performance evaluation of tcp congestion control ... - Semantic Scholar

PERFORMANCE EVALUATION OF TCP CONGESTION CONTROL ALGORITHMS OVER UMTS G. Boggia, P. Camarda, L. A. Grieco, and A. M. Mebabi DEE – Politecnico di Bari V. Orabona, 4 – 70125 Bari, Italy Ph. +39 080 596 3301 – Fax +39 080 596 3410 E-mail: {g.boggia, camarda, a.grieco, msabdul}@poliba.it

Abstract TCP congestion control was designed and optimized for wired networks. During recent years, however, the popularity of wireless network infrastructures enabling ubiquitous Internet access, such as Universal Mobile Telecommunication System (UMTS), has grown considerably. As a consequence, many improved TCP congestion control algorithms, able to effectively operate over hybrid wired/wireless networks, as well as, layer 2 optimizations, hiding to higher layers the inefficiencies of radio links, have been proposed by the scientific community. This work investigates the performance of TCP congestion control over UMTS networks. We have focused the attention on the interactions between UMTS RLC error recovery, and New Reno or Westwood+ TCP congestion control algorithms. Using the ns-2 network simulator, we have shown that a proper tuning of RLC parameters can ensure an efficient usage of the UMTS wireless channel.

1. Introduction The popularity of wideband wireless networks enabling ubiquitous Internet access, such as Universal Mobile Telecommunication System (UMTS) [1], has grown considerably during last years. At transport layer, TCP [14] is being exploited for supporting a large quota of the Internet traffic [21]. Unfortunately, such a protocol has been designed and optimized for wired networks with low Bit Error Rate (BER), less than 10-7 [13]. In particular, TCP congestion control follows the Additive Increase Multiplicative Decrease (AIMD) paradigm, which is made of a probing phase, during which the sending rate is increased to probe the network available bandwidth, and a shrinking phase, which is triggered as soon as a network congestion is revealed by the reception of duplicated ACKs (DUPACKs) or a timeout expiration. In this way, TCP congestion control [15] treats each segment loss as a symptom of congestion, thus reducing the sending rate even if it is not due, as in the case of segment losses due to unreliable radio links. As a consequence, improved TCP congestion control algorithms, able to effectively operate over hybrid wired/wireless networks, as well as, layer 2 optimizations, hiding to higher level layers the inefficiencies of radio links, have been proposed by the scientific community [6-8, 17, 18]. In UMTS networks, to support reliable upper layer protocols, the Radio Link Control (RLC) protocol [11] is utilized to partially recover errors at link level. However the retransmission strategy of UMTS RLC protocol competes with TCP [9]. As a consequence many research efforts have been recently devoted to the study of standard TCP over UMTS RLC protocol [9, 13, 24-28], which have basically highlighted that the use of RLC in acknowledged mode is highly recommended to improve TCP performance, but the RLC protocol parameters have to

be carefully tuned. Moreover, these studies have been mainly focused on standard TCP congestion control algorithms, such as Tahoe and Reno [15]. TCP Westwood [17] is a new congestion control algorithm that proposes an Additive Increase ADaptive Decrease (AIADD) paradigm to enhance classic AIMD. AIADD paradigm leaves unchanged the probing phase of the AIMD, but employs an end-to-end estimate of the available bandwidth to adaptively set the sending rate after a congestion episode. TCP Westwood significantly improves the fairness in wired networks and the utilization of wireless links [17]. TCP Westwood+ [18] is a recent modification to TCP Westwood that employs an enhanced bandwidth estimation scheme, robust with respect to ACK compression [18]. This work investigates the interactions between the error recovery scheme of UMTS RLC and New Reno [16] or Westwood+ TCP congestion control algorithms. Using the ns-2 network simulator [12], we have shown that a proper tuning of RLC parameters can ensure an efficient usage of the UMTS wireless channel to both New Reno and Westwood+ TCP . The rest paper is organized as follows. Section 2 briefly describes UMTS RLC error recovery mechanisms. Section 3 gives a brief background of NewReno [16] and Westwood+ [17, 18] congestion control algorithms. Section 4 presents simulation results. Finally, last section draws the conclusion.

2. RLC overview In the layered architecture of UMTS radio interface protocols, RLC sits in layer 2, between the Medium Access Control (MAC) and the Radio Resource Control (RRC) layers [3, 11]. There are three operational modes for UMTS-RLC, namely, transparent, unacknowledged, and acknowledged. With the transparent mode, RLC layer does not add a separate header to the data. It simply forwards the data to the MAC layer. This mode is used principally for the transmission of multimedia data streams. In the Unacknowledged mode, RLC blocks are transmitted without any guarantees. As a consequence, data transfer over the radio interface is heavily affected by errors but no additional delay due to retransmission is experienced by RLC blocks, which are provided with a sequence number to allow segmentation, reassembly, and concatenation of Service Data Units (SDUs). Unacknowledged mode is used for cell broadcast service or for certain RRC signaling procedures. The acknowledged mode provides reliable data transfer over the error-prone radio channel. This is accomplished by retransmitting erroneous RLC protocol data units (PDUs) by using automatic repeat request (ARQ) mechanisms. The acknowledged mode, which is considered in this article, is the most sophisticated one. It can provide both in sequence and out of sequence delivery of SDUs to higher layers. RLC entity delivers only SDUs, that are error free and ensures that a SDU is delivered only once. In this work, the Selective Repeat ARQ scheme is the retransmission strategy considered in the acknowledged mode. The transmission buffer in the RLC protocol stores not only all the packets arriving from the upper layer and to be transmitted for the first time, but also RLC PDUs to be retransmitted. It is a priority queue, in which RLC PDUs requested for retransmission have a higher transmission priority than the RLC PDUs to be transmitted for the first time. When a negative acknowledgement is received, the adopted ARQ scheme can retransmit PDUs up to a

maximum number of times, i.e., the MaxDAT protocol parameter [11]. A status message is used by the receiver for notifying loss or corruption of a RLC PDU. Several mechanisms can trigger a status message. Either the sender or the receiver can trigger the status message. In addition, a timer is required at both sender and receiver for handling the triggering mechanisms. At the sender, the timer (called poll timer) is started when a request for status messages is sent to the receiver. The timer is cancelled if a status report is received before its expiration. If the timer expires and no status reports have been received, a PDU with the poll bit set is sent to trigger the transmission of the status report. At the receiver side, instead, the Move Receiving Window (MRW) timer is used to trigger the retransmission of a status report. The timer is started when the status PDU is transmitted. Each time the timer expires the status report is retransmitted and the timer is restarted. RLC allows a maximum number of blocks to be outstanding, i.e., sent but not yet acknowledged, over the radio link. This limit is known as transmission window.

3. TCP Algorithms overview This section briefly describes the NewReno and Westwood+ congestion control algorithms. 3.1. NewReno TCP As well-known, TCP congestion control is a sliding-window algorithm based on the additive increase multiplicative decrease (AIMD) paradigm. It employs two variables to throttle the sending rate: the congestion window (cwnd), which is the sliding window, and the slow start threshold (ssthresh), which varies the way cwnd is increased. In particular, during the probing phase, cwnd is linearly (exponentially) increased over the time, when cwnd ≥ ssthresh (cwnd < ssthresh). When a congestion is revealed via the reception of 3 DUPACKs (duplicated ACKs), cwnd is halved and ssthresh is set equal to cwnd, the segment that appears to be lost is retransmitted, and TCP enters in the fast-recovery phase. With Reno TCP [21], the fast-recovery phase is leaved when the retransmitted segment is successfully acknowledged. Instead, in the NewReno TCP variant, the fast recovery phase does not end until all the segments within the current window are acknowledged; this improves performance when several segments within the same window get lost [16]. Finally, when a timeout expires, ssthresh is set to the half of the cwnd, and cwnd is set equal to one. 3.2. Westwood+ TCP The key idea of Westwood+ TCP is to exploit the stream of returning acknowledgment packets to estimate the bandwidth B that is available for the TCP connection. When a congestion episode happens at the end of the TCP probing phase, the used bandwidth corresponds to the definition of best effort available bandwidth in a connectionless packet network [21]. The bandwidth estimate is used to adaptively decrease cwnd and ssthresh after a timeout or three duplicated ACKs, as it is described below: • •

When 3 DUPACKs are received by the sender: ssthresh = max{2, (B*RTTmin) /seg-size}; When a coarse timeout expires: ssthresh = max{2, (B*RTTmin) /seg-size};

cwnd = ssthresh. cwnd =1.

4. Simulation Results Fig. 1 illustrates the network scenario considered in this work. User Equipments (UEs) communicate with a base station (node B) over the UMTS radio interface Uu with a capacity of 2 Mbps [4]. The UMTS network is connected to an external packet-switched IP network through a router, with a link capacity of 200 Mbps a delay equal 10 ms. Links connecting fixed hosts to the router have 40 Mbps capacity and delay equal to 50 ms. In the radio access network, the UMTS-TDD (Time Division Duplex) mode is used because is more suitable for data traffic such as TCP flows. In fact, the TDD mode of operation is flexible in sharing the channel capacity and data traffic requires significantly different bandwidth in downlink and uplink [23]. The RLC is in acknowledged mode [11].

sink 1 FTP 1

sink 2 . . .

UMTS network

FTP 2 . . .

Node B

sink M FTP M

Fig. 1 Reference hybrid wired/wireless network scenario.

A number M of TCP receivers placed on mobile nodes are connected to M TCP senders placed within the wired section of the topology. The mobile nodes move randomly within an area whose boundary is defined in this simulation as 200 m × 200 m. The default node mobility model provided with the ns-2 simulator has been used [12]. A uniformly distributed Packet Error Rate (PER) affects radio links. Main parameters used in simulations are summarized in Table I. In the following sub-sections the impact of the MaxDAT parameter, the size of the RLC transmission buffer, and the number of mobile station (M) in one cell will be investigated. Table I. Main Simulation Parameters RLC mode

Acknowledged

Multiplexing mode

TDD

RLC PDU size

40 Byte

Traffic model

FTP

RLC buffer size

Variable: 60÷1500 PDUs

TCP flavors

New Reno/ Westwood+

Poll Timeout

0.02 sec

TCP Segment size

1000 Bytes

MWR Timeout

0.2 sec

Packet Error Rate (PER)

Uniformly distributed

MaxDAT

Variable from 1 to 4

Number of mobile stations

Variable from (1 to M )

RLC transmission window size

600 PDUs

4.1. The impact of maximum number of retransmissions The MaxDAT parameter represents the maximum number of retransmission attempts of a RLC-SDU before it is discarded. With a sufficiently high number of RLC retransmissions, the SDU error ratio provided by the RLC layer can be negligible. However, this reliability benefit is obtained at the expense of a variable transmission delay, which can potentially lead to TCP timeouts.

Figs. 2 and 3 show TCP goodput achieved by New Reno and Westwood+, respectively, i.e., the number of TCP segments correctly received over the simulation time, as a function of the packet error rate (PER), for different values of MaxDAT. They show that TCP goodput, as expected, decreases for increasing PER. On the other hand, the higher is MaxDAT and the higher becomes TCP goodput, because, for increasing values of MaxDAT, the error recovery of UMTS RLC becomes more effective. This is also evident in Figs. 4 and 5, where TCP segment retransmission ratios are reported. In particular, segment retransmission ratios get smaller for increasing MaxDAT because RLC error recovery hides to TCP the PER of the UMTS radio link. 350

400 MaxDAT=1 MaxDAT=2 MaxDAT=3 MaxDAT=4

300 250 200 150 100

250 200 150 100 50

50 10-4

10-3

10-2 Packet Error Rate

10-1

0

1

10-4

10-3


10-1

1

Fig. 2 TCP Goodput obtained by New Reno.

Fig. 3. TCP Goodput obtained by Westwood+.

1

1

Retransmission Ratio [%]


0

MaxDAT=1 MaxDAT=2 MaxDAT=3 MaxDAT=4

300 TCP Goodput [kbit/s]

TCP Goodput [kbit/s]

350

10-1

10-2

10-3 10-4

MaxDAT=1 MaxDAT=2 MaxDAT=3 MaxDAT=4 10-3


10-1

Fig. 4. TCP Retransmission ratio with New Reno.

1

10-1

10-2

10-3 10-4

MaxDAT=1 MaxDAT=2 MaxDAT=3 MaxDAT=4 10-3

-2

10 Packet Error Rate

10-1

1

Fig. 5. TCP Retransmission ratio with Westwood+.

Therefore, the retransmissions at RLC layer are highly beneficial to TCP based applications since it is always preferable to perform retransmissions at data link layer than using packet recovery at TCP layer. Therefore, the number of maximum allowed link layer retransmissions should be set to the highest possible value. 4.2. The impact of RLC buffer size RLC buffer has an important impact on TCP performance because packets that arrive when it is full are dropped, thus forcing TCP to decrease the sending rate. Figs. 6 and 7 show TCP goodputs versus PER for different values of RLC buffer size, obtained using New Reno and Westwood+, respectively. In both cases a larger RLC buffer leads to a larger TCP goodput. The benefits deriving from having a large RLC buffer can be observed also in Figs. 8 and 9, where segment retransmission ratios of New Reno and Westwood+ have been reported. From

these figures it is straightforward to note that a small RLC buffer lead to unacceptable segment retransmission ratios. In the case of PER=10-3, we have also considered values of the RLC buffer size ranging from 60 to 1500 PDU (see Fig. 11), to show that TCP goodput monotonically increases with RLC buffer size until it reaches a maximum value.

350

350 buffer = 2.4 kByte buffer = 4.8 kByte buffer = 10 kByte buffer = 24 kByte

250 200 150 100 50 0

250 200 150 100 50

10

-4

10

-3

-2

10 Packet Error Rate

10

-1

0 -4 10

1

Fig. 6. TCP Goodput with New Reno.


10-1

1

1


10-1

10-2

10-3 -4 10

10-3

Fig. 7. TCP Goodput with Westwood+.

1


buffer = 2.4 kByte buffer = 4.8 kByte buffer = 10 kByte buffer = 24 kByte

300



300

buffer = 2.4 kByte buffer = 4.8 kByte buffer = 10 kByte buffer = 24 kByte 10-3


10-1

1

Fig. 8. TCP Retransmission ratio with New Reno.

10-1

10-2

10-3 10-4

buffer = 2.4 kByte buffer = 4.8 kByte buffer = 10 kByte buffer = 24 kByte 10-3


10-1

1

Fig. 9. TCP Retransmission ratio with Westwood+.

4.3. The impact of the number of station in one cell This sub-section evaluates TCP performance in the scenario shown in Fig. 1, when the number of mobile nodes M ranges from 1 to 20, MaxDAT = 4, RLCBuffer = 600 bytes, and PER= 10 −3 . Fig. 11 reports the aggregate goodput achieved by the M TCP connections as a function of M. Obviously, as M increases the goodput increases up to the saturation of the UMTS capacity. The reason is that as the number of competing TCP connections increases their aggregate behavior becomes more aggressive and network resources are more utilized. This is confirmed by Fig. 12 where a segment retransmission ratio increasing with M is shown.

700

300

600

250

500



350

200 150 100

New Reno Westwood+

50

400 300

NewReno Westwood+

200 100

0

0

0

10

20

30 40 RLC Buffer [kByte]

50

60

70

0

Fig. 10. TCP goodput for PER=10-3.

5

10 Number of mobile stations

15

20

Fig. 11. TCP aggregate goodput.

Finally, to investigate the fairness in bandwidth sharing provided by the wireless UMTS access network we have adopted the Jain fairness index [22]. It is computed as follows:

(∑ b ) = M

J FI

2

i =1 i M 2 i =1 i

(1)

M∑ b

where bi is the goodput of the ith connection and M is the number of connections sharing the bottleneck. The Jain fairness index belongs to the interval [0,1] and increases with fairness reaching the maximum value at one. Hence, a higher fairness index indicates better fairness between flows. The fairness index is 1 when all the throughputs are equal. Otherwise, the fairness index drops below 1. Fig. 13 shows that a high JFI, which is roughly equal to 0.9, has been obtained regardless of the number of competing TCP connections, which means that UMTS access network is able to provide fairness in bandwidth sharing to TCP connections. 1 0.9 TCP Goodput [kbit/s]


10-1

10-2

NewReno Westwood+ 10-3

0.8 0.7 NewReno Westwood+

0.6 0.5

0

5

10 15 Number of Connections

Fig. 12. Rtx ratio.

20

25

0

5

10 15 Number of Connections

20

Fig. 13. Fairness Index.

5. Conclusion

This work has investigated the performance of TCP congestion control over UMTS networks. The attention has been focused on the interactions between UMTS RLC error recovery, and New Reno or Westwood+ TCP congestion control algorithms. Using the ns-2 network simulator, we have shown that a proper tuning of RLC parameters can ensure an efficient usage of the UMTS wireless channel.

25

References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [21] [22] [23] [24] [25] [26] [27] [28]

B. Walke, P. Seidenberg, M. P. Althoff, UMTS The Fundamentals, Wiley, June 2003. L. L. Peterson, B. S. Dave Computer Networks: A System Approach, 3rd ed., Morgan Kaufmann, 2004. H. Holma, A Toskala, WCDMA for UMTS, Wiley, 2001. J. Korhonen, Introduction to 3G mobile communication, Artech House, 2002. F. Lefevre, G. Vivier, “Understanding TCP Behavior Over Wireless Links,” in Proc. of Symposium on Commun. Veh. Technol., IMEC, Leuven, Belgium, Oct. 2000. H. Balakrishnan, V. N. Padmanabhan, S. Sechan, and R. H. Katz, “A Comparison of Mechanisms for Improving TCP performance Over Wireless Links,” IEEE/ACM Trans. Networking, vol. 5, no. 6, pp. 756−769, Dec. 1997. H. Balakrishnan, S. Seshan, and R. Katz, “Improving Reliable Transport and Handoff Performance in Cellular Wireless Networks”, ACM Wireless Networks, Dec. 1995. A. Bakre, B.R Badrinath, “I-TCP: Indirect TCP for Mobile Hosts,” in Proc. of 15th Int. Conf. on Distributed Computing Systems, Vancouver, Canada, May-Jun., 1995. F. Lefevre, G. Vivier, “Optimizing UMTS Link Layer Parameters for a TCP connection,” in Proc. of VTC 2001 Spring, Rhodes, Greece, May 2001. D. Huang, J. Shi, “Performance of TCP over radio link with adaptive channel coding and ARQ,” in Proc. of VTC 1999, Houston, Texas, May 1999. ETSI TS 125 322 V3.18.0, Universal Mobile Telecommunications System (UMTS); Radio Link Control (RLC) protocol specification (3GPP TS 25.322 version 3.18.0 Release 1999), Jun. 2004. Ns-2, Network Simulator, available at http://www.isi.edu/nsnam/ns. R. Bestak, P. Godlewski, and P. Martins, “RLC buffer occupancy when using a TCP connection over UMTS,” in Proc. of IEEE PIMRC, Lisbon, Portugal, 2002. J. Postel, ed., “Transmission Control Protocol,” IETF RFC 793, Sep. 1981, M. Allman, V. Paxson, and W. Stevens, “TCP congestion control,” IETF RFC 2581, Apr. 1999. S. Floyd, T. Henderson, and A. Gurtov, “NewReno modification to TCP’s fast recovery,” IETF RFC 3782, Apr. 2004. S. Mascolo, C. Casetti, M. Gerla, M. Sanadidi, and R. Wang.“TCP Westwood: End-to-end bandwidth estimation for efficient transport over wired and wireless networks,” in Proc. of ACM Mobicom, Rome, Italy, 2001. L. A. Grieco, S. Mascolo, “Performance evaluation and comparison of Westwood+, New Reno and Vegas TCP congestion control,” ACM Computer Communication Review, vol. 34, no. 2, pp. 25–38, Apr. 2004. R. Cáceres, L. Iftode, “Improving the performance of reliable transport protoclos in mobile computing environments,” IEEE J. Select. Areas Commun., vol. 13, no. 5, Jun. 1995. CAIDA, the Cooperative Association for Internet Data Analysis, available at http://www.caida.org. R. Jain, The art of computer systems performance analysis, John Wiley and Sons, 1991. R. Esmailzadeh, M. Nakagawa, TDD-CDMA for wireless communications, Artech House Publishers, 2003. A. Baiocchi and F. Cuomo, “Efficiency and fairness trade-off for TCP over UMTS-TDD,” Proc. of IEEE Int. Conf. on Communications (ICC 2001), Helsinki, Finland, June 2001. R. Bestak, P. Godlewski, and P. Martins, “A TCP connection over uplink UMTS radio access bearer in RLC acknowledged mode,” in Proc. of VTC 2002 Spring, Birmingham, England, May 2002. A. Baiocchi and F. Vacirica, “End-to-end evaluation of WWW and File Transfer performance for UMTS-TDD,” in Proc. of IEEE Globecom, Taipei, Taiwan, Nov. 2002. F. Vacirica, A. De Vendictis, A. Todini, and A. Baiocchi, “On the effects of ARQ mechanisms on TCP performance in wireless environments,” in Proc. of Globecom, San Francisco, California, Dec. 2003. M. Rossi, L. Scaranari, and M. Zorzi, “On the UMTS RLC parameters setting and their impact on higher layers performance,” in Proc. of VTC fall 2003 fall, Orlando, Florida, Oct. 2003.