Professional Documents
Culture Documents
1, FEBRUARY 2016
AbstractMultipath TCP (MP-TCP) has the potential to greatly Various congestion control algorithms have been proposed
improve application performance by using multiple paths trans- as an extension of TCP NewReno for MP-TCP. A straightfor-
parently. We propose a uid model for a large class of MP-TCP al- ward extension is to run TCP NewReno on each subpath, e.g.,
gorithms and identify design criteria that guarantee the existence,
uniqueness, and stability of system equilibrium. We clarify how al-
[3] and [4]. This algorithm, however, can be highly unfriendly
gorithm parameters impact TCP-friendliness, responsiveness, and when it shares a path with a single-path TCP user. This mo-
window oscillation and demonstrate an inevitable tradeoff among tivates the Coupled algorithm, which is fair because it has the
these properties. We discuss the implications of these properties same underlying utility function as TCP NewReno, e.g., [5] and
on the behavior of existing algorithms and motivate our algorithm [6]. It is found in [7]however that the Coupled algorithm re-
Balia (balanced linked adaptation), which generalizes existing al- sponds slowly in a dynamic network environment. A different
gorithms and strikes a good balance among TCP-friendliness, re-
sponsiveness, and window oscillation. We have implemented Balia
algorithm is proposed in [7] (which we refer to as the Max al-
in the Linux kernel. We use our prototype to compare the new al- gorithm), which is more responsive than the Coupled algorithm
gorithm to existing MP-TCP algorithms. and still reasonably friendly to single-path TCP users. Recently,
opportunistic linked increase algorithm (OLIA) is proposed as
Index TermsComputer networks, convergence, nonlinear
dynamical systems, TCPIP. a variant of Coupled algorithm that is as friendly as the Coupled
algorithm but more responsive [8]. See [9] for more references
to early work on multipath congestion control.
I. INTRODUCTION Our goal is to develop structural understanding of MP-TCP
algorithms so that we can systematically trade off different prop-
1063-6692 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
PENG et al.: MULTIPATH TCP: ANALYSIS, DESIGN, AND IMPLEMENTATION 597
Linux kernel, and we evaluate its performance using our Linux . Let represent the sending rate on route
prototype. . Each link maintains a congestion price at time . Let
We now summarize our proposed Balia MP-TCP algorithm. be the aggregate price on route . In
Each source has a set of routes . Each route maintains a this paper, represents the packet-loss probability at link ,
congestion window and measures its round-trip time . The and represents the approximate packet-loss probability on
window adaptation is as follows. route .
For each ACK on route We associate three state variables
for each route . Let
(1) .
Then, represents the corresponding
For each packet loss on route state variables for each source . For each link , let
be its aggregate trafc rate.
(2) Congestion control is a distributed algorithm that adapts
and in a closed loop. Motivated by the AIMD algorithm of
where and .
TCP NewReno, we model MP-TCP by
The rest of the paper is structured as follows. In Section II, we
develop a uid model for MP-TCP and use it to model existing (3)
algorithms. In Section III, we prove several structural proper-
ties, focusing on design criteria that determine the existence, (4)
uniqueness, and stability of system equilibrium, TCP-friend-
where for and for . We
liness, responsiveness, window oscillation, and an inevitable
omit the time in the expression for simplicity. Equation (3)
tradeoff among these properties. In Section IV, we discuss the
models how sending rates are adapted in the congestion avoid-
implications of these properties on existing algorithms. This
ance phase of TCP at each end system, and (4) models how
motivates our new MP-TCP algorithm, and we explain our de-
the congestion price is (often implicitly) updated at each link.
sign rationale. In Section V, we compare the performance of the
The MP-TCP algorithm installed at source is specied by
proposed algorithm with existing algorithms using Linux imple-
, where and
mentations of these algorithms. We conclude in Section VI.
. Here, is a vector of positive gains
II. MULTIPATH TCP MODEL that determines the dynamic property of the algorithm.
determines the equilibrium properties of the algorithm. The link
In this section, we rst propose a uid model of MP-TCP algorithm is specied by , where is a positive gain that
and then use it to model MP-TCP algorithms in the lit- determines the dynamic property. This is a simplied model for
erature. Unless otherwise specied, a boldface letter the RED algorithm that assumes the loss probability is propor-
denotes a vector with components . We use tional to the backlog and is used in, e.g., [10] and [11].
to denote the -di-
mensional vector without and to denote B. Existing MP-TCP Algorithms
the -norm of . Given two vectors , means
We rst show how to relate the uid model (3) to the window-
for all components . A capital letter denotes a matrix
based MP-TCP algorithms proposed in the literature. On each
or a set, depending on the context. A symmetric matrix is
route , the source increases its window at the return of each
said to be positive (negative) semidefinite if
ACK. Let this increment be denoted by , where is
for any , and positive (negative) definite if
the vector of window sizes on different routes of source . The
for any . For any matrix , dene
source decreases the window on route when it sees a packet
to be its symmetric part. Given two arbitrary matrices and
loss on route . Let this decrement be denoted by . Then,
(not necessarily symmetric), means is
most loss-based MP-TCP algorithms take the form of the fol-
positive semidenite. For a vector is a diagonal
lowing pseudocodes.
matrix with entries given by .
For each ACK on route .
A. Fluid Model For each loss on route .
We now model the above pseudocodes by the uid model
Consider a network that consists of a set (3). Let be the net change to window on route in each
of links with nite capacities . The network is shared by a set round-trip time. Then, is roughly
of sources. Available to source is a
xed collection of routes (or paths) . A route consists of a set
of links . We abuse notation and use both to denote a source
and the set of routes available to it, depending on the context.
Likewise, is used both to denote a route and the set of links since the loss probability is small. On the other hand
in the route. Let be the collection of all
routes. Let be the routing matrix: if
link is in route (denoted by ), and 0 otherwise.
For each route denotes its round-trip time (RTT). Hence
For simplicity, we assume are constants. Each source main-
tains a congestion window at time for every route
598 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 24, NO. 1, FEBRUARY 2016
From (3) we have where we have ignored taking the minimum with the term
since the performance is mainly captured by .
(5) Recently, OLIA MP-TCP algorithm [8] is shown to achieve
good performance in many scenarios. OLIA uses complicated
We now apply this to the algorithms in the literature. We rst feedback congestion control signals and cannot be modeled by
summarize these algorithms in the form of a pseudocode and (3)(4). We do, however, include OLIA in our Linux-based per-
then use (5) to derive parameters and of the uid formance evaluation in Section V.
model (3).
Single-Path TCP (TCP-NewReno): Single-path TCP is a spe- III. STRUCTURAL PROPERTIES
cial case of MP-TCP algorithm with . Hence, is Throughout this paper, we assume, for all
a scalar, and we identify each source with its route . and only if for some .A
TCP-NewReno adjusts the window as follows. point is called an equilibrium of (3)(4) if it satises, for
For each ACK on route . all and
For each loss on route .
From (5), this can be modeled by the uid model (3) with
or equivalently
We now summarize some existing MP-TCP algorithms, all
of which degenerate to TCP NewReno if there is only one route and if (6)
per source. and if (7)
EWTCP [3]: EWTCP algorithm applies TCP-NewReno-like
algorithm on each route independently of other routes. It adjusts We make two remarks. First, an equilibrium does not
the window on multiple routes as follows. depend on , but only on . The design however
For each ACK on route . affects dynamic properties such as stability and responsiveness
For each loss on route . as we show below. Second, since and
From (5), this can be modeled by the uid model (3) with only if for some by assumption, any nite
equilibrium must have for all . In the following,
we always restrict ourselves to nite equilibria.
In this section, we denote an MP-TCP algorithm by
where is a constant. . We characterize MP-TCP designs
Coupled MPTCP [5], [6]: The Coupled MPTCP algorithm that guarantee the existence, uniqueness, and stability of
adjusts the window on multiple routes in a coordinated fashion system equilibrium. We identify design criteria that determine
as follows. TCP-friendliness, responsiveness, and window oscillation and
For each ACK on route . prove an inevitable tradeoff among these properties. We discuss
in Section IV the implications of these structural properties on
For each loss on route . existing algorithms. All proofs are relegated to the appendices.
From (5), this can be modeled by the uid model (3) with
A. Summary
We rst present some properties of an MP-TCP algorithm
that we have identied. We then interpret them and sum-
Semicoupled MPTCP [7]: The Semicoupled MPTCP algo- marize their implications.
rithm adjusts the window on multiple routes as follows. C0: For each and each , the Jacobians of
For each ACK on route . are continuous and symmetric, i.e.,
For each loss on route .
From (5), this can be modeled by the uid model (3) with
s.t. (8)
where , and
is evaluated at .
The stability and responsiveness of the linearized system (10)
Fig. 1. Test network for the denition of TCP friendliness. The link in the (how fast does the system converges to the equilibrium locally)
middle is the only bottleneck link with capacity . is determined by the real parts of the eigenvalues of . Speci-
cally the linearized system is stable if the real parts of all eigen-
values of are negative; moreover, the more negative the real
Our proposed algorithm does not satisfy parts are, the faster the linearized system converges to the equi-
even though it seems to be stable in our experiments. This con- librium. We now show that the linearized system (10) is stable
dition is only sufcient and needed in our Lyapunov stability (i.e., converges exponentially fast to locally) and characterize
proof; see Appendix C. When depends on , one can its responsiveness in terms of the design choices .
replace in the denition of the Lyapunov function in Let .
(21) with evaluated at the equilibrium, and the same ar- Theorem 3.5 (Responsiveness): Suppose C1C3 hold. Then,
gument there proves that is (locally) asymptotically we have the following.
stable. Also see Theorem 3.5 for an alternative proof of local 1) The linearized system (10) is stable, i.e., for
stability. any eigenvalue of . Moreover, where
D. TCP Friendliness
Informally, an MP-TCP ow is said to be TCP friendly
if it does not dominate the available bandwidth when it shares where and are evaluated at the equilibrium point
the same network with a SP-TCP ow [2]. To dene this pre- .
cisely, we use the test network shared by a SP-TCP ow and an 2) For two MP-TCP algorithms and
MP-TCP ow under test as shown in Fig. 1. provided
All paths traverse a single bottleneck link with capacity ,
with all other links with capacities strictly higher than . The and for all
links have xed but possibly different delays. To compare the
friendliness of two MP-TCP algorithms and Theorem 3.5 motivates the following denition of respon-
, suppose that when shares the test network siveness. Given two MP-TCP and , we say that is more
with an SP-TCP, it achieves a throughput of in equilib- responsive than if . Theorem 3.5 (2) implies
rium aggregated over the available paths (the SP-TCP there- that an MP-TCP algorithm with a larger or more nega-
fore attains a throughput of ). Suppose achieves tive denite is more responsive, in the sense that the
a throughput of in equilibrium when it shares the test net- real parts of the eigenvalues of the Jacobian have a smaller,
work with the same SP-TCP. Then, we say that is friendlier more negative upper bound.
(or more TCP-friendly) than if , i.e., if re- Then, the next result suggests an inevitable tradeoff between
ceives no more bandwidth than does when they separately responsiveness and friendliness.
share the test network in Fig. 1 with the same SP-TCP ow. Theorem 3.6 (Tradeoff): Consider two MP-TCP algorithms
From the theory for single-path TCP ( for all ), and with the same gain . Suppose both satisfy
it is known that a design is more TCP-friendly if it has a smaller C1C3 and C5. Then, for all
marginal utility . The same intuition holds for
MP-TCP algorithms even though the utility functions may not
exist for MP-TCP algorithm.
Theorem 3.4 (Friendliness): Consider two MP-TCP algo-
In light of Theorems 3.4 and 3.5, Theorem 3.6 says that a more
rithms and . Suppose both satisfy
responsive MP-TCP design is inevitably less friendly if they
C1C4. Then, is friendlier than if for
have the same .
all .
The theorem is easier to understand in the case of SP-TCP,
i.e., when for all and . Then, it
E. Responsiveness Around Equilibrium implies that a more concave utility function has a larger
Suppose conditions C1C3 hold and there is a unique equi- marginal utility, and hence is less friendly.
librium . Assume all links in are active with
; otherwise remove from all links with prices . F. Window Oscillation
Let . The behavior of (3)(4) around the equi- Window oscillations are inherent in loss-based additive
librium is dened by the linearized system increase multiplicative decrease (AIMD) TCP algorithms. We
close this section by discussing informally why a larger design
(10) generally creates more severe window oscillations.
PENG et al.: MULTIPATH TCP: ANALYSIS, DESIGN, AND IMPLEMENTATION 601
This implies a tradeoff between responsiveness (which is The discussion motivates a new design that generalizes the ex-
enhanced by a large and oscillation (which is reduced isting MP-TCP algorithm.
with a small .
The effect of on window uctuations can be under- A. Implications on Existing Algorithms
stood by studying how it affects the decrease per packet Recall Table I that summarizes the conditions satised by
loss in the following packet-level model. the various algorithms. Only EWTCP and Coupled algorithms
For each ACK on route . satisfy C0. Their equilibrium properties can be studied in the
For each loss on route . standard utility maximization model as done for single-path
Let be an indicator variable of whether a packet TCP. Semicoupled and Max algorithms do not satisfy C0, and
loss is observed on route at an arbitrary time in steady state. therefore analysis through utility maximization is not appli-
Then cable. However, Theorem 4.1 implies that both Semicoupled
and Max algorithms satisfy C1C3 provided they enable no
more than eight routes. Theorems 3.2 and 3.3 then imply
that they have a unique and globally stable equilibrium. It is
also easy to show that EWTCP satises C1C3. The Coupled
represents the expected relative reduction in aggregate algorithm does not satisfy C2 and is found to have multiple
throughput , given that there is at least one equilibria in [5].
packet loss on some route . It is a measure of throughput Next, we discuss friendliness of existing MP-TCP algo-
uctuation for each packet loss that an application experiences. rithms. It can be shown that the corresponding to these
For TCP-NewReno (for which and is a scalar), the algorithms satisfy
window size is halved on each packet loss, ,
and hence .
To understand for MP-TCP algorithms, we need the
following result. for all if all routes have the same round-trip
Lemma 3.1: Let with elements. time. Since all of them satisfy C4, Theorem 3.4 implies that their
Each element is an independent binary random variable with friendliness will be in the same order, i.e., their throughputs in
. Dene the test network of Fig. 1 are ordered as follows:1
. Then
TABLE II
HOW DESIGN CHOICES AFFECT MP-TCP PERFORMANCE
B. Generalized Algorithm
Consider the class of algorithms parametrized by as
follows: Fig. 2. Network for our Linux-based experiments on TCP friendliness and re-
sponsiveness, with MP-TCP ows and single-path TCP ows sharing
two links of capacity and propagation delay (single trip) . MP-TCP
(12) ows maintain two routes with rate . Single-path TCP ows maintain one
route with rate .
This class includes the Max , Coupled
, and Semicoupled
algorithms as special cases when all RTTs on different paths of by subpaths with large throughput. The parameter in the
the same source are the same, i.e., . generalized algorithm (12) scales in the right way: A
The next result characterizes a subclass that has a unique and path that has a large has and hence a
locally stable equilibrium point. similar degree of window oscillation as existing algorithms,
Theorem 4.1: Fix any and . For any , while a path with a small has larger than that under
the in (12) satises the following: a design with zero and therefore is more responsive.
1) C1 if ; Our experiments show that Max algorithm
2) C2 and C3 if and are the same for overtakes too much of the competing single-path
all (assuming has full row rank). TCP ows. Hence, we can only use a smaller since is
The requirement that is not restrictive since in practice already innite in order to improve friendliness. To compensate
a device may typically enable no more than three paths. The the responsiveness performance, we will use a larger , which
requirement that are the same for all is used in proving will sacrice window oscillation performance. The Balia
the negative deniteness of the (symmetric part of the) Jacobian MP-TCP algorithm given at the end of Section I corresponds
of . Since a negative denite matrix remains negative to the choice . Instead of allowing the
denite after small enough perturbations of its entries, Theorem window size to drop to 1 for a packet loss, we add a cap for the
4.1 holds if the RTTs of the subpaths do not differ much. This decrement of window size, which improves the performance
(sufcient) condition seems reasonable as two paths between according to our experiments. Note that there is no best
the same sourcedestination pair often have similar RTTs if both parameter settings since there are tradeoffs among all the
are wireline paths. Note that our experiments in Section V show performance metrics and we choose
that the algorithm also converges even if the RTTs on different based on our experiments in Section V, which show that this
paths differ dramatically, e.g., the RTT of WiFi is usually much parameter setting strikes a good balance among responsiveness,
smaller than that of 3G. friendliness, and window oscillation.
For the class of algorithms specied by (12), Theorem 4.1
motivates a design space dened by , V. EXPERIMENT
where and control the tradeoff between friendliness and re- In this section, we summarize our experimental results that
sponsiveness and controls the tradeoff between responsive- illustrate the above analysis. In addition to the MP-TCP algo-
ness and window oscillation. In Table II, we summarize how rithms illustrated in Section II-B, we also include the recently
the parameters affect the performance. developed OLIA MP-TCP algorithm [8]. We evaluate the
We now describe our design philosophy. As discussed above MP-TCP algorithms using a reference Linux implementation
the design of MP-TCP algorithms involves inevitable trade- of MP-TCP, Multipath TCP v0.88 [13]. Since it currently in-
offs among responsiveness, friendliness, and the severity of cludes only Max and OLIA algorithms, we implement EWTCP,
window oscillation. Specically, a design is more responsive Semicoupled, Coupled, and the proposed Balia algorithm in the
if it has a higher gain or a more negative denite Jacobian reference implementation. For the Coupled and our algorithms,
(Theorem 3.5). However, a larger usually the minimum ssthresh is set to 1 instead of 2 when more than
creates a bigger window oscillation; a more negative denite one path is available.
implies a larger , usually hurting friendliness The network topology is shown in Fig. 2. In the testbed, all
(Theorems 3.6 and 3.4). This is summarized in Table II. Since nodes are Linux machines with a quad-core Intel i5 3.33 GHz
enabling multiple paths already reduces window oscillation processor, 4 GB RAM, and multiple 1 Gbps Ethernet inter-
compared to single-path TCP (Section IV-A), MP-TCP can faces, running Ubuntu 13.10 (Linux kernel 3.11.8). The net-
afford to use a relatively large gain for responsiveness. This work parameters such as , and are controlled by
does not compromise too much on window oscillation, but Dummynet [14].
allows us to use a less negative denite Jacobian Our experiments are divided into three parts. First, we com-
with a smaller to maintain sufcient TCP friendliness. pare TCP friendliness of Balia algorithm and prior algorithms.
Moreover, responsiveness is mainly affected by subpaths with The result conrms that the Couple algorithm is the friendliest,
small throughput, while window oscillation is mainly affected and Balia algorithm is close to the Coupled algorithm and
PENG et al.: MULTIPATH TCP: ANALYSIS, DESIGN, AND IMPLEMENTATION 603
TABLE V
RESPONSIVENESS: CONVERGENCE TIME (S) OF MP-TCP AND
TOTAL THROUGHPUT (Mb/s) OF ALL SINGLE-PATH TCP USERS
( ms, Mb/s, AND )
friendlier than the other algorithms. Second, we compare the
responsiveness of each algorithm in a dynamic environment
where ows come and go. The result shows that the Coupled
and OLIA algorithms are unresponsive (illustrating the tradeoff
between responsiveness and friendliness). EWTCP is the most
responsive; Balia is similar in responsiveness but friendlier
the dynamic performance of each algorithm, we assume the
to single-path TCP ows. Finally, we show that all MP-TCP
MP-TCP ow is long-lived while the single-path TCP ows
algorithms have smaller average window oscillations than
start at 40 s and end at 80 s. We record the aggregate throughput
single-path TCP.
of the single-path TCP ows from 40 to 80 s, which measures
These experiments conrm our analytical results and suggest
the friendliness of MP-TCP. We also measure the time for the
our design choice strikes a good balance among friendliness,
congestion window on the second path to recover2 of MP-TCP
responsiveness, and window oscillation.
users. It measures the responsiveness of MP-TCP. These mea-
surements are shown in Table V, and the congestion window
A. TCP Friendliness
and throughput trajectories of all algorithms are shown in Fig. 3.
We study TCP friendliness of each algorithm, rst with paths To clearly show the responsiveness performance, we record the
of similar RTTs and then with paths of different RTTs, which longest convergence time found in our experiment in Table V,
emulates the wireless scenario. We assume all the ows are and the corresponding trajectories are shown in Fig. 3.
long-lived and focus on the steady-state throughput. EWTCP is the most responsive among all the algorithms.
In the rst set of experiments, we let ms, Ours is as responsive as the Max algorithm, yet signicantly
Mb/s, and . We repeat the experi- friendlier than EWTCP. Both Coupled and OLIA algorithms
ments 20 times. The average aggregate throughput of MP-TCP take an excessively long time to recover. For Coupled algo-
and single-path TCP users and the 95% margin of error for con- rithm, the excessively slow recovery of the congestion window
fidence interval (CI) are shown in Table III. The Coupled algo- on the second path (see Fig. 3) is due to the design that in-
rithm is the friendliest, and Balia algorithm is closer to Coupled creases the window roughly by on each ACK
algorithm than the others. assuming the RTTs are similar. After the single-path TCP ow
In the second set of experiments, we assume highly heteroge- has left, is small while is large, so that
neous RTTs by emulating the scenario of a mobile device with is very small. It therefore takes a long time for to increase
both 3G and WiFi access. WiFi access usually has higher ca- to its steady-state value. In general, under the Coupled algo-
pacity and lower delay compared to 3G. Specicially, we set rithm, a route with a large throughput can greatly suppress the
ms, Mb/s for the rst link to emulate WiFi throughput on another route even though the other route is un-
access and ms, Mb/s for the second link to derutilized. The reason of the poor responsiveness performance
emulate 3G access. When there exists single-path TCP ows, of OLIA can be explained using a similar argument as Cou-
i.e., , the behaviors of all the algorithms are similar to pled algorithm since they have the same increment/decrement
the equal-RTT case in the rst set of simulation. The Coupled for each ACK/loss in this scenario.
algorithm is the friendliest, and Balia algorithm is closer than
other algorithms. However, when there is no single-path TCP C. Window Oscillation
ow, i.e., and , the performance of OLIA is
not stable to effectively take all the available capacity, while the We use a single-link network model to compare window os-
other algorithms do not have such a problem. We repeat the ex- cillation under MP-TCP and single-path TCP. First, an MP-TCP
periments 20 times, and we nd sometimes OLIA does not use ow initiates two subpaths through that link, and we measure
the 3G access link. The average throughput of MP-TCP user the window size of each subpath and their aggregate window
and the 95% margin of error for condence interval is shown in size. Then, a TCP-Reno ow traverses the same link and we
Table IV. measure its window size. The results are shown in Fig. 4 for our
algorithm in comparison to single-path TCP (other MP-TCP al-
B. Responsiveness gorithms have a similar behavior). They conrm that enabling
We use the network in Fig. 2 with Mb/s, 2Dened as the rst time the congestion window on the second path reaches
ms, and . To demonstrate the average congestion window (e.g., 60) after the single-path users have left.
604 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 24, NO. 1, FEBRUARY 2016
VI. CONCLUSION
We have presented a model for MP-TCP and identied
designs that guarantee the existence, uniqueness and stability
of the network equilibrium. We have characterized the de-
sign space and studied the tradeoff among TCP friendliness,
responsiveness, and window oscillation. We have proposed
Balia MP-TCP algorithm that generalizes existing algorithms
and strikes a good balance among these properties. We have
implemented Balia in the Linux kernel and used it to evaluate
the performance of our algorithm.
APPENDIX A
PROOF OF THEOREM 3.1 (UTILITY MAXIMIZATION)
The Lagrangian of (9) is
and
(13)
and (14)
APPENDIX B
PROOF OF THEOREM 3.2 (EXISTENCE AND UNIQUENESS)
A. Proof of Part 1
For any link , let Lemma B.2: Suppose C3 holds. Then, at equilibrium
for all .
Proof: Suppose . Then, by C3,
whose component composes of all the elements in except .
and hence there is a link with . However, then for
For , let
all paths and hence by C3. This implies
, and hence by (7), contradicting .
Recall the vector notations that
and . According to C1, we have the following and
two facts, which will be used in the proof. . To prove uniqueness of the equilibrium, suppose for
is a nondecreasing function of on since the sake of contradiction that there are two distinct equilibrium
is a nonincreasing function of . points and . By Lemma B.2, we have and
since . . Hence, (6) implies and
Next, we will show that is a quasi-concave function . By Lemma B.1 and Assumption C2, we then have
of . In other words, for any xed , the set
is a convex set. If , then
The rst inequality holds since where the inverse exists by Condition C2. C2 also guaran-
if and if . The tees the inverse of , denoted by ; C4 ensures
last equality holds since by Lemma B.2 and (6). . Let
Hence
and
Then
where the last inequality holds since
by Lemma B.1 and Assumption C2. Similarly
Thus
where the last inequality holds since by the equi-
librium condition (7). Hence
(22)
Hence
and is the congestion price at the bottleneck link. Applying
the implicit function theorem, we get
The denominator is real and positive, and When , it is easy to show that satises C1 and
in the numerator is imaginary. Hence is negative semidenite under the conditions of the theorem.
We hence prove the theorem for .
A. Proof of Part 1
Fix any and . Fix any nite such that
for all . Fix any . We now show that there exists
an that satises (6), in particular , in two
where the last inequality holds because the numerator is nega- steps.
tive by Condition C2 and the denominator is positive. Since this First, there exists an that satises if and only
holds for all eigenvalues of , the linearized system (10) is if
stable. Moreover, as desired.
B. Proof of Part 2
(27)
Consider two MP-TCP algorithms and such
that which is equivalent to
For any (nonzero) , we have Since this holds for all , we have
(23)
(24)
(29)
Hence, .
Clearly, as . Let
APPENDIX F (30)
PROOF OF THEOREM 3.6 (TRADEOFF)
Fix an . Let and Then, since for all by assumption. Moreover,
. Suppose for the sake for all and hence
of contradiction that but
does not hold, i.e., there exists a nite
and a such that
where
APPENDIX G
PROOF OF THEOREM 4.1
We will show the results hold for any . Since
, the results also hold for . and is the unique value at which .
608 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 24, NO. 1, FEBRUARY 2016
B. Proof of Part 2
To prove satises C2 and C3 for , we will show provided .
that the Jacobian is negative denite if Proof: Given any , let . It
and are the same for . Other properties then sufces to show that, for every for
of C2 and C3 are easy to prove and we omit the proof. Fix an . Given any , consider
and let , the common round-trip time for all .
Let and (35)
Its Lagrangian is
(33)
Then, and .
3If , the Jacobian degenerates to
(34)
where the last equality follows from the independence of
which is merely negative semidenite. and
PENG et al.: MULTIPATH TCP: ANALYSIS, DESIGN, AND IMPLEMENTATION 609