using buffer management in 3g radio bearers to …ait.upct.es/~jjalcaraz/pub/ijwcmc_2009.pdf ·...

Int. J. Wireless and Mobile Computing, Vol. X, No. Y, XXXX

Copyright © 200X Inderscience Enterprises Ltd.

Using buffer management in 3G radio bearers to enhance end-to-end TCP performance

Juan J. Alcaraz* and Fernando Cerdán Department of Information Technologies and Communications, Technical University of Cartagena, Plaza del Hospital, 1, 30202, Spain Email: [email protected] Email: [email protected] *Corresponding author

Abstract: TCP performance over 3G links can be enhanced by means of Active Queue Management (AQM) techniques at the downlink buffer of the mobile network link layer (RLC). In this paper we describe Slope Based Discard (SBD), a novel deterministic buffer management technique with a discarding policy based on the observation of the filling rate of the buffer. Compared with the classic Random Early Detection (RED) scheme, SBD has a simpler operation and is best suited to multiplex short- and long-lived flows simultaneously. By means of extensive simulation experiments we compare the performance of TCP over three different RLC buffer management schemes: drop-tail (no AQM), SBD and RED. The simulation scenarios are configured to evaluate how the end-to-end goodput and delay are influenced by the number of multiplexed flows, the TCP flavour, the buffer size and the wireless channel error rate.

Keywords: TCP over 3G links; active queue management; radio link control; cross-layer design.

Reference to this paper should be made as follows: Alcaraz, J.J. and Cerdán, F. (XXXX) ‘Using buffer management in 3G radio bearers to enhance end-to-end TCP performance’, Int. J. Wireless and Mobile Computing, Vol. X, No. Y, pp.xx–xx.

Biographical notes: Juan J. Alcaraz obtained his engineering degree in telecommunication from the Polytechnic University of Valencia in 1999. His professional experience includes working at Siemens S.A. and at the Spanish 3G operator Xfera Móviles S.A. in the area of network design. He obtained his PhD in 2007 from the Polytechnic University of Cartagena, Spain. His research interests include cellular access networks design and performance evaluation.

Fernando Cerdán obtained his engineering degree in telecommunication in 1994 from the Polytechnic University of Catalonia and received his PhD from the same university in 2000. He is a Full Associate Professor at the Polytechnic University of Cartagena. Most of his published papers are in the field of service integration and QoS in wired IP networks, although his current interest includes security and QoS in 3G and heterogeneous networks.

1 Introduction

Third generation cellular networks (3G) provide internet access at data rates of up to 384 Kbit/s for wide area coverage. In contrast with wired networks, wireless links may suffer frequents frame losses due to signal corruption. The link layer protocol of the 3G radio access network, the Radio Link Control (RLC), can recover from frame losses by means of an Automatic Repeat reQuest (ARQ) algorithm, when configured in Acknowledged Mode (TS 24.322).

Many internet applications like e-mail, web surfing and file downloading rely on TCP for end-to-end data transport. TCP congestion control, originally developed for wired networks, considers a packet loss as an indication of congestion at the network routers that should be handled with a rate reduction in

the TCP source. To avoid bandwidth underuse in wireless connections, where propagation errors cause frequent frame losses (Tsaoussidis and Matta, 2002), the RLC protocol of 3G links should operate in Acknowledged Mode, providing a reliable radio bearer to TCP connections. Many previous works showed that a reliable link layer benefits TCP in this environment (e.g. Xylomenos and Polyzos, 2003). However, several characteristics of 3G radio bearers like high and variable latency and buffer overflow of the downlink buffers (Yavuz and Khafizov, 2002; Inamura et al., 2003; Chakravorty et al., 2005), have undesired effects on TCP performance.

There are several approaches to enhance the TCP–RLC interaction. Some of them, surveyed by Barakat et al. (2002) propose the modification of the TCP protocol itself or the introduction of an intermediate proxy (e.g. Meyer et al.,

J.J. Alcaraz and F. Cerdán

2003; Chakravorty et al., 2005). Many of these solutions are not practical in the short term because they may need changes on the protocol stack at the end users, and some proposals break the end-to-end semantics of TCP.

A less developed approach is the application of Active Queue Management (AQM) techniques at the downlink RLC buffers. Simulation experiments have shown that this solution improves TCP performance over 3G links (Agfors et al., 2003; Alcaraz et al., 2006) with a small change at the RLC operation in the Radio Network Controller (RNC) nodes. One of the most extended AQM mechanisms in the internet routers, Random Early Discard (RED) (Floyd and Jacobson, 1993), can be adapted to the particularities of 3G links (Alcaraz et al., 2006). However, Agfors et al. (2003) argues that the complex parameter configuration of RED in 3G links is a drawback that deterministic approaches may overcome. It should be mentioned that although RED is extensively used for internet routers, its implementation in wireless link layer buffers is relatively novel. Thus, the effect of its parameter configuration on TCP performance is still not fully known. In this paper we also contribute by providing further insight on RED parameter setting for this environment.

In this paper we describe and evaluate a novel deterministic AQM algorithm called Slope Based Discard (SBD) especially suitable to the characteristics of 3G radio bearers. By means of simulation, SBD is compared to RED and to the simple drop-tail scheme in several scenarios. The protocols involved are implemented in detail in the simulator which was developed and configured according to the recommendations of previous works. Analysing the results, we disclose the influence of TCP flavours, the number of concurrent TCP flows and the wireless error rate.

SBD deterministic operation and lower parameter sensitivity compared to RED makes it easier to implement in a highly variable environment like a 3G network. Moreover, SBD is best suited to multiplex long-lived and short-lived TCP flows together, thanks to its faster reactions against changes in the buffer occupancy.

The rest of the paper is organised as follows. Section 2 describes the characteristics of a 3G radio bearer that may degrade TCP performance. Section 3 surveys previous work about buffer management in radio bearers. Section 4 explains the SBD algorithm in detail. Section 5 provides a brief description of the simulation environment, compares the performance figures of RED, SBD and drop-tail schemes in several scenarios and provides configuration guidelines for each scheme. Section 6 deals with the performance of short- and long-lived flows multiplexed together at the RLC. The findings of this paper are summarised in Section 7 where future improvements are suggested.

2 Characteristics of 3G links

Previous experimental (Chakravorty et al., 2005) and simulation (Bestak et al., 2002) results provide a clear view of the characteristics of 3G wireless links. The behaviour of the link buffer occupancy has shown a great impact on TCP performance.

2.1 Buffer overflow

3G links employ per-user buffering. Flows going to a single mobile terminal share a single buffer. Therefore, radio bearers are expected to multiplex a number of simultaneous connections ranging form 1 to 4 TCP flows (Gurtov and Floyd, 2004). At a reliable RLC layer, the upper layer packets will be stored in the downlink buffer until they are fully acknowledged by the receiver side. The consequence is that, as described by Chakravorty et al. (2005) and Rossi et al. (2003), frame losses in the downlink channel result in higher buffer occupancy at the RLC network side. Considering that current RLC implementations use drop-tail buffers, the buffer may overflow causing consecutive packet losses. This situation is especially harmful in the first stages of a TCP connection (slow start) and has a higher impact in TCP Reno, which can only recover from consecutive losses with a Retransmission TimeOut (RTO). An RTO reduces TCP transmission window to one, causing the highest reduction of the source rate.

2.2 Overbuffering

The buffer should be large enough to avoid frequent overflow. However, excessive queueing cause some additional problems (Chakravorty et al., 2005) like Round Trip Time (RTT) inflation, unfairness between competing flows and viscous web surfing.

Figure 1 illustrates the effect of the buffer size on TCP performance over an RLC connection without AQM and a Frame Error Ratio (FER) in the wireless channel of 10%. The buffer size is given in RLC Service Data Units (SDU) of 1500 bytes. The figure shows the end-to-end goodput, defined as the number of successfully received packets at the receiver per time unit, and the delay of a 384 Kbit/s radio bearer in four different scenarios. Each scenario is defined by the number of simultaneous TCP flows at the user terminal, ranging from 1 to 4. We present the goodput and delay metrics for the most extended TCP flavours in the internet (Gurtov and Floyd, 2004): Reno, New Reno and SACK. Table 1 shows the parameter configuration for the RLC and TCP protocols. More details about the simulation environment are given in Section 5.

Using buffer management in 3G radio bearers

Figure 1 TCP performance over RLC. Goodput: (1) Reno, (2) New Reno (3) SACK. Delay: (4) TCP, (5) TCP Reno (6) SACK

210

240

270

300

330

30 40 50 60 70

Goo

dput

(K

bit/

s)

Buffer Size (SDU)

3 TCP flows

(1)(2)

(3)

30 40 50 60 70

0.5

1

1.5

2

2.5

Del

ay (

s)

Buffer Size (SDU)

4 TCP flows

(4)(5)

(6)

210

240

270

300

330G

oodp

ut (

Kbi

t/s)

1 TCP flow

0.5

1

1.5

2

2.5

Del

ay (

s)

2 TCP flows

Table 1 Simulation parameters

TCP parameters Setting Maximum TCP/IP packet size 1500 bytes Maximum allowed window 64 Kbytes Initial window 1 Fixed Round Trip Delay 200 ms 3G link parameters Setting PDU payload size 320 bits PDUs per TTI 12 TTI (Transm. Time Interval) 10 ms RB nominal bit rate 384 Kbit/s Transmission window 2047 PDUs maxDAT 15 In-order-delivery True Status Prohibit Timer 60 ms Missing PDU detection true Poll Timer 60 ms Wireless Round Trip Delay 50 ms Normalised doppler frequency 0,01 Poll window 50 % Last PDU in buffer Poll yes Last retransmitted PDU Poll yes

From Figure 1 we conclude that larger buffer sizes benefit the goodput performance, especially when the radio bearer only carries one TCP flow. On the other hand, an oversized buffer causes higher latency values. These results also show the improvement achieved by the more advanced TCP versions, New Reno and SACK, compared to TCP Reno. This better performance is in agreement with the recommendations of Inamura et al. (2003).

Figure 2 shows the trace of a TCP connection over a 384 Kbit/s radio bearer multiplexing a single TCP flow. The curve at the top shows the RLC buffer occupancy. The buffer size is limited to 60 SDUs. The TCP congestion window (cwnd) is depicted below, and the curve at the bottom shows the sequence number of the packets when they are sent, received and dropped.

At the first stages of the connection, multiple packets are dropped due to buffer overflow, triggering an RTO. The buffer is drained because of the rate reduction, with the consequent underutilisation of the resources. After that, the rate is recovered slowly and the overbuffering appears again, causing high delay and additional consecutive packet losses.


Figure 2 Trace of a TCP Reno connection over RLC

0

10

20

30

40

50

60

70

80

90

100

0 10 20 30 40 50 60 70 80 90 100

Sequ

ence

Num

ber

Time (s)data sent data rx drop

10 20 30 40 50 60 70

cwnd

0 10 20 30 40 50 60

BO

3 Related work

The objective of AQM techniques is to detect congestion before the buffer is completely filled, so that certain measures can be taken to avoid overflow. In general, these measures consist of sending a congestion signal (e.g. a packet discard) to the source in an early stage of congestion, allowing the TCP congestion control algorithm to reduce the data sending rate. AQM techniques were developed for internet routers, but now they are considered a feasible strategy to enhance TCP performance over 3G links because they are able to maintain the buffer occupancy around certain level, thus avoiding consecutive packet losses and reducing the delay. One of the most extended AQM schemes for internet routers is Random Early Detection (RED) (Floyd and Jacobson, 1993). Alcaraz et al. (2006) shows how RED can be adapted to 3G links. Previously, Agfors et al. (2003) had pointed out the possibility of using deterministic discard arguing the difficult parameter tuning of RED at an RLC buffer. Following this trend, we presented, in a shorter version of this paper (Alcaraz and Cerdán, 2006) an evaluation of a deterministic algorithm, SBD.

According to the useful definition given by Srivastava and Motanim (2005) the use of AQM at RLC to enhance TCP over 3G links is a clear case of cross-layer design, because it implies the design of an algorithm at a layer based on the details of how other layer operates. Section 6 addresses other cross-layer issues related to parameter configuration.

4 Slope-based discard algorithm

Slope-Based Discard (SBD) is a novel deterministic AQM algorithm that can be implemented at the downlink RLC buffer. SBD is based on the following ideas:

1 A packet discard is a congestion signal directed to the TCP sender side that takes a certain amount of time, TS, to arrive to the TCP source (see Figure 3). The rate reduction caused by this signal is perceived at the buffer after the propagation time, Tf, of the fixed (wired) network.

2 The RLC buffer occupancy process reflects the quality of the channel. When the frame loss rate at the channel increases, the rate at which the buffer is drained decreases.

3 The discarding policy is driven by the buffer filling rate, r. When r exceeds a critical value, rc, a packet will be drop. The buffer occupancy level determines the value of rc.

4 After a packet drop, additional packet discarding should be avoided until the rate reduction at the TCP source can be noticed at the RLC buffer. In Figure 3, this reaction time equals TS + Tf. Only after this period it is possible to know if the rate reduction is enough to prevent buffer overflow.

5 The value of rc represents the filling rate that, if sustained, would fill the buffer entirely in a time equal to the reaction time.


6 The packet chosen for discard will be as close as possible to the front of the queue, thus the signalling time TS does not include the queueing time, TQ. Additionally, the algorithm should not discard a packet whose transmission over the RLC link has already started. Otherwise, upon a packet discard, the RLC would start the signalling procedure required to synchronise RLC sender and receiver sides (TS 25.322, 2006). Consequently, our protocol discards the first packet whose transmission has not yet started, thus reducing complexity and avoiding changes in the 3GPP specification itself.

Figure 3 Schematic diagram of the end-to-end connection with the delay times of each section

Source

User

TS

TQ Tf

RLC entity

3G link

Fixed network Packets already sent in RLC frames that are still not acknowledged.

The following parameters control SBD operation:

• minth is the occupancy level above which packets can be dropped.

• maxth is the maximum occupancy allowed in the buffer.

• Tr is the estimated reaction time. • α is the occupancy interval used for the estimation of

the slope of BO process curve, r.

4.1 Description of the algorithm Figure 4 shows a graphical example of the algorithm. For the sake of clarity, the curve that represents the buffer occupancy is always increasing until a certain point, where it starts to decrease. In reality, there will be a fast oscillation around the average occupancy caused by the instantaneous arrival and departure of packets.

Figure 4 Buffer occupancy curve of a buffer implementing SBD

minth

maxth

time

α

DROP

TS

C0

th1

th2

th3

C1 C2 C3

th0

Tf

Buffe

r Occ

upan

cy

When the buffer occupancy (BO) is higher or equal to minth−α, the algorithm sets the value of the measuring interval threshold (thn = BO + α), and calculates with equation (1) the value of rc corresponding to this thn.

nc

R

maxth thr

T−

= (1)

The time Cn that it will take the buffer to store α additional bits at rc is Cn = α/rc. A timer for Cn is started. If the timer expires and the buffer occupancy is below the threshold thn, then the actual filling rate is lower than rc, and no packet is dropped. On the other hand, if the threshold thn is reached before the expiration of the timer, the current filling rate surpasses rc and therefore a packet will be discarded.

In Figure 4, the dotted segments starting at each measuring interval represent the buffer filling at the critical rate (critical curve). In the intervals C0, C1 and C2, no packet is discarded because the filling curve is below the critical curve. In contrast, in the C3 period, the filling rate is above rc. Hence, the threshold th3 is reached before the timer C3 expires. When th3 is reached, a packet is dropped. The monitoring and discarding algorithm is deactivated for a period Tr, avoiding consecutive packet discards. A description of the algorithm in pseudocode can be seen in Figure 5.

Figure 5 Pseudocode of SBD algorithm

for each packet arrival update BO if timer Cn on

if (BO ≥ thn) drop packet deactivate timer Cn start timer Tr

if timer Cn off if (BO ≥ minth − α)

calculate Cn thn = BO + α start timer Cn

upon timer (Cn or Tr) expiration

if (BO ≥ minth − α) calculate Cn thn = BO + α start timer Cn

4.2 Implementation issues

One of the main advantages of SBD compared to other RED-like mechanisms is that it does not need to generate random numbers to compute the discarding probability because of its deterministic operation. This reduces the computational cost of the algorithm, and makes it more feasible for its implementation at the RLC level where the buffering is done in a per-user basis.

The maintenance of a new timer, Cn, does not add too much complexity to the RLC operation, which already handles several timers, e.g. Poll Timer and Status Prohibit


Timer (TS 25.322, 2006). The RLC can synchronise Cn to the Transmission Time Interval (TTI), which is equivalent to a clock signal with a granularity of 10 ms (similar to that of other RLC timers).

The protocol could be further simplified in terms of implementation with the use of pre-calculated Cn values. Therefore, the buffer could be considered as divided in gaps, each one having a corresponding Cn value. In our simulations, the number of Cn values equals the number of SDUs fitting in the buffer space delimited by minth and maxth.

5 Performance evaluation

5.1 Simulation environment

The simulation environment employed in this research was developed in OMNeT++ (www.omnetpp.org) and comprises a complete implementation of TCP and RLC protocols. Similar simulators were described by Bestak et al. (2002) and Rossi et al. (2003). The simulation scenario consists of one or several TCP sources connected to their respective receivers in a mobile equipment (see Figure 6). The end-to-end connection consists of two sections, the fixed network and the radio bearer. The fixed network considers the internet and the 3G core network. The radio bearer has a round trip time of 50 ms (Holma and Toskala, 2004) and a bidirectional nominal rate of 384 Kbit/s, representing the bottleneck link, which is the situation expected in most cases (Gurtov and Floyd, 2004). The fixed network is modelled with a fixed rate of 1 Mb/s and a round trip delay of 200 ms.

Figure 6 Schematic representation of the simulation environment

Wireless Channel

RLC

Frame Error Rate: 10% - 15% - 20% Normalizad Doppler Freq.: 0.01

Network Side

User Side

RLC

TCP sources

#1

TCP receivers

#3 #2 #4

#1 #3 #2 #4

RTT Internet: 200 ms

The wireless channel generates error bursts according to the model described by Chockalingam and Zorzi (1999) where the Doppler frequency, fd, of the wireless link determines the average burst length. Lower fd causes longer bursts of errors. It is usual to employ the normalised Doppler frequency, equal to the product of fd and the radio frame duration (10 ms).

In order to obtain more realistic results, the error probability in our model is the same in the uplink and in the downlink direction. The FER is fixed to 10%, a typical UMTS

design value (Meyer et al., 2003), except for scenarios evaluating the effect of higher FER values (see Section 6).

The simulation results are obtained averaging 20 runs per sample. Each run is a 60-s download session, except for the performance evaluation of short-lived TCP flows (see Section 7). The radiuses of the confidence intervals are obtained with a confidence degree of 90% according to a t-student distribution.

RLC and TCP configuration is shown in Table 1. The description of RLC parameters can be found in (TS 25.322, 2006). RLC parameter values have a considerable impact on TCP performance. Our RLC configuration follows the guidelines provided by previous works (Rossi et al., 2003; Alcaraz et al., 2006) to achieve the best performance at the RLC layer and to avoid protocol stalling.

5.2 Simulation scenarios

Our performance simulation analysis compares two metrics, goodput and delay, for the simple drop-tail scheme (DT) and two AQM mechanisms: RED and SBD. Each AQM scheme was evaluated for several combinations of parameter values, in order to find the optimum configuration of the algorithms in terms of end-to-end performance in each scenario.

For the simulations implementing the RED algorithm, the following values for RED parameters were evaluated:

• minth: from 5 to 55 SDUs in steps of 5 SDUs.

• maxp: from 0.005 to 1.

It should be mentioned that, as stated by Agfors et al. (2003), an RED implementation at the RLC layer should not perform buffer occupancy averaging. Using the instantaneous queue size instead makes the reaction of the algorithm to sudden increases faster, simplifies its configuration and reduces computing load.

For the simulations with SBD buffer, the SBD parameter values evaluated are:

• minth: from 5 to 55 SDUs in steps of 5 SDUs.

• Tr: from 50 ms to 900 ms.

• α: from 1 to 10 SDUs.

The different scenarios are generated with the combination of the following values of the simulator parameters:

• TCP flows multiplexed: 1, 2, 3 and 4 TCP flows.

• TCP flavours: Reno and New Reno.

• FER in the wireless channel: 10%, 15% and 20%.

The RLC buffer size is fixed to 60 SDUs (90 Kb). For most AQM configurations, this size is enough to prevent buffer overflow.

5.3 SBD parameter setting

We start by selecting the value of α. Figure 7, shows the goodput and delay performance for different values of α, minth = 20 and Tr = 250 ms. Lower values of α imply a fast detection of congestion but excessive sensitivity to occupancy oscillations, causing more packet drops than


needed. If α is too large, the reaction against buffer increments is too slow and therefore is less effective against overbuffering. For the rest of the experiments in this paper, α is set to 5, which is a good compromise value.

Figure 7 Influence of α on SBD performance

50

100

150

200

250

300

350

0 2 4 6 8 10 0

0.25

0.5

0.75

1

1.25

1.5

1.75

2

Goo

dput

(K

bit/s

)

End

-to-

end

dela

y (s

)

α

goodp. flows=1goodp. flows=2goodp. flows=3goodp. flows=4

delay flows=1delay flows=2delay flows=3delay flows=4

We define the aggressiveness of the algorithm as the tendency to reduce the source’s rate given certain BO level. It is clear that, if the protocol discards a packet at low BO values (low minth) or if Tr is large (rc value is small), the aggressiveness is high.

Figure 8 shows the performance figures for one and several TCP Reno connections to illustrate the effect of Tr in SBD (minth = 20). Higher Tr values yield lower packet latency, because the buffer occupancy is kept at lower levels. If Tr is too small, the algorithm operation approaches a drop-tail scheme. Considering that the goodput for a single TCP Reno flow over a drop-tail RLC buffer is 276.8 ± 4.1 Kbit/s and the delay is 0.72 ± 0.02 s, there is a range of configuration values for SBD where the goodput and the delay performance are simultaneously improved in the single flow scenario. The goodput improvement is more noticeable for Tr values near the Round Trip Propagation Delay of the connection. If Tr is too high, the frequency of packet drops is excessive, causing goodput degradation.

If the number of concurrent TCP connections in the RLC buffer increases, the effect of Tr is slightly different. If the number of multiplexed flows augments, the goodput reduction is less severe for aggressive SBD configurations. For example, within the range of Tr values considered in Figure 8, the goodput is almost constant for the case of four TCP connections. This fact is easily explained considering TCP cwnd dynamics. Using AQM, an early packet discard halves the window of the TCP connection. Obviously, in a single flow scenario this measure halves the overall user rate. This avoids buffer overflow but limits the goodput improvement if packet drops are too frequent. In a multiple flow scenario the overall user rate reduction is less severe because it only affects one connection upon each packet discard. In the multiple flow scenario, the goodput improvement compared to that of drop-tail is less noticeable, but the delay reduction can be of up to 40%.

Figure 8 Influence of Tr on SBD performance

50

100

150

200

250

300

350

100 200 300 400 500 600 700 800 900 1000 0

0.25

0.5

0.75

1

1.25

1.5

1.75

2

Goo

dput

(K

bit/

s)

End

-to-

end

dela

y (s

)

Tr (ms)



In SBD, the impact of minth is lower than in RED. Figure 9 shows the effect of minth (Tr = 250 ms). In this figure, the goodput is almost constant. The reason is that SBD discards packets if the buffer’s filling rate surpasses the critical rate, rc, on each BO level. Assuming that there is a maximum rate rmax, at which packets directed to a single user can arrive to the RNC, the maximum buffer filling rate is equal to rmax. Therefore, although BO is above minth, no packet is discarded if rc is higher than rmax, i.e. if rc can not be reached. Tr, determines the BO value where rc equals rmax. We may consider this BO value as the effective minimum buffer threshold, mintheff, above which a packet discard may occur (see Figure 10). For Tr = 250 ms, and rmax = 1000 Kbit/s (83 SDU/s), according to (1), mintheff = maxth − Tr·rc = 39 SDUs. For more aggressive Tr values, minth acts as as ‘security’ limit, assuring that the goodput does not fall bellow certain threshold.

Figure 9 Influence of minth on SBD performance

50

100

150

200

250

300

350

0 10 20 30 40 50 60 0

0.25

0.5

0.75

1

1.25

1.5

1.75

2

Goo

dput

(K

bit/

s)

End

-to-

end

dela

y (s

)

minth (SDUs)




Figure 10 Critical filling rate rc respect the Buffer Occupancy

maxthminth mintheff

No packet is dropped in this area

rc

rmax

BO

5.4 RED parameter setting

In RED, the aggressiveness is determined by the combination of maxp and minth. Figure 11 shows the performance of RED respect maxp, with minth = 30. As in the SBD case, there are several configuration values that improve goodput while reducing the delay respect to those of a drop-tail scheme. For the single flow scenario, these values are around maxp = 0.05. When more connections are multiplexed, maxp can be set to higher values. Like in SBD, when more flows are multiplexed, a more aggressive setting provides a better overall performance.

Figure 11 Influence of maxp on RED performance

50

100

150

200

250

300

350

1 0.8 0.5 0.2 0.1 0.05 0.02 0.01 0.005 0

0.25

0.5

0.75

1

1.25

1.5

1.75

2

Goo

dput

(K

bit/

s)

End

-to-

end

dela

y (s

)

maxp



Figure 12 shows the influence of minth in RED performance for maxp = 0.05. It can be seen that RED performance is more dependant on minth than SBD, especially for the single-flow case. In this case, minth determines the minimum cwnd to allow a packet discard. To avoid link underuse, minth should be above the Bandwidth Delay Product (BDP) of the connection, defined as the product of the Round Trip Propagation Delay (RTPD = RTTw + RTTf)

with the bottleneck link rate. Our results show that the optimum minth value in terms of goodput for a single TCP flow is around 4 × BDP.

Figure 12 Influence of minth on RED performance

50

100

150

200

250

300

350

0 10 20 30 40 50 60 0

0.25

0.5

0.75

1

1.25

1.5

1.75

2

Goo

dput

(K

bit/s

)

End

-to-

end

dela

y (s

)

minth (SDUs)



5.5 Performance comparison In order to compare the performance of both AQM schemes and evaluate its improvement respect to the drop tail buffer, we have selected different parameter settings regarding the number of multiplexed TCP flows and the FER. These configuration values, shown in Table 2, do not provide, in general, the highest goodput in each scenario, because our objective is to balance goodput and delay improvement. For this reason, we choose configurations providing a goodput near the optimum (98% or more) with the lowest delay.

Table 2 AQM parameter setting for each scenario

FlowsFER

1 2 3 4

10% 20, 250 20, 250 20, 300 20, 500 15% 20, 250 20, 250 20, 300 20, 500 SDB*

20% 25, 200 25, 200 25, 300 25, 300 10% 30, 0.05 20, 0.05 15, 0.05 15, 0.05 RED** 15% 30, 0.05 25, 0.08 25, 0.08 20, 0.08 20% 25, 0.02 25, 0.05 25, 0.05 25, 0.08

Notes: *SBD parameter values: minth (SDUs), Ts (ms). **RED parameter values: minth (SDUs), maxp.

Figure 13, shows that the improvement in goodput and delay is noticeable for the case of TCP Reno, while for TCP New Reno the improvement is focused on the delay reduction. TCP New Reno is less affected by buffer overflow than TCP Reno. Therefore, the avoidance of consecutive packets losses helps especially Reno sources. These results are extensible to TCP SACK, which operation is similar to that of TCP New Reno.


Figure 14 shows the performance of TCP Reno for different FERs. Obviously, higher FER values cause worse performance in every scenario. AQM buffers helps to increase the goodput even for higher FER values, although the main improvement is the delay reduction. AQM algorithms counteract the buffer occupancy increment caused by a higher frame loss ratio in the

channel. However, it is advisable to relax slightly the aggressiveness of AQM schemes, as shown in Table 2. Obviously, the average BO tends to increase for higher FER. In consequence, an AQM scheme should reduce its tendency to discard a packet at each BO level in order to avoid excessive packet drops.

Figure 13 Performance figures of different buffer schemes for different scenarios (FER = 10%)

240

260

280

300

320

340

360

4 flows3 flows2 flows1 flow

Goo

dput

(K

bit/

s)

(a) Goodput TCP Reno

DTREDSBD

240

260

280

300

320

340

360


Goo

dput

(K

bit/

s)

(b) Goodput TCP New Reno

DTREDSBD

0.4 0.5 0.6 0.7 0.8 0.9

1 1.1 1.2 1.3


Del

ay (

s)

(c) Delay TCP Reno

DTREDSBD

0.4 0.5 0.6 0.7 0.8 0.9

1 1.1 1.2 1.3


Del

ay (

s)

(d) Delay TCP New Reno

DTREDSBD

Figure 14 Performance figures of different buffer schemes for different FERs in the radio bearer

250 260 270 280 290 300 310 320 330 340


Goo

dput

(K

bit/s

)

(a) FER 10%

DTREDSBD

200

220

240

260

280

300

320


Goo

dput

(K

bit/s

)

(b) FER 15%

DTREDSBD

160 180 200 220 240 260 280 300


Goo

dput

(K

bit/s

)

(c) FER 20%

DTREDSBD

0.4 0.5 0.6 0.7 0.8 0.9

1 1.1 1.2 1.3


Del

ay (

s)

(d) FER 10%

DTREDSBD

0.4 0.5 0.6 0.7 0.8 0.9

1 1.1 1.2 1.3


Del

ay (

s)

(e) FER 15%

DTREDSBD

0.4 0.5 0.6 0.7 0.8 0.9

1 1.1 1.2 1.3


Del

ay (

s)

(f) FER 20%

DTREDSBD


It should be remarked that the configuration proposal of this section relies on a cross-layer approach. According to Srivastava and Motanim (2005), the joint tuning of parameters across layers is a type of cross-layer design. The configuration of AQM based on the FER requires an upward information flow from the physical layer. In the 3GPP architecture, this is already possible thanks to the Radio Resource Control (RRC) protocol that handles the resource management algorithms, setting up and modifying layer 1 and layer 2 protocol entities (TS 25.331, 2006). On the other hand, the number of multiplexed flows can be provided in a downward information flow from the Packet Data Convergence Protocol (PDCP) which is above RLC (TS 25.232, 2006). It should be noticed that these information flows do not require additional signalling because they operate only on the RNC side, and therefore could be implemented without introducing changes in 3GPP specifications.

6 Performance with short-lived flows

In this section we show how each scheme performs when long- and short-lived flows are multiplexed simultaneously in the RLC buffer. The scenario for the simulations consists of a TCP connection that is continuously transmitting (long-lived flow). After 40 s, the second TCP connection starts the transference of a 64 kB file (short-lived flow).

Considering Reno sources, the file’s transfer time using a drop-tail buffer is 35 ± 2.99 s. Figure 15 shows the transfer time for different SBD configurations and Figure 16 considers RED. The average transfer time of drop-tail (DT) configuration is shown as a reference value. It is clear that SBD outperforms RED in this task. If Tr is equal or above 250 ms (as recommended in Section 6) the transfer time is below that of most RED configurations. The reason is that SBD is more responsive against sudden increases in BO.

Figure 15 Transfer time for a 64 Kb file with SBD

20

25

30

35

40

45

50

0 200 400 600 800 1000 1200

Tra

nsfe

r T

ime

(s)

Ts (ms)

DTminth=15

minth=20minth=25

minth=30

In order to illustrate this, Figure 17 shows a trace of one run of the experiment using RED and Figure 18 shows a trace

for SBD with the same channel conditions. Both RED and SBD are configured for the one flow scenario. As can be seen, RED is not able to avoid buffer overflow when the short transfer starts, which suffers multiple packet drops, with the consequent RTO. In SBD, which is conceived to react against BO increments, the number of packet drops during the first stages of the second connection is higher. Some of these packet discards affect to the new flow which reduces its rate preventing buffer overflow.

Figure 16 Transfer time for a 64 Kb file with RED

20

25

30

35

40

45

50

1 0.8 0.5 0.2 0.1 0.05 0.02 0.01 0.005

Tra

nsfe

r T

ime

(s)

maxp

DTminth=15

minth=20minth=25

minth=30

Figure 17 Transference of a 64 Kb file over a RED buffer with a pre-existing connection

0

10

20

30

40

50

60

70

80

90

100

0 10 20 30 40 50 60 70 80 90 100

SN F

low

1


0 10 20 30 40 50 60 70 80 90

100

SN F

low

2

10 20 30 40 50 60 70

cwnd

1

10 20 30 40 50 60

cwnd

2

0 10 20 30 40 50 60

BO


Figure 18 Transference of a 64 Kb file over a SBD buffer with a pre-existing connection

0

10

20

30

40

50

60

70

80

90

100

0 10 20 30 40 50 60 70 80 90 100

SN

Flo

w 1


0 10 20 30 40 50 60 70 80 90

100

SN

Flo

w 2

10 20 30 40 50 60 70

cwnd

1

10 20 30 40

cwnd

2

0 10 20 30 40 50 60

BO

7 Conclusion and future work

By means of simulation experiments we show that AQM techniques implemented at RLC downlink buffers improve goodput and delay performance of TCP over 3G radio bearers. The relative improvement of AQM schemes respect to drop-tail depends on several features. For example, AQM avoidance of buffer overflow benefits especially the goodput performance of TCP Reno because New Reno and SACK are more robust to consecutive packet losses. On the other hand, the delay is reduced in every scenario because AQM maintains the buffer occupancy around a certain level. AQM is more effective for large buffers and when the radio bearer multiplexes more than one TCP flow.

SBD performance is similar to that of RED, although in certain scenarios SBD show better figures. The main advantages of SBD are its easier implementation, its lower parameter sensitivity and its better performance for short-lived flows when multiplexed with long-lived connections. SBD is deterministic; hence it does not need the generation of random numbers required for the computation of a discarding probability. This fact makes SBD more feasible for buffer management at RLC level, where the buffering is done in a per-user basis.

This simplicity can be applied in the design of automatic configuration algorithms for SBD. The wireless links are highly variable; the delay and the bandwidth can experience sudden changes due to several mobile network phenomena, e.g. handovers, outages and scheduling. TCP-based applications may have different QoS objectives (e.g. higher goodput or less delay). Following a cross-layer approach, a dynamic algorithm could rearrange SBD parameters adapting it to the changing link characteristics while meeting the application requirements.

Acknowledgements

This work was supported by project grant TEC2007-67966-01/TCM (CON-PARTE-1) and it was also developed in the framework of ‘Programa de Ayudas a Grupos de Excelencia de la Región de Murcia, Fundación Seneca’.

References 3GPP TS 25.322 (2006) ‘Radio Link Control (RLC) protocol

specification’, v. 6.7.0. 3GPP TS 25.331 (2006) ‘Radio Resource Control (RRC) protocol

specification’, v. 6.9.0 3GPP TS 25.232 (2006) ‘Packet Data Convergence Protocol

(PDCP) specification’, v. 6.5.0. Agfors, M., Ludwig, R., Meyer, M. and Peisa, J. (2003) ‘Queue

management for TCP traffic over 3G links’, Proceedings of IEEE WCNC 2003, Vol. 3, pp.1663–1668.

Alcaraz, J.J. and Cerdán, F. (2006) ‘Using buffer management in 3G buffers to enhance end-to-end TCP performance’, Proceedings of IEEE AINA 2006, Vol. 2, pp.437–441.

Alcaraz, J.J., Cerdán, F. and García-Haro, J. (2006) ‘Optimizing TCP and RLC interaction in the UMTS radio access network’, IEEE Network, Vol. 20, No. 2, pp.56–64.

Barakat, C., Altman, E. and Dabbous, W. (2000) ‘On TCP performance in a heterogeneous network: a survey’, IEEE Communications Magazine, Vol. 38, No. 1, pp.40–46.

Bestak, R., Godlewski, P. and Martins, P. (2002) ‘RLC buffer occupancy when using a TCP connection over UMTS’, Proceedings of IEEE PIMRC, Vol. 3, pp.1161–1165.

Chakravorty, R., Clark, A. and Pratt, I. (2005) ‘Optimizing web delivery over wireless links: design, implementation and experiences’, IEEE Journal on Selected Areas in Communication, Vol. 23, No. 2, pp.402–416.

Chockalingam, A. and Zorzi, M. (1999) ‘Wireless TCP performance with link layer FEC/ARQ’, Proceedings of IEEE ICC, Vol. 2, pp.1212–1216.

Floyd, S. and Jacobson, V. (1993) ‘Random early detection gateways for congestion avoidance’ IEEE/ACM Transactions on Networking, Vol. 1, pp.397–413.

Gurtov, A. and Floyd, S. (2004) ‘Modeling wireless links for transport protocols’, ACM SIGCOMM Computer Communication Review, Vol. 34, No. 2, pp.85–96.

Holma, H. and Toskala, A. (2004) ‘WCDMA for UMTS: radio access for third generation mobile communications’, 3rd ed., Wiley.


Inamura, H., Montenegro, G., Ludwig, R., Gurtov, A. and Khafizov, F. (2003) ‘TCP over second (2.5G) and third (3G) generation wireless networks’ IETF RFC 3481.

Meyer, M., Sachs, J. and Holzke, M. (2003) ‘Performance evaluation of a TCP proxy in WCDMA networks’, IEEE Wireless Communication, Vol. 10, No. 5, pp.70–79.

Rossi, M., Scaranari, L. and Zorzi, M. (2003) ‘On the UMTS RLC parameters setting and their impact on higher layers performance’, Proceedings of IEEE 57th VTC, Vol. 3, pp.1827–1832.

Srivastava, V. and Motanim, M. (2005) ‘Cross-layer design: a survey and the road ahead’, IEEE Communications Magazine, Vol. 43, No. 12, pp.112–119.

Xylomenos, G. and Polyzos, G.C. (2003) ‘Multi-service link layer enhancements for the wireless internet’, Proceedings of IEEE ISCC, pp.1147–1152.

Yavuz, M. and Khafizov, F. (2002) ‘TCP over wireless links with variable bandwidth’, Proceedings of IEEE 56th VTC, Vol. 3, pp.1322–1327.

using buffer management in 3g radio bearers to …ait.upct.es/~jjalcaraz/pub/ijwcmc_2009.pdf ·...

Documents