Вы находитесь на странице: 1из 4

Fairness Comparison Based on Queue Schedule between Fast and Reno

Wei Liang 1,2 , Sulei Xu 2 , Jinyi Chang 1 1. Department of Computer, Changshu Institute of Technology, Changshu, China 2. Nanjing University of Post and Telecommunications, Nanjing, China
Abstract: In the current TCP congestion control algorithm,
TCP Reno and its variants in the high-speed networks, has encountered many constraints. It is difficult to meet the higher requirements of QoS guarantee. The emergence of the Fast TCP may be used as the alternative TCP in congestion control in high-speed network environment. So it is necessary to make a comparison between TCP Reno and FAST TCP on the fairness. This paper puts the focus on considering a number of TCP Reno and FAST TCP sharing a bottleneck link. By comparing the TCP Reno and FAST TCP based on queue management with the two typical scheduling algorithms Drop-Tail and RED, the throughput analysis of the two versions of TCP can show fairness between the TCP Reno and FAST TCP. The numerical analysis and NS2 simulation results can give the guidance for deploying of FAST TCP. queue management algorithms. Queue management

algorithms can be divided into two categories: active management and passive management. Drop-tail and RED are the typical representatives of the two types. Therefore, these two algorithms are set as the premise of a fair comparison. By comparing the throughput between the TCP Reno and FAST TCP based on the queue scheduling algorithm (both Drop-Tail and RED), we derive from the fair comparison effect between the two versions of TCP in general. The numerical analysis and NS2 simulation results can illustrate the fair comparison between TCP Reno and FAST TCP in details. 2. Network Model

Keywords: Congestion control; Queue schedule; Fast TCP;


TCP Reno. 1. Introduction Fairness is considered to be an important measure in the design of network, which gives rise to discussion on how to define and achieve fairness. Taken into account the fairness of the importance of the network, it is often required to reach an agreement to ensure that one protocol is more equitable than the other, or to make the protocol more equitable by setting some parameters. The current TCP congestion control algorithm, TCP Reno and its variants have encountered many problems in the high-speed network environment. In recent years, FAST TCP began to gain the popularity and may be used as an alternative to TCP. Similarly, it has a good performance as to the system stability, throughput, convergence, and fairness. Therefore, it is necessary to make a comparison between TCP Reno and FAST TCP on the fairness. This paper focuses on considering this situation: multiple TCP Reno and FAST TCP sharing a bottleneck link under the
Figure 1: Network model

In order to realize the comparison of two TCP versions in the paper, we designed the network model shown in Figure 1. Figure 1 consists of N r sender host(s) running TCP Reno,

N f sender host(s) running FAST TCP, a receiver host, an


intermediate router, and links connecting the senders/receiver to the router. Each link bandwidth and the propagation delay between the sender hosts and router are bw [Mbps] and delay sx [Sec] respectively; and the link bandwidth and the propagation delay between the receiver host and router are bw [Mbps] and delay sd [Sec] respectively. We denote the propagation delay as [Sec] between sender hosts and the receiver host, being equal to: = sx + sd . In addition, the

router's buffer size is B[packets].We consider two queue scheduling algorithms (both Drop-Tail and RED) using in the

978-1-4244-6252-0/11/$26.00 2011 IEEE

management of routers buffer and assume the sender hosts have the ability of sending the infinite amount of data.
3. Numerical analysis comparison

its behavior as follows:

= N

+
f

In the following, we will use the network model depicted in Figure 1 to carry out a mathematical analysis according to the average throughput of each TCP connection. In the analysis, we assume that the throughput to the buffer occupancy for each connection is equal portions at the Drop-Tail. The same interpretation is applied in the following RED queue.
3.1 Case of Drop-Tail

TCP Reno links continue to increase the window size until the router buffer is full and some packets losses occur. Accordingly, when a packet is lost in the router buffer, the overall window size of TCP Reno in all links can be obtained as follows:

Wr = 2 + B W f

In Figure 2, we explain the typical change of the total package queued in the router buffer when the Drop-Tail algorithm is utilized. Here, assuming that all TCP Reno connections have the same performance. The window size of TCP Reno connections will be increased until the package loss occurs at the router buffer and its change has the cycles trigged by packet losses, even when the TCP Reno connections share the same link with FAST TCP connections. By assuming that all packet losses can be detected by the fast retransmit algorithm, it takes one RTT [sec] for the sender side TCP to detect the packet loss after the packet loss actually occurs at the route buffer. It corresponds to the flat part of buffer occupancy shown in Figure 2.
B Buffer occupancy [packets] RTT

In the buffer overflow period the number of lost packets becomes N r [packets], according to TCP Reno algorithm in adjusting the window and the window size of the TCP Reno link becomes increased by 1 [packets / RTT] in the congestion avoidance phase. By assuming that each packet loss probability is proportional to its window size, we can obtain the respective number of packet losses of TCP Reno and FAST TCP during buffer overflow are Lr [packets] and
L f [packets]:

Lr = N r

Wr Wr + W f

Lf = N f

Wf Wr + W f

According to the fast retransmit algorithm, when packet loss is detected by each of TCP Reno connections, the size of its window will be halved. So the total window size of TCP Reno connections Wr' [packets] can be derived from equation (4) as
Time [sec]

TCP Reno packets FAST TCP packets 1 Cycle

follows:
Wr' = 2W f + Wr W 1 Wr Lr + r ( N r Lr ) = Wr 2 Nr Nr 2(Wr + W f )

Figure 2: The change of buffer occupancy at Drop-tail router

FAST TCP connections on the other hand control their window size according to the observed sending packet RTTS. The number of queued packets of each connection is kept between and in the router buffer. Therefore, from the above observations N f FAST TCP connection for the total window size W f [packets], we can obtain the following relations:

From equation (1) and Figure 2, we can obtain the average window size:

W Wr' 1 (Wr' + Wr ) r + Wr 2 Nr Wr = Wr Wr' +1 Nr scheduling algorithm can be derived as follows:


Br = Wr B 2 + B

Accordingly, the respective average number of two queue

N f < W f < N f

f B

= Wf

We can get W f [packets], the average value of W f , from

B 2 + B

By assuming that they become proportional to the buffer

occupancy at the router, we finally have r [packets/sec] and

neglected. Therefore, we must consider the time-out expiration into reducing network throughput. We denote the probability of occurring timeout expiration within the window by pto . By using Wr , the average value of the window size of a certain

f [packets/sec], the average throughput of the connections of


two versions of TCP is as follows:

r =

Br Br + B f

f =

Bf Br + B f

TCP Reno connection when packet loss is detected, we determine pto by the following simple equation:
Wr W pto = r p i (1 p)Wr +1i i =2 i

3.2 Case of RED

The incoming packets will be dropped by the RED algorithm at the preset probability when the number of packets in the buffer exceeds a certain threshold value. To simplify the following analysis, we assume that all the data packets are lost to the same probability P of occurrence, and buffer overflow situation is not considered. Even with the RED algorithm, TCP Reno connections continue to increase their window sizes until packet loss occurs. Therefore, as in the drop-tail case, the FAST TCP connections cannot open their window sizes and keep them ranging from to . Therefore, the following equations yield for W f and W f :
+ Nf <Wf < Nf ; Wf = N f
2

(11)

In follows, we distinguish two cases of detecting packet loss; retransmission timeout expiration (TO case) and the fast retransmit (FR case), because in each of two cases, a different algorithm of changing the window size is used. If retransmission timeout expiration detects packet loss, the window size is reset to 1 [packet]. It is then updated according to
Wr 2

the

slow

start

phase

until

it

reaches

[packets]. We can determine Tto ,1 [sec], the time duration

of the slow start phase, and Ato ,1 [packets], the number of

(9)

packets transmitted in the slow start phase, by the following equations:

On the other hand, in the Drop-Tail, when there is packet loss, it will also trigger change in its window size. On the assumption, the arrival of the packets is possible to discard with the probability P. On average, each link will send 1 P

Tto,1 = rtt log2

Wr ; 2

Ato,1 =

Wr 1 2

12

Where rtt is the value of RTTS of sending packets. Furthermore, we can easily obtain the time duration Tto ,2 (sec) and the number of transmitted packets Ato ,2 [packet] in the following congestion avoidance phase, respectively as:

packets in each cycle (2 packet loss event interval), and we define the number of transmission data packets as N p during a cycle and the following is given:

Np = 1

(10)

Tto ,2 = rtt (Wr


Ato,2 =

Different from the drop-tail router case, we focus on a certain TCP Reno connection because we assume that all TCP Reno connections behave identically under the stochastic packet dropping algorithm employed by RED. Although the RED algorithm can reduce the robust packet loss, retransmission timeout expiration cannot be perfect to avoid. Even if the timeout expiration rarely occurs, the impact of the timeout expiration on the network throughput cannot be

Wr ); 2
( 13)

W W 1 (Wr + r ) (Wr r ) 2 2 2

These equations are based on the premise of a fact: In the congestion avoidance phase, the window size will increase 1 (packet) per RTT. On the other hand, if the packet loss is detected by the TCP Reno connection through the fast retransmit algorithm, the

window size will be halved to

Wr

and the congestion

but there is still inevitable unbalance between fairness and throughput. And when the FAST TCP competes with the TCP version at the router buffer, its performance is better than TCP Reno at the low buffer size. However, with the rising of buffer size, the throughput of Fast TCP becomes weak, which is beyond our expectation.
Reference
[1] D. Wei, C. Jin, S. H. Low, S. Hegde. FAST TCP: Motivation, Architecture, Algorithms, Performance. IEEE/ACM Trans. Networking,

avoidance will start again. And then the duration and the number of transmitted packets during slow start phase are zero, i.e.:

T fr ,1 = 0 A fr ,1 = 0

14

Similarly, time duration and the number of transmitted packets in the congestion avoidance phase can be represented by:

T fr ,2

W = rtt (Wr r ) 2

14(6):1246-1259, Dec. 2006 [2] A. Tang, K. Jacobsson, L. L. H. Andrew and S. H. Low An accurate link

model and its application to stability analysis of FAST TCP. In Proc. IEEE INFOCOM, 2007 [3] K. Koo, J. Choi, and J.S. Lee. Parameter Conditions for Global Stability

Afr ,2 =

W W 1 (Wr + r ) (Wr r ) (5) 2 2 2 Then, from equation (12) (15), the first phase of the

of FAST TCP. IEEE Communications Letters. Vol. 12, NO.2, Feb. 2008 [4] A. Tang, L. L. H. Andrew, K. Jacobsson, K. Johansson, S. H. Low and

number of transmitted packets and the average window size can be met by the following equation:

H. Hjalmarsson. Window Flow Control: Macroscopic Properties from

Np = pto(A ,1 + A ,2) +(1 pto)(Afr,1 + Afr,2) to to


A +A A +A W =rttpto to,1 to,2 +(1pto) to,1 to,2 r T +T Tfr,1 +Tfr,2 +rto fr,1 fr,2

16

Microscopic Factors In Proc. IEEE INFOCOM, 2008. [5] J. Wang, D. X. Wei, S.H, Low. Modeling and Stability of FAST TCP.

IEEE 2005

17

[6]

F. Zhao, J. Zhou and N. Lu. Stability Analysis of FAST TCP Based on

Lyapunov Function. In Proc. 7th Word Congress on Intelligent Control and Automation. Jun.2008 [7] A. Baiocchi, F. Vacirca. TCP fluid modeling with a variable capacity

Where rto [sec] is the timeout value of the retransmission connection since you can derive the p to and Wr from the equation (16) (17). And the average window size of all the TCP Reno links can be easily obtained as follows:

bottleneck link. IEEE INFOCOM 2007 [8] K. Jacobsson, L. L. H. Andrew, Ao Tang, S. H. Low, H. Hjamlmarsson, An Improved Link Model for Window Flow Control and Its Application to FAST TCP[J], IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 54, NO. 3, MARCH 2009, pp.551-564 [9] R. Shorten, F. Wirth and D. Leith. Modelling TCP in droptail and other

Wr = N r wr
Finally,

18

f and r in the RED case can be determined

environments. Automatica, To appear, 2007 [10] K. Jacobsson, L. L. H. Andrew, A. Tang, K. Johansson, H. Hjalmarsson and S. H. Low. ACK-clocking dynamics: Modeling the interaction between

similarly to the Drop-Tail router case from equations (7) and (8), (9) and (18).
4. Conclusion

windows and the network. In proc. IEEE Infocom, 2008. [11] J. S. Lee and J. T. LIM. Window-Based Congestion Control to Improve TCP Fairness and Utilization for Wide-Bandwidth Networks. In procaine Infocom, 2008

In this paper, we do the investigation of fairness between the two versions of TCP: FAST TCP, TCP Reno in the bottleneck link. We gained the following results through numerical analysis and simulation experiments: FAST TCP performance has been severely affected by the Drop-Tail circumstances, because of the different buffer occupancy at the router. To a certain extent, RED routers can improve fairness,

Вам также может понравиться