while the NI bit is set in the RM cells belonging to the connections for which f (i) < f . This algorithm, that will be called fair, is an extremely simple attempt to implement in a distributed manner the `Max-Min' fairness criterion described in [4] when connections have identical PCR values, and to extend the same approach to the case of connections with di erent negotiated PCR values.

2 The Simulation Scenarios We base our simulation experiments on the TCP source model available in CLASS. For details on the TCP implementation in CLASS, the reader is referred to [5]. The TCP source peak cell transmission rate is controlled via a UNI shaping device according to the feedback signals conveyed by the network, assuming that the cell delay variation tolerance (CDVT) negotiated for this shaper is always zero. We consider two network topologies: one comprising only two ATM switches connected by one channel that is the system bottleneck. This will be called the bottleneck topology. The second topology comprises four nodes arranged in the layout generally known in the literature with the name `parking lot'. In the considered topologies, the data rate on all channels, both user-node and node-node, is set to 150 Mbit/s, and the size of the bu ers inside ATM switches is set to 1000 cells. TCP connections are unidirectional: TCP transmitters send data segments, and TCP receivers return only ACK segments. The connections lengths are taken to be all equal to 1000 km. TCP sources always transmit segments of 9180 bytes, the suggested maximum segment size for TCP/IP over ATM. Two types of background trac are considered. In the rst case the background trac results from the segmentation of user messages generated according to a Poisson process, with a truncated geometric message 2

0.50 0.40 0.30 0.20 0.10 0.00 Background Traffic [Mbit/s] Unfair, L=10, H=500, RIF=2048, RDF=1024

60 55 50 45 40 35 30 25 20 15 10 5 0

0.30

Goodput [Mbit/s]

0.80

0.40

0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00

Fair, L=10, H=500, RIF=2048, RDF=1024

0.90

0.50

0.80

Background Traffic [Mbit/s]

1.00

0.60

0.90

0 10 20 30 40 50 60 70 80 90 100

Goodput Efficiency User 1 User 2 User 3

0.70

1.00

0.20 0.10 0.00

60 55 50 45 40 35 30 25 20 15 10 5 0

1.00 0.90 0.80 0.70 0.60 0.50 0.40

Efficiency

0.60

Efficiency

0.70

Goodput [Mbit/s]

0.80

60 55 50 45 40 35 30 25 20 15 10 5 0

Efficiency

0.90

0 10 20 30 40 50 60 70 80 90 100

Goodput [Mbit/s]

Fair, L=10, H=500, RIF=128, RDF=128 1.00

Efficiency

Goodput [Mbit/s]

Unfair, L=10, H=500, RIF=128, RDF=128 60 55 50 45 40 35 30 25 20 15 10 5 0

0.30 0.20 0.10 0.00

0 10 20 30 40 50 60 70 80 90 100

0 10 20 30 40 50 60 70 80 90 100

Background Traffic [Mbit/s]

Background Traffic [Mbit/s]

Figure 1: Goodput and eciency of 3 TCP connections in the bottleneck topology, with linear decrease rate and moderately bursty background trac length distribution whose mean is equal to 20 cells and whose maximum is equal to 200 cells. The background trac is then shaped with an allowed peak cell rate equal to 1.2 times the average cell rate, thus introducing only a small burstiness. We refer to this rst type of background trac with the term `moderately bursty'. In the second case, the background trac is generated by a single ON-OFF source, sending a CBR cell

ow during ON periods, with geometrically distributed ON and OFF-period lengths. The average duration of the OFF and ON periods is set so that on the average the background trac source transmits for 2/3 of the simulation time; the transmission rate during the ON period is set to the value that produces the required average background trac load. The use of such a bursty background trac allows us to focus phenomena that less severe background trac scenarios can hide. Moreover, with this bursty background trac, phenomena and behaviors typical of transient phases, such as the ability of sources to quickly increase and decrease their transmission rates, have a signi cant impact on the time-averaged performance indices.

3 Numerical Results Numerical results are presented as curves referring to two performance indices: the useful throughput, called goodput, at the TCP receivers, obtained considering the received data, but discarding all the faulty and the retransmitted segments, and the eciency of the TCP connections, i.e., the ratio between the goodput and 3

0.50 0.40 0.30 0.20 0.10 0.00 Background Traffic [Mbit/s] Fair, L=10, H=500, RIF=2048, RDF=128

60 55 50 45 40 35 30 25 20 15 10 5 0

0.30

Goodput [Mbit/s]

0.80

0.40

0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00

Fair, L=10, H=500, RIF=4096, RDF=128

0.90

0.50

0.80

Background Traffic [Mbit/s]

1.00

0.60

0.90

0 10 20 30 40 50 60 70 80 90 100

Goodput Efficiency User 1 User 2 User 3

0.70

1.00

0.20 0.10 0.00

60 55 50 45 40 35 30 25 20 15 10 5 0

1.00 0.90 0.80 0.70 0.60 0.50 0.40

Efficiency

0.60

Efficiency

0.70

Goodput [Mbit/s]

0.80

60 55 50 45 40 35 30 25 20 15 10 5 0

Efficiency

0.90

0 10 20 30 40 50 60 70 80 90 100

Goodput [Mbit/s]

Fair, L=10, H=500, RIF=2048, RDF=1024 1.00

Efficiency

Goodput [Mbit/s]

Fair, L=10, H=500, RIF=128, RDF=128 60 55 50 45 40 35 30 25 20 15 10 5 0

0.30 0.20 0.10 0.00

0 10 20 30 40 50 60 70 80 90 100

0 10 20 30 40 50 60 70 80 90 100

Background Traffic [Mbit/s]

Background Traffic [Mbit/s]

Figure 2: Goodput and eciency of 3 TCP connections in the bottleneck topology, with linear decrease rate and ON-OFF background trac the total o ered load of TCP connections. The goodput and eciency curves are plotted on the same charts; solid lines refer to goodput values, dashed lines to eciency results; a dot-dashed line plots the bandwidth available to each TCP connection. Fig. 1 reports a comparison between the unfair and fair control algorithms with 1000 km connections in the bottleneck topology with the moderately bursty trac (hence the least severe scenario under study); the decrease rate is linear. The two plots on the left refer to the unfair control algorithm, while the two plots on the right refer to the fair control algorithm. Plots in the higher and lower parts di er for the values of RIF and RDF (upper plots have RIF = RDF = 128; lower plots have RIF = 2048, RDF = 1024). The two plots on the left show that the unfair algorithm can indeed allow the control of ABR sources, but only with a very careful selection of the values for the increase and decrease factors; otherwise, extreme unfairness can result (upper left plot) and for most sets of values of the system parameters, one of the TCP connections is forced to close by the backo algorithm, as indicated by the black diamond markers on the top side of the frame. This untoward behavior even for this simplest scenario that we are considering raises serious doubts about the suitability of ABR switch algorithms that do not carefully consider the fairness issue. The two plots on the right instead show that the fair algorithm is capable of providing acceptable performance and fairness to the three TCP connections for a much wider range of values of the increase and decrease factors. Still, the results for zero background trac load in the top right plot show the danger of 4

Efficiency

Goodput [Mbit/s]

Fair, L=10, H=500, RIF=2048, RDF=128, MCR=1 60 1.00 55 0.90 50 0.80 45 40 0.70 35 0.60 30 0.50 25 0.40 20 0.30 15 0.20 10 0.10 5 0 0.00 0 10 20 30 40 50 60 70 80 90 100

Goodput Efficiency User 1 User 2 User 3

Background Traffic [Mbit/s]

Figure 3: Goodput and eciency of 3 TCP connections in the parking lot topology, with linear decrease rate and ON-OFF background trac a too fast rate increase and decrease. The loss in throughput and eciency in this case is in fact due to the following phenomenon. The three TCP connections do not start transmission exactly at the same time, so that each one initially nds the network very lightly loaded and can very quickly grow its window size. After a few transmission cycles the TCP windows grow large enough to allow continuous transmissions, and all connections transmit at the same time with very high speed; unfortunately, the ABR control mechanism is not fast enough to reduce their speeds before cells are lost. When cells are lost, TCP reduces its transmission window to 1, and the cycle starts again. It must be noted that this phenomenon is very similar to the synchronization e ect of TCP transmissions observed in standard packet networks [6]. Fig. 2 reports results for the bottleneck topology with ON-OFF background trac and the fair ABR control algorithm with linear decrease. In this more severe environment, the performance of the ABR control is less satisfactory and an accurate tuning of the system parameters must be adopted to avoid unacceptable behavior. Setting respectively RIF to 2048 and RDF to 128 seems to provide acceptable performance also in this scenario. Fig. 3 reports results for the parking lot topology with ON-OFF background trac and the fair ABR control algorithm with linear decrease. In this topology the ON-OFF background can be quite annoying for connections traversing several links. The RIF and RDF parameters are set to 2048, and 128, respectively. This situation is quite demanding for the ABR transport capability, but the curves show that also in this case satisfactory performance can be obtained. No signi cant bias can be observed against connections traversing several links. Fig. 4 reports results for the bottleneck topology with ON-OFF background trac and exponential decrease rate. It is interesting to observe that fairness control seems not necessary to obtain a fair behavior among connections; the optimal values of the RDF and RIF parameters have to be changed with respect to the case of linear decrease rate in order to obtain reasonable performance. It is interesting to note that no major di erence exists between the cases of exponential and linear decrease rates.

4 Conclusions The pervasive di usion of TCP in present networks is such that the investigation of its performance when used over ATM networks is a matter of paramount importance. Our simulation study aims at providing indications about the suitability of the ABR ATM transfer capability for le transfer applications based on TCP in di erent ATM network scenarios. Numerical results show that the TCP performance depends quite drastically on the burstiness of the background trac with which the TCP trac shares the transmission channels, and that the fairness issue is very important in the ABR context; a very careful choice of the ABR parameters is necessary to ensure 5

0.50 0.40 0.30 0.20 0.10 0.00 Background Traffic [Mbit/s] L=10, H=500, RIF=2048, RDF=32

60 55 50 45 40 35 30 25 20 15 10 5 0

0.30

Goodput [Mbit/s]

0.80

0.40

0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00

L=10, H=500, RIF=2048, RDF=128

0.90

0.50

0.80

Background Traffic [Mbit/s]

1.00

0.60

0.90

0 10 20 30 40 50 60 70 80 90 100

Goodput Efficiency User 1 User 2 User 3

0.70

1.00

0.20 0.10 0.00

60 55 50 45 40 35 30 25 20 15 10 5 0

1.00 0.90 0.80 0.70 0.60 0.50 0.40

Efficiency

0.60

Efficiency

0.70

Goodput [Mbit/s]

0.80

60 55 50 45 40 35 30 25 20 15 10 5 0

Efficiency

0.90

0 10 20 30 40 50 60 70 80 90 100

Goodput [Mbit/s]

L=10, H=500, RIF=512, RDF=8 1.00

Efficiency

Goodput [Mbit/s]

L=10, H=500, RIF=256, RDF=16 60 55 50 45 40 35 30 25 20 15 10 5 0

0.30 0.20 0.10 0.00

0 10 20 30 40 50 60 70 80 90 100

0 10 20 30 40 50 60 70 80 90 100

Background Traffic [Mbit/s]

Background Traffic [Mbit/s]

Figure 4: Goodput and eciency of 3 TCP connections in the bottleneck topology, with exponential decrease rate and ON-OFF background good performance.

References [1] ATM Forum/95-0013R8, \ATM Forum Trac Management Speci cation", Version 4.0, October 1995 [2] M.Ajmone Marsan, A.Bianco, T.V. Do, L.Jereb, R.Lo Cigno, M.Munafo \ATM Simulation with CLASS", Performance Evaluation, vol.24 (1995), pp.137-159 [3] ATM Forum, \ATM User-Network Interface Speci cation", First Edition, Prentice-Hall [4] D. Bertsekas, R. Gallager, \Data networks", Prentice-Hall, 1987 [5] M.Ajmone Marsan, A.Bianco, R.Lo Cigno, M.Munafo, \Shaping TCP Trac in ATM Networks", IEEE ICT'95, Bali, Indonesia, April 1995 [6] S. Floyd, Van Jacobson, \On Trac Phase E ects in Packet-Switched Gateways", Internetworking: Research and Experience, Vol. 3, n. 3, pp. 116-156, September 1992 6