Making Intra-Domain Routing Robust to Changing and Uncertain Traffic Demands: Understanding Fundamental Tradeoffs

smashlizardsΔίκτυα και Επικοινωνίες

29 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

59 εμφανίσεις

Making Intra
-
Domain Routing Robust to
Changing and Uncertain Traffic Demands:

Understanding Fundamental Tradeoffs




Presented by Bharath Balasubramanian


Introduction


Intra
-
Domain Traffic Engineering:


Understanding Traffic Demands


Configuring Routing Protocols


There is no quantified relation between traffic demands
and good utilization.


Is knowledge of traffic demands crucial?


Can we obtain a routing protocol robust to variations in
demand?


Measuring Traffic Demands


Accurate knowledge of Traffic Matrix (TM) leads to
better utilisation of link capacities.


Traffic Demands cannot be measured accurately:


Flow Measurements not available at all points.


Demands change over time.


Internal and external failures.


We require robust routing strategy which performs for
a wide range of traffic demands.

Robust Routing




How well can a routing strategy be designed without (or
with very little) knowledge of TM?


When traffic demands change what range of change is
tolerable?


Will a routing designed to be optimal for one TM
behave well if the actual traffic deviates?


How will a routing perform in the event of link failures?

Data


Topologies: Six ISP maps from the Rocketfuel dataset
along with OSPF/IS
-
IS weights.


Capacities: Derived from OSPF/IS
-
IS weights using
inverse proportionality of capacity and weights.


Traffic Matrices:


Bimodal TMs


Gravity TMs

Metrics and Methodology


Maximum link utilisation
: Maximum ,over all links
,of the total flow on the link divided by the capacity of
the link.


Optimal Routing
: For a certain TM,
D
it is defined as
the routing which minimises maximum utilisation.


Performance Ratio
: It is defined as the maximum link
utilisation of
f

on
D
divided by the minimum possible
link utilisation on this TM.


Performance Ratio for a set of TMs.


When
D
includes all set of TMs, the performance ratio
is the
Oblivious Performance Ratio.


Optimal Oblivious Routing
: Routing which
minimises the Oblivious Performance Ratio.


Metrics and Methodology

Limitations


Parameters defined only for a given network topology.


Metrics do not capture relation between traffic
demands and resulting throughput.


Focus is only on point
-
to
-
point and not point
-
to
-
multipoint.


We may want to optimise beyond link utilisation (Eg.
MPLS
-
label stack size or number of provisioned paths)

Experiments and Results


Three different routings were considered:


Optimal Oblivious Routing.


OSPF routing.


Optimal Routing for gravity TMs


Optimal oblivious routing produced the best results by
far: Performance ratio ranging between 1.425
-
1.972.


Other two techniques produced 2
-
3 digit ratios.


43%
-
97%(overhead) in max utilisation is not negligible
but it has been obtained with no knowledge of traffic
demands.



Experiments and Results


Experiments with TM varying over a range( error
margin).


Routings evaluated:


opt


no
-
margin opt


OSPF


global
-
opt



Nm
-
gravity opt


Results:


Routing optimised for the set of TMs varying over a range
gives the best results.


For larger margins global
-
opt performs as well as opt.


Optimal routing allows for fairly sizable error margins with
performance ratio close to 1.



Experiments and Results


Performance under link failures: Oblivious routing
outperforms OSPF.


Are the results the same when networks scale??


Graphs with asymmetric link capacities have optimal
oblivious ratio that is Ω(n^0.5).


Some symmetric graphs have optimal oblivious ratio that is
Ω(logn).


Cliques and cycles have optimal oblivious ratio that is 2
-
2/n.







LP Models for Oblivious Routing




Optimal oblivious routing can be can be computed in
polynomial time in the size of the network.


A simpler model was obtained with faster running times.


Basic Lemmas used:


Removal of degree one nodes does not affect the oblivious ratio
of the network.


Consider an undirected network G, and a directed network G`
derived from it by replacing each edge e by two anti
-
parallel arcs
that have the same capacity as e.


The two networks, G and G`have the same optimal oblivious
ratio. Moreover, G and G` have (the same) symmetric optimal
oblivious routing.



Conclusion



Good traffic engineering requires demand oblivious
routing.


It is possible to obtain such techniques.


Standard tools will not provide the solutions. Hence the
algorithmic tools provided here are required.



THANK YOU!