A Fair and Dynamic Load
Balancing Mechanism
F. Larroca and J.L. Rougier
International Workshop on Traffic
Management and Traffic Engineering for the
Future Internet
Porto, Portugal, 11-12 December, 2008
Agenda
Introduction
Utility
Maximization Load-Balancing
Distributed Algorithm
Simulations
• Packet-Level Simulations
• Fluid-Level Comparison
Conclusions
page 1
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Introduction
 Network
Convergence:
• Traffic increasingly unpredictable and dynamic
 Classic TE techniques (i.e. over-provisioning)
inadequate:
• Ever-increasing access rates
• New emerging architectures with low link capacities
 Possible answer: Dynamic Load-Balancing
• Origin-Destination (OD) pairs with several paths:
how to distribute its traffic?
• Paths configured a priori and distribution dependent
on current TM and network condition
page 2
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Introduction
 Network
operator interested OD pairs obtained
performance
• Why not state the problem in their terms?
 Analogy with Congestion Control (TCP):
• End-hosts = OD pairs
• Rate = OD performance indicator
 Differences:
• Decision variable: portion of traffic sent through each
path (total traffic is given)
• Much larger time-scale
page 3
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Introduction
 Previous
proposals:
• Define a link-cost function Fl(rl) for each link l=1..L
• Minimize the total network’s cost
 Example:
Limitations:
• Indirect way of proceeding
• Cannot prioritize an OD pair or enforce fairness
page 4
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Agenda
Introduction
Utility
Maximization Load-Balancing
Distributed Algorithm
Simulations
• Packet-Level Simulations
• Fluid-Level Comparison
Conclusions
page 5
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Utility Maximization Load-Balancing
 Define
a single performance indicator per OD pair
• us(d): performance perceived by OD pair s when
traffic distribution is d
 “Distribute” us(d) among OD pairs to maximize total
Utility (à la Congestion Control)
S
max
d
d U
s 1
s
s
(u s (d ))
• ds = total demand of OD pair s (given)
• dsi = traffic sent through path i of OD pair s (∑dsi= ds)
• d = [ d11 d12 .. dS1 .. dSnS ]T
 How to define us(d)?
page 6
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Utility Maximization Load-Balancing
choice for us(d): mean path’s Available
Bandwidth (ABW)
 Our
u s(d)   psi ABWsi  psi arg min ABWl 
lsi
 Assumptions:
• Majority of traffic is elastic (i.e. TCP)
• Path choice considered propagation delay
 Advantages:
• Mean ABW rough approximation of rate obtained by
TCP flows (ABW is the most important indicator)
• Sudden increases in demand may be
accommodated
page 7
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Utility Maximization Load-Balancing
 Final
version of the problem:
 ns

max  d sU s   psi ABW si 
d
s 1
 i 1

S
ns
s.t. Rd  c, d  0,  d si  d s s
i 1
 If
ABWsi is the flow obtained rate, the problem is
very similar to Multi-Path TCP
• By only changing ingress routers, users may be
regarded as if they used MP-TCP: improved
performance and more supported demands
page 8
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Agenda
Introduction
Utility
Maximization Load-Balancing
Distributed Algorithm
Simulations
• Packet-Level Simulations
• Fluid-Level Comparison
Conclusions
page 9
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Distributed Algorithm
 The
optimization problem is not convex
 However, not too “unconvex”
 The distributed algorithm solves the dual problem
and results in a good approximation
 Based on the Harrow-Hurwitz method: greedy on
path utility (PU) minus path cost (PC)
PU si  U ' (u s ) ABWsi PCsi  ˆl
lsi
where
ˆl    sil and
s i:lsi
d siU ' ( ABWl ), if l  arg min ABWl 
lsi
 sil  
0, otherwise
page 10
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Agenda
Introduction
Utility
Maximization Load-Balancing
Distributed Algorithm
Simulations
• Packet-Level Simulations
• Fluid-Level Comparison
Conclusions
page 11
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Packet-Level Simulations
A
simple example: all links have the same capacity
and probabilities are updated every 50 seconds
page 12
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Fluid-Level Simulations
 In
Comparison
with two previous
two real topologies
and TMs:proposals:
• MATE: minimize total M/M/1 delay
1
min 
l ABWl
• TeXCP: greedy on the path’s maximum utilization
 Two
performance indicators:
• Mean ABW (us) (weighted mean, 10% quantile and
minimum)
• Link Utilization (mean, 90% quantile and maximum)
page 13
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Fluid-Level Simulations – Abilene
 Mean
ABW (us)
UM/MATE
 Link
Utilization
TeXCP - MATE
page 14
UM/TeXCP
F. Larroca and J.L. Rougier
TeXCP - UM
FITRAMEN 08, Dec. 2008
Fluid-Level Simulations – Géant
 Mean
ABW (us)
UM/MATE
 Link
UM/TeXCP
Utilization
TeXCP - MATE
page 15
F. Larroca and J.L. Rougier
TeXCP - UM
FITRAMEN 08, Dec. 2008
Agenda
Introduction
Utility
Maximization Load-Balancing
Distributed Algorithm
Simulations
• Packet-Level Simulations
• Fluid-Level Comparison
Conclusions
page 16
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Conclusions



page 17
Performance as perceived by OD pairs is always better
in UM than in MATE or TeXCP
• MATE: relatively small differences in mean, but
significant in the worst case
• TeXCP: more significant differences
Link utilization results for TeXCP and UM are very
similar
• MATE: although similar in mean and quantile, the
maximum link utilization may increase significantly
Future Work:
• Stability
• Other simpler methods or objective function that obtains
similar results
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008
Thank you
Questions?
page 18
F. Larroca and J.L. Rougier
FITRAMEN 08, Dec. 2008