Beruflich Dokumente
Kultur Dokumente
I. I NTRODUCTION
Routing plays a clear role in the efficient use of network
capacity, since routing choices can overload certain paths
while under-utilizing others. Addressing this issue is the focus
of Traffic Engineering [1], focusing mostly on methods to
distribute a given traffic matrix of demands within network
capacity. Solutions involve use of optimization together with
practical techniques to implement the results via OSPF weights
[4], [13], [15] or MPLS tunnels (see [1]). When external demands are unkwown or time-varying, dynamic load balancing
methods are required (e.g., [2]).
The impact of routing becomes more subtle when combined
with the TCP layer: congestion-controlled sources adapt rate
to whatever capacities are made available, effectively reducing
the demand on a congested path, so superficially the demand
appears to be served. It is only when the TCP/IP layers are
analyzed jointly that efficiency can be precisely defined, and
integrated solutions can be sought. In recent years, research on
Multipath TCP [6] has tackled this problem from the transport
layer side: here, TCP sources manage multiple congestion
windows for different fixed routes provided by IP. This approach has generated interest in the standards community [3].
Its impact is, however, constrained by the limited path diversity
in the first-hop seen by the source (e.g., a multi-homed server).
It is difficult to overcome this limitation without a stronger
breach of layer separation, namely source routing inside an
operator network. The latter would also not be scalable since
the number of internal paths is exponential.
Far more efficiency impact and scalability can be obtained
if IP routers are endowed with a dynamic multipath funcResearch supported in part by AFOSR-US under grant FA9550-09-1-0504.
P
k
Problem 1 (NUM): Maximize
k Uk (x ), subject to link
capacity constraints yl cl , and flow balance constraints at
each node.
To consider all paths would require controlling route choices
at all possible intermediate hops. Source-based Multipath TCP
cannot exploit this path diversity, so it cannot reach the
efficiency of Problem 1.
In [12] a method is proposed that is able to fully exploit
path diversity by controlling the aggregate rate xk of each TCP
d
that
flow (as is current practice), plus the split fractions i,j
specify multipath routing to destination d at router i. Namely,
if xdi is the traffic rate reaching router i destined to d, the
router sends to neighbor j the rate portion
d
d
yi,j
= i,j
xdi .
(1)
d
Control laws are presented in [12] for xk , i,j
so that the
equilibrium allocation of Problem 1 is achieved, or alternatively that of
P
P
Problem 2: Maximize S := k Uk (xk ) l l (yl ) subject to flow balance constraints.
This approximation amounts to replacing capacity constraints
with a barrier function; it also can be seen as combining the
utility maximization approach with cost minimization used in
Traffic Engineering [4], [13], [15].
Control is based on congestion feedback, based on a link
congestion measure or price pl . This variable can be generated
as in standard congestion control (primal or dual versions,
see [14]), which will lead to solutions of either Problem 1 or
Problem 2. For the purposes of this paper it suffices to say
that either packet loss fraction or queuing delay can serve as
congestion price, provided TCP responds to this quantity.
In the multipath setting, congestion prices must be averaged
among paths to destination. We introduce node prices qid ,
representing the average price of sending packets from node
i to destination d, defined through the recursion
X
d
qdd = 0,
qid =
i,j
[pi,j + qjd ], i 6= d.
(2)
j
d
) is the mean price experi(also denoted i,j
Here pi,j +
enced from i to d when routing through next hop j. The overall
mean node price qid is obtained by averaging these prices with
their corresponding routing fractions.
The overall price qsd from the source node to destination
serves as congestion control signal for the TCP sources. For
the overall network to solve the desired optimization, split
fractions must also respond to the same congestion prices.
One control law proposed in [12] for this purpose is
qjd
d
d
i,j
= i (id i,j
);
(3)
d
here id is the average of the i,j
over j. So we reduce transmission on paths with higher than average price, transferring
to lower price paths. In [12], a saturation is further imposed to
the right-hand side of (3) so that traffic fractions remain nonnegative. It is shown in [12] that this algorithm, together with
prices pl = 0l (yl ) provides global convergence to the solution
B. ns2 implementation
We supplemented the standard ns2 distribution that contains
a module for TCP-Reno, with multipath router modules.
Many features are common to the implementation in [12],
Fig. 1.
Fig. 3.
Fig. 4.
2 For this to happen, RTTs must be the same; this is set up by making
external delays predominant with respect to those inside this loop.
Fig. 7.
Fig. 6.
R EFERENCES
[1] D. Awduche, A. Chiu, A. Elwalid, I. Widjaja, X. Xiao, Overview and
Principles of Internet Traffic Engineering, RFC3272, IETF.
[2] A. Elwalid, C. Jin, S. Low, and I. Widjaja, MATE: MPLS Adaptive
Traffic Engineering, Proc. IEEE INFOCOM 2001.
[3] A. Ford, C. Raiciu, M.Handley TCP Extensions for Multipath Operation with Multiple Addresses, Internet Draft, Oct. 2009.
[4] B. Fortz and M. Thorup, Internet Traffic Engineering by Optimizing
OSPF Weights, Proc. IEEE INFOCOM 2000.
[5] R. G. Gallager, A minimum delay routing algorithm using distributed
computation, IEEE Trans. on Comm., Vol Com-25 (1), pp. 73-85, 1977.
[6] H. Han, S. Shakkottai, C.Hollot, R. Srikant and D. Towsley, Multi-Path
TCP: A joint congestion and routing scheme to exploit path diversity in
the Internet, IEEE/ACM Trans. Netw. Vol. 14(6), pp. 1260-1271, 2006.
[7] C. Jin, D. X. Wei and S. H. Low, FAST TCP: motivation, architecture,
algorithms, performance; Proc. IEEE INFOCOM 2004.
[8] V. Jacobson, Congestion avoidance and control, Proc. ACM SIGCOMM 88.
[9] F. P. Kelly, A. Maulloo, and D. Tan, Rate control for communication
networks: Shadow prices, proportional fairness and stability, Jour. Oper.
Res. Society, vol. 49(3), pp 237-252, 1998.
[10] http://athenea.ort.edu.uy/publicaciones/mate/en/index.html.
[11] M. Mathis, J. Semke, J. Mahdavi, T. Ott The Macroscopic Behavior of
the TCP Congestion Avoidance Algorithm, Computer Communication
Review, volume 27, number 3, July 1997.
[12] F. Paganini, E. Mallada, A unified approach to congestion control and
node-based multipath routing, IEEE/ACM Trans. on Networking, Vol.
17, no. 5, pp. 1413-1426, Oct. 2009.
[13] A. Sridharan, R.Guerin, C. Diot. Achieving Near-Optimal Traffic
Engineering Solutions for Current OSPF/ISIS Networks. IEEE/ACM
Transactions on Networking, March 2005.
[14] R. Srikant, The Mathematics of Internet Congestion Control, Birkhauser,
2004.
[15] D. Xu, M. Chiang, J. Rexford, Link-state routing with hop-by-hop
forwarding can achieve optimal traffic engineering, Proc. IEEE INFOCOM 2008.