Sie sind auf Seite 1von 7

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/321854404

A Fairness-based Congestion Control Algorithm


for Multipath TCP

Conference Paper · April 2018

CITATIONS READS

0 28

3 authors:

Reem Melki Mohammad M. Mansour


American University of Beirut American University of Beirut
2 PUBLICATIONS 2 CITATIONS 129 PUBLICATIONS 1,379 CITATIONS

SEE PROFILE SEE PROFILE

Ali Chehab
American University of Beirut
236 PUBLICATIONS 1,008 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Pruned Interleavers View project

Software Defined Networking View project

All content following this page was uploaded by Mohammad M. Mansour on 16 December 2017.

The user has requested enhancement of the downloaded file.


A Fairness-based Congestion Control Algorithm for
Multipath TCP
Reem Melki, Mohammad M. Mansour and Ali Chehab
Department of Electrical and Computer Engineering, American University of Beirut
Email: {rmm71, mmansour, chehab}@aub.edu.lb

Abstract—Multipath TCP (MP-TCP) has been introduced as and window oscillation which has prevented existing CCAs
an extension to the legacy TCP transport protocol to support from achieving optimal performance. Moreover, the existing
communication through multiple paths under a single connection fluid model in [8] mathematically models a limited class of
session. The target is to improve both resource utilization and
connection robustness. Several congestion control algorithms CCAs, and only relies on experimental simulations to assess
(CCAs) have emerged in the literature to adapt subflow rates and study the performance of the existing CCAs.
to congestion conditions on the various paths without negatively In this paper, a generalized fluid model is proposed. Unlike
impacting competing single-path TCP sources. The challenge is previous work, the proposed fluid model mathematically mod-
to provide a trade-off among three factors, namely, fairness, els all existing CCAs and provides an accurate interpretation
responsiveness, and window oscillation. In this paper, we propose
a new fairness-based CCA (FCCA) based on the fluid model that of the performance of each algorithm, which is validated
improves fairness without degrading the other two metrics. The through experimental simulations. In addition, a new Fairness-
FCCA tracks the performance on each route and dynamically based CCA (FCCA) is derived based on an optimized set of
adapts the respective congestion windows, enhancing the overall parameters that define the proposed fluid model. The proposed
performance. The proposed algorithm is implemented in a Linux FCCA outperforms existing CCAs by improving fairness with-
kernel. Simulation results demonstrate that FCCA is capable
of achieving almost maximal fairness (98%) while maintaining out degrading responsiveness and window oscillation.
responsiveness, unlike existing CCAs. II. R ELATED W ORK
I. I NTRODUCTION Most existing CCAs follow a common methodology for
The advancement of multi-homed devices, supporting both adjusting the congestion window on multiple routes during
cellular and WiFi technologies, has led to the rapid increase the congestion avoidance phase. They essentially perform two
in traffic demand. Legacy transport protocols such as Trans- operations: 1) when an ACK is received on route r, the
mission Control Protocol (TCP) and User Datagram Protocol congestion window wr is incremented by I(wr ), and 2) when
(UDP) [1], [2] use only one access interface exclusively there is a loss of packets on route r, the congestion window
although multiple IP addresses which can connect to multiple is decremented by D(wr ):
networks are available [3]. This has led to the emergence of Increment wr : wr ← wr + I(wr ) (1)
a new transport protocol known as Multipath TCP (MP-TCP) Decrement wr : wr ← wr − D(wr ) (2)
[4], which utilizes multiple paths simultaneously. These increment and decrement factors differ between CCAs,
MP-TCP has the potential to significantly enhance appli- and have a major impact on both, fairness and responsiveness.
cation performance in terms of throughput, robustness and Some of the popular existing MP-TCP congestion control
network resource utilization [5]–[7]. However, with multiple algorithms are reviewed in this section. In the following, we
subflows, it is more challenging to manage traffic congestion assume the source s belongs to a set of sources s ∈ S that
in the network. As a result, congestion control has become an share the network, and has a set of routes R that can be used
essential requirement for MP-TCP to operate efficiently. to send data. Each route j ∈ R maintains a congestion window
Various MP-TCP CCAs have emerged in the literature [4], wj and measures its own round-trip time τj . We denote the rate
[8]–[13]. One simple approach is to implement TCP NewReno of route j by xj , where xj = wj /τj . Note that in the sequel,
[1]; however, this algorithm results in an unfair distribution of we refer to a particular route, r ∈ R, on which congestion
resources among users. In [11] and [12], a coupled congestion control is to be applied to, and the parameters wr , xr , and
control algorithm (also referred to as the linked increases τr designate its congestion window, rate, and round-trip time,
algorithm (LIA)) and the opportunistic linked-increases al- respectively. We use the words route, subpath, and subflow
gorithm (OLIA) have been presented, respectively. Both of interchangeably throughout the text.
these algorithms are “fair” algorithms; however, they suffer The EWTCP algorithm [4] adjusts its congestion window in
from large delays. On the other hand, the Semi-Coupled [9], a manner similar to NewReno [1]. It assigns a weight µ to each
MAX [13], and EWTCP algorithms [4] are more responsive route, and in turn each route modifies its window according
than LIA and OLIA, but they are more aggressive to TCP to these weights independently from other routes:
users on shared subflows. As it can be inferred, there exists µ wr
IEWTCP (wr ) = and DEWTCP (wr ) = .
an inevitable tradeoff between TCP fairness, responsiveness wr 2
The LIA algorithm [11], also known as the Coupled CCA, MP-TCP congestion control algorithm can be designed and
adjusts the congestion window on each subflow based on the evaluated using appropriate equations for k(xr ) and φ(xr ).
ratio of the congestion window wr over the square of the The increment and decrement factors I(wr ) and D(wr ) of
round-trip time τr2 of the current route r, divided by the sum
of rates of all R routes that belong to source s: the general MP-TCP algorithm can be modelled as:
xr
wr /τr2 wr ẋr = [I(wr )(1 − qr ) − D(wr )qr ] (4)
ILIA (wr ) = P 2 ; DLIA (wr ) = τr
2 Assuming that the packet-loss probability qr is very small
j∈R wj /τj
(< 1% for wired links and 6% for wireless links), then the
The Semi-Coupled MP-TCP algorithm [9] adjusts the win- expression I(wr )(1 − qr ) can be approximated as I(wr ).
dow on multiple paths according to the inverse of the product Rearranging terms, (4) can then be written as:
of the round-trip time τr and the sum of rates xj : xr I(wr )

1 ẋr = D(wr ) − qr (5)
ISemi−Coupled (wr ) = P ; τr D(wr )
τr j∈R wj /τj Now, comparing (3) and (5), the expressions for the gain k
wr and the term φ are derived as:
DSemi−Coupled (wr ) = xr I(wr )
2 k(xr ) = D(wr ) and φ(xr ) = . (6)
τr D(wr )
The OLIA algorithm [12] takes a more complex form. Its In [8], an extended fluid model is derived after examining
increment factor is based on that of the LIA algorithm, in the expressions of k(xr ) and φ(xr ) (equation (6)) specific to
addition to a term that is inversely proportional to wr . This each congestion control algorithm. The general form of this
ratio is modulated by some factor δ as specified in [12]: fluid model is parameterized by (β, η, n), where β ∈ R+ ,
δ wr η ∈ R+ , and n ∈ N:
IOLIA (wr ) = ILIA (wr ) + ; DOLIA (wr ) = 1 
wr 2 k(x) = xr xr + η kxk∞ − xr , (7)
2
The MAX algorithm [13] takes the minimum between two 
quantities. The first is based on the LIA algorithm but uses 2 (1 − β) xr + β kxkn
φ(x) = , (8)
instead the maximum ratio max{wj /τj2 } in the numerator τr2 xr kxk21
j∈R n
where the boldface letter x ∈ R denotes a vector P with n
instead of wr /τr2 pertaining to the route of interest. The second
factor, 1/wr , is based on the TCP NewReno algorithm [1]. The components xj , kxk∞ = maxj |xj | and kxkn = ( j xnj )1/n .
update equations are given  by: The effect of the design parameters (β, η, n) on the MP-TCP
performance is summarized in the Table I [8].

2
 max{wj /τj }
 
1

j∈R
IMAX (wr ) = min P 2 , ; Table I
wr  E FFECT OF D ESIGN PARAMETERS ON P ERFORMANCE
j∈R wj /τj

 
w Performance Parameters
DMAX (wr ) = r Fairness ↑ β ↓, n ↑
2
Responsiveness ↑ β ↑, n ↓, η ↑
The balanced linked adaptation (BALIA) algorithm [8] Window Oscillation ↓ η↓
introduces the parameter αr , max{xj }/xr to modulate the The positive gain parameter k(x) controls the responsive-
j∈R
increment and decrement factors of the window size as given ness property; an increase in k(x) (more specifically in the
by the LIA algorithm as: value of η) results in a more responsive congestion control
IBALIA (wr ) = ILIA (wr ) × 1+α
 4+α 
2
r
5
r
; algorithm and vice versa. Similarly, φ(x) is a positive gain
DBALIA (wr ) = DLIA (wr ) × min{αr , 3/2} parameter that controls the fairness property; increasing β
and n increase φ(x), which leads to a more friendly (fair)
III. F LUID M ODEL congestion control algorithm (Table I).
In [8], the authors model the MP-TCP algorithm by
+ IV. G ENERALIZED F LUID M ODEL
ẋr = k(xr ) (φ(xr ) − qr )xr , r ∈ R (3)
where ẋr is the time-derivative of the rate xr (t) of the sender A. Extended Fluid Model
on route r at time t, which is equal to the ratio of the The fluid model [8] as given by (7) and (8) models only
congestion window wr (t) by the roundtrip time τr ; k(xr ) is a class of MP-TCP congestion control algorithms, namely
a positive gain that controls the responsiveness property of the MAX, Semi-Coupled, and LIA algorithms. The other
the MP-TCP algorithm and its window oscillation; φ(xr ) is a algorithms cannot be represented by this fluid model; hence
factor that determines the equilibrium or fairness properties of one would rely on experimental simulations to analyze and
the MP-TCP algorithm; qr (t) is the approximate packet-loss compare the performance of these algorithms. To solve this
probability on route r; and the notation (δ)+x is equal to δ when limitation, we extend the model of [8] to span all MP-
x > 0 and to max {0, δ} for x ≤ 0. For simplicity, the time TCP congestion control algorithms presented in Section II by
variable t has been omitted from the equations. The model, generalizing k(x) in (7) and φ(x) in (8) as follows:
from which equation (3) is taken, is a simplified version of 1
k(x) = C[xr (xr + η(kxk∞ − xr ))], (9)
the more general RED algorithm [14], [15]. 2
QM
As seen from (3), both k(xr ) and φ(xr ) are the main 2 m=1 [(1 − βm ) xr + βm kxkn ]
φ(x) = 2 (10)
parameters for modeling a MP-TCP algorithm; therefore a (C) τr2 xM
r kxk1
This generalized fluid model is parameterized by a new set moreover, the performance of OLIA depends on the value of ρ,
of parameters ({β1 , β2 , ..., βM }, η, n). The product in (10) which is not fixed and which directly affects the expressions of
is composed of M factors, where each factor depends on k(x) and φ(x). However we can interpret its performance from
a different parameter βm . Moreover, a constant gain C is the results presented in [8]. Since all of the above algorithms
introduced in (9) and (10), leading to a more flexible and have the same k(x) = 12 x2r , they can be ordered as follows:
accurate modeling of various algorithms. Using (6) and (9), φEWTCP (x) ≥ φSemi−Coupled (x) ≥ φMAX (x) ≥ φLIA (x)
the decrement factor, D(wr ), can be generalized as D(wr ) = EWTCP is said to be the most responsive algorithm, while
C
2 wr , where 0 < D(wr ) < wr , and thus, 0 < C < 2. LIA algorithm is said to be the friendliest. On the other
Applying the above model to existing MP-TCP algorithms, hand, BALIA has a different value of gain k(x), which is
one can find expressions for both gains k(x) and φ(x), and 3 2
4 xr , so it cannot be compared to the above algorithms using
afterwards compare algorithms using these expressions. The the previously mentioned criteria. However, using the results
EWTCP, LIA, Semi-Coupled, OLIA and MAX algorithms presented in [8], BALIA is said to be friendlier than the MAX
have the same gain k(x) = 21 xr 2 with C = 1, whereas BALIA algorithm and more responsive than the Coupled algorithm.
has a gain k(x) = 34 x2r with C = 1.5. On the other hand, each The Coupled algorithm and OLIA have similar performance
algorithm has a different expression for φ as discussed below. in terms of fairness; however, both suffer from large delays,
The EWTCP algorithm has a value of M = 2, and which is a major drawback for both algorithms. Unlike the
consequently its parameters ({β1 , β2 }, η, n) are ({1, 1}, 0, 1): Coupled algorithm and OLIA, the MAX, EWTCP and Semi-
2 Coupled algorithms are very responsive but are not as fair as
2 kxk1 2
φEWTCP (x) = 2 = τ 2 x2 the OLIA and the Coupled algorithms.
2 2
τr xr kxk1 r r
For LIA, M = 1; hence (β1 , η, n) ≡ (0, 0, “don’t care”): B. Proposed Algorithm
2xr 2
φLIA (x) = 2 = 2 2
In this section we present a new fairness-based congestion
2
τr xr kxk1 τr kxk1 control algorithm (FCCA). We set the parameters M and
(β, η, n) by retaining the desired features of all previously
The Semi-Coupled algorithm also has M = 1, but its param- discussed algorithms as follows.
eters (β1 , η, n) are (1, 0, 1): • Choice of M : We model our φ(x) parameter using M =
2 kxk1 2 1 factor.
φSemi−Coupled (x) = 2 = τ 2 x kxk
2
τr xr kxk1 r r 1 • Choice of η: From equation (6), the gain k(x) is mainly
related to the decrement factor D(wr ); therefore, we set
η to zero in order to reduce window oscillation.
The MAX algorithm has similar parameters as the Semi-
• Choice of n: The parameter n directly affects φ(x)
Coupled algorithm, but with n = ∞ instead of n = 1,
in (10). We set n = ∞ in order to introduce a handle
(β1 , η, n) ≡ (1, 0, ∞):
2 kxk∞ in φ(x) to control the rate of the current subflow relative
φMAX (x) = 2 to the maximum rate between subflows. The ratio of the
2
τr xr kxk1 maximum rate between subflows over the rate of the
current subflow is an indicator of performance; the closer
BALIA introduces a variable αr2 in its increment factor; hence this ratio is to one the higher the rate of the current flow
M = 2 and the parameters ({β1 , β2 }, η, n) correspond to is. This generalizes the term αr introduced in [8], where
{ 12 , 15 }, 0, ∞ : max{xj }
αr = , (11)
2x2 ( 1+αr )( 4+α
5 )
r
xr
φBALIA (x) = r3 2 2 and max{xj } is the maximum rate among all subflows
2 2
2 τr xr kxk1 and xr is the rate of the current subflow.
• Choice of β: An infinite value of n requires β, where
OLIA has a more complex form. It can be modeled by setting 0 ≤ β ≤ 1, to be small (close to zero) in order to
√ √ 
M = 2 and ({β1h, β2 }, η, n) ≡ { ρ, ρ}, 0, 1 : achieve an acceptable fairness level. The smallest value
2 √ √ i
that can be considered is β = 0, which results in
2 ρ(kxk1 ) + (1− ρ)2 x2r + 2( ρ)xr kxk1
2
φOLIA (x) = 2 φ(x) = φLIA (x) = τ 2 kxk 2 (Section IV). In this case, the
τr2 x2r kxk1 r 1
effect of parameter n is eliminated and the performance
of MP-TCP is suboptimal, as is the case of the coupled
Now assume two MP-TCP algorithms modeled as algorithm (LIA), since both fairness and responsiveness
(k(x), φ1 (x)) and (k(x), φ2 (x)) have the same gain k(x). are sacrificed. In Fig. 1, experimental simulation shows
Algorithm 1 is said to be more responsive and less friendly that maximum fairness (98%) can be attained for a value
than Algorithm 2 if φ1 (x) > φ2 (x). In this way the above MP- of β equal to 0.1. On the other hand, too small a value
TCP algorithms can be compared and evaluated. It should be of β requires the constant gain C to be relatively large
noted that OLIA has a complex form that can be approximately (close to the upper bound), in order to increase k(x) and,
modeled by the generalized fluid model presented above; in turn, improve responsiveness.
MP−TCP Fairness as a function of Beta pared with other existing congestion control algorithms,
100
namely OLIA, LIA and BALIA.
98
MP−TCP Percentage Fairness (%)
The MP-TCP kernel v0.90 was installed on four Linux
96 machines running Ubuntu 14.04 LTS, two of which are used
94
as servers and the other two as clients. Each of these machines
is installed with two Ethernet cards with rate 100 Mbps each.
92
In Scenario 1, we created two parallel paths from the sender
90 to the receiver by connecting four machines to two routers
88
(with a rate of 100 Mbps) as shown in Fig. 2(a). One pair
implements MP-TCP and utilizes two subflows for sending
86
data, and the other pair uses regular TCP. The two MP-TCP
84
0 0.2 0.4 0.6 0.8 1
subflows traverse two different routes via two different routers;
Fairness Parameter Beta
the first path is unshared and the MP-TCP user can use it fully,
Figure 1. MP-TCP fairness as a function of different values of β
while the other path is shared between the MP-TCP user and
the TCP user. Both users are sharing an 11 Gbyte file in this
Hence, we choose β = 0.1 and C = 1.5 accordingly. Now
1 scenario using an FTP application over Ethernet links.
substituting (β, η, n) = ( 10 , 0, ∞) in equations (9) and (10)
Scenario 2 is similar to Scenario 1, but provides a more
with M = 1 yields:
realistic setting. Here we have 20 TCP users (ten pairs) instead
k(x) = C2 (x2r ) = 34 x2r (12)
of two. In order to have ten pairs using regular TCP instead
9 1 9+αr
 
2 10 xr + 10 kxk∞ 2 10 of one, we use Linux containers (LXC virtualization) and
φ(x) = 3 2 2 = 3 2 (13)
generate 20 containers, 10 containers on the client side and
2
P
2 τr xr kxk1 2 τr ( xj )
where αr is given in (11). The increment and decrement fac- 10 on the server side. These containers act as 20 independent
tors of the proposed algorithm can be obtained by comparing TCP users sharing data over an FTP connection (Fig. 2(b)).
equations (12) and (13) on one hand, and equation (6) on the Finally, Scenario 3 shown in Fig. 2(c) represents a hetero-
other. The resulting factors are geneous system. The same simulation setup as Scenario 1 is
x 9 + αr used; however, the MP-TCP user sends traffic on two links that
IFCCA (wr ) = Pr
τr ( xj )2 10 have different bandwidths, and as a result different round-trip
3wr times. The bandwidth of the first link, which is unshared, is
DFCCA (wr ) = . reduced to 10 Mbps, while that of the second link, which is
4
1

The choice of parameter values (β, η, n) = 10 , 0, ∞ pro- shared with another TCP user, is kept at 100 Mbps.
vides a good balance among the previously mentioned fac- In all scenarios, the TCP users start sending at time 40 sec
tors, and leads to an enhanced performance as demonstrated and stop at time 80 sec. During [0 sec-40 sec] and [80 sec-
through simulations in the following section. 120 sec], the MP-TCP user utilizes subflows 1 and 2 fully with
The proposed FCCA algorithm is listed in Algorithm 1. a total throughput equal to 186 Mbps (93 Mbps on each sub-
Algorithm 1 A Fairness-Based Congestion Control Algorithm flow), since the TCP users remain idle during these intervals.
Input: xr (t), τr (t) of flow r; subflow rates x1 (t), x2 (t), · · ·
Each simulation was repeated 10 times and an average value
Output: window size wr (t) of the throughput (between 40 sec and 80 sec) is reported. All
1: wr (t) ← 1 . initialize window simulations run over a period of 120 sec. The experiments
2: λr (t) ← advertised window size . initialize threshold show the performance gains of the proposed algorithm in terms
3: while user is active do of fairness, responsiveness and window oscillation.
4: if ACK is received then . packet received
5: if in congestion avoidance phase: wr (t) > λr (t) then A. Performance Evaluation
max{xj (t)}
6: αr (t) ← xr (t)
1) Scenario 1: From the results reported in Table II,
7: wr (t) ← wr (t) + τr (t)(xP
r (t)
xj (t))2
9+αr (t)
10 all MP-TCP algorithms (OLIA, LIA, BALIA, FCCA) utilize
8: else . slow start: wr (t) < λr (t)
9: wr (t) ← wr (t) + 1
subflow 1 fully, which is an unshared resource, and the
10: end if throughput attained on that subflow is maximum (93 Mbps).
11: else . packet loss: 3 duplicate ACKs or timeout OLIA and LIA have close throughput values on subflow 2
12: wr (t) ← wr (t) − 3wr (t)/4 . decrease window (3–4 Mbps respectively), and their total MP-TCP throughput
13: λr (t) ← wr (t)/2 . update threshold is within the same range (96 Mbps and 97 Mbps, respectively).
14: end if
15: end while
Both of these values are close to the throughput of the TCP
16: return wr (t) user on the shared subflow, which is around 90 Mbps, thus
attaining an acceptable level of fairness. However both of
these algorithms suffer from slow responsiveness as shown
V. E XPERIMENTAL S ETUP AND R ESULTS
in Figs. 3(a) and 3(d). BALIA, on the other hand, is more
To assess the efficacy of the proposed FCCA algorithm, responsive than the two. However, a MP-TCP user using
three scenarios are simulated and their performance is com- BALIA is less friendly on the shared subflow, consuming up
Link bandwidth= Link bandwidth=
100Mbps 10Mbps

(a) Scenario 1 (b) Scenario 2 (c) Scenario 3


Figure 2. Simulation setup for the three scenarios considered

Table II Table III


AVERAGE THROUGHPUT (M BPS ): S CENARIO 1 AVERAGE THROUGHPUT (M BPS ): S CENARIO 2
MP-TCP MP-TCP MP-TCP MP-TCP MP-TCP MP-TCP TCP/ Total
Algorithm TCP Algorithm
subflow 1 subflow 2 total subflow 1 subflow 2 total user TCP
Proposed FCCA 93 2.5 95.5 91 Proposed FCCA 93 2.5 95.5 9.1 91
BALIA [8] 93 11 104 8.5 85
BALIA [8] 93 11 104 85 OLIA [12] 93 3 96 9 90
OLIA [12] 93 3 96 90 LIA [11] 93 4 97 8.9 89
LIA [11] 93 4 97 89
Table IV
AVERAGE THROUGHPUT (M BPS ): S CENARIO 3
to 12% (11 Mbps) of the shared bandwidth which is slightly
MP-TCP MP-TCP MP-TCP
high. The overall MP-TCP throughput (104 Mbps) is larger Algorithm
subflow 1 subflow 2 total
TCP
than that of the TCP user (85 Mbps) (see Fig. 3(g)). Proposed FCCA 9 15 24 78
The proposed FCC algorithm strikes a good balance be- BALIA [8] 9 20 29 80
tween responsiveness and fairness. The results in Table II show OLIA [12] 9 4 13 85
that the total MP-TCP throughput is close to that of the TCP LIA [11] 9 4 13 85
users, which are 95.5 Mbps and 91 Mbps, respectively. The
FCCA algorithm utilizes only 2.6% of the shared bandwidth (29 Mbps). Moreover, LIA and OLIA have similar perfor-
(2.5 Mbps) as shown in Fig. 3(j). mance, both having a total MP-TCP throughput of around
2) Scenario 2: The results of this scenario are plotted in 13 Mbps. The reason is that in a heterogeneous system each
Figs. 3(b), 3(e), 3(h), and 3(k). They validate the results of link has a different propagation delay, which imposes more
Scenario 1. Table III presents the average throughput values challenges on allocating traffic to multiple links efficiently.
attained and summarizes the results of Fig. 3. Figure 3(k)
VI. C ONCLUSION
shows that the total throughput of the 20 TCP users is around
90 Mbps, which means that the shared subflow 2 is mainly A new fairness-based congestion control algorithm for MP-
used by the ten TCP pairs and that each user had a share of TCP has been proposed. It has been demonstrated that by
9.1 Mbps. The MP-TCP user, on the other hand, has utilized generalizing the fluid model, one can model the dynamics of a
the first unshared subflow fully (93 Mbps) and around 2% broader class of CCAs. Furthermore, by carefully setting the
(2 Mbps) of the bandwidth of the second (shared) subflow, key parameters of this generalized model, one can design a
which is a total of 95 Mbps. The 20 TCP users start sending CCA that ensures maximal fairness without negatively impact-
at time 40 sec and stop at time 80 sec; hence, it can inferred ing responsiveness and window oscillation. Future extensions
that FCCA is very responsive since the throughput of the TCP of the work will focus on optimizing the performance of MP-
users jumped to a maximum after few seconds (5 sec) and TCP CCAs in heterogeneous networks.
dropped back to zero within the same range of time, when data R EFERENCES
transfer stopped. This shows that the proposed FCCA reaches
[1] J.Postel, “Transmission control protocol,” RFC 793, September 1981.
stability within an acceptable range of time, while noting that [2] ——, “User datagram protocol,” RFC 768, August 1980.
the application simulated here is transfers a large file using [3] S. Baidya and R. Prakash, “Improving the performance of multipath tcp
FTP, which is not a real-time application and which is not time over heterogeneous paths using slow path adaptation,” in Proc. IEEE
Int. Conf. Commun. (ICC), 2014, pp. 3222–3227.
sensitive. The plot of Fig. 3(h) is similar to that of Fig. 3(k) [4] A. Ford and et. al, “Tcp extensions for multipath operation with multiple
in terms of responsiveness, but in terms of fairness, BALIA is addresses,” draft-IETF-mptcpmultiaddressed-12, October 2012.
less friendly and exhibits significant window oscillations. [5] ——, “Tcp extensions for multipath operation with multiple addresses,”
Tech. Rep., 2009.
3) Scenario 3: Figure (3) and Table IV show that the [6] C. Raiciu and et. al, “How hard can it be? designing and implementing
performance of MP-TCP degrades in a heterogeneous system. a deployable multipath tcp,” in 10th USENIX Symp. on Net. Sys. Design
However, simulations results show that FCCA performs better and Implement., 2012, pp. 29–29.
[7] P. Dong and et. al, “Performance enhancement of multipath tcp for
than LIA and OLIA in such situations; the total MP-TCP wireless communications with multiple radio interfaces,” IEEE Trans.
throughput is around 24 Mbps, which is close to BALIA on Comm., vol. 64, no. 8.
OLIA OLIA OLIA
200 200 90
MPTCP MPTCP−subflow1
180 TCP 180 80 MPTCP−subflow2
TCP
160 160
70

140 140
60

Throughput (Mbps)

Throughput (Mbps)
Throughput (Mbps)

120 120
50
100 100
40
80 80
30
60 60
20
40 40

20 20 10

0 0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time (sec) Time (sec) Time (sec)

(a) OLIA in scenario 1 (b) OLIA in scenario 2 (c) OLIA in scenario 3

LIA LIA LIA


200 200 90
MPTCP MPTCP−subflow1
180 TCP 180 80 MPTCP−subflow2
TCP
160 160
70

140 140
60
Throughput (Mbps)

Throughput (Mbps)
Throughput (Mbps)

120 120
50
100 100
40
80 80
30
60 60
20
40 40

20 20 10

0 0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time (sec) Time (sec) Time (sec)

(d) LIA in scenario 1 (e) LIA in scenario 2 (f) LIA in scenario 3

BALIA BALIA BALIA


200 200 90
MPTCP MPTCP−subflow1
180 TCP 180 80 MPTCP−subflow2
TCP
160 160
70

140 140
60
Throughput (Mbps)

Throughput (Mbps)
Throughput (Mbps)

120 120
50
100 100
40
80 80
30
60 60
20
40 40

20 20 10

0 0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time (sec) Time (sec) Time (sec)

(g) BALIA in scenario 1 (h) BALIA in scenario 2 (i) BALIA in scenario 3

Proposed congestion control algorithm Proposed congestion control algorithm Proposed congestion control algorithm
200 200 90
MPTCP−subflow1
180 180 80 MPTCP−subflow2
TCP
160 160
70

140 140
60
Throughput (Mbps)

Throughput (Mbps)

Throughput (Mbps)

120 120
50
100 100
40
80 80
30
60 60
20
40 40

20 20 10

0 0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120
Time (sec) Time (sec) Time (sec)

(j) Proposed FCCA in scenario 1 (k) Proposed FCCA in scenario 2 (l) Proposed FCCA in scenario 3
Figure 3. Variation of throughput as a function of time using different congestion control algorithms in the three scenarios

[8] Q. Peng and et. al, “Multipath tcp: Analysis, design, and implementa- amd a possible solution,” IEEE/ACM Trans. Netw., vol. 21, no. 5, pp.
tion,” IEEE/ACM Trans. Netw., vol. 24, no. 1, pp. 596–609, 2016. 1651–1665, 2013.
[9] D. Wischik and et. al, “Design, implementation and evaluation of [13] M. Allman and et. al, “Tcp congestion control,” RFC 5681, 2009.
congestion control for multipath tcp,” in USENIX Symp. on Net. Sys. [14] F. Kelly and et. al, “Rate control for communication networks: Shadow
Design and Implement., 2011, pp. 8–8. prices, proportional fairness and stability,” J. Oper. Res. Soc., vol. 49,
[10] ——, “Dynamic window coupling for multipath congestion control,” in no. 3, p. 237–252, 1998.
IEEE Int. Conf. on Net. Protocols (ICNP), 2011, pp. 341–352. [15] S. Low and D. Lapsley, “Optimization flow control. i: Basic algorithm
[11] C. Raiciu and et. al, “Coupled congestion control for multipath transport and convergence,” IEEE/ACM Trans. Netw., vol. 7, no. 6, p. 861–874,
protocols,” RFC 6356, October 2011. 1999.
[12] R. Khalili and et. al, “Mptcp is not pareto-optimal: Performance issues

View publication stats

Das könnte Ihnen auch gefallen