Title: Routing and Congestion Problems in General Networks
1Routing and Congestion Problems in General
Networks
- Presented by Jun Zou
- CAS 744
2Outline
- 1. Introduction to Routing and Congestion
- 2. Network Model and Objective
- 3. Construction of Tree
- 4. Simulation Graph on Tree
- 5. Simulation Tree on Graph
- 6. Conclusion and Application
- 7. References its full paper and improved one
31. Introduction to Routing
- The function of routing is to find a best path
from source to destination for each incoming
packet. - What is best? Minimum hop count, minimum
delay, security, etc - In this paper, our goal is to minimize the
congestion of the whole network links.
42.1 Network Model
- Network a weighted graph G(V, E)
- Vn nodes and Em edges
- Bandwidth a function b(e) E? R
- Absolute load amount of data transmitted on a
edge e - Relative load L(e) Absolute load/bandwidth
- Congestion C Maximum over the relative load of
all links in the network
52.2 Two approaches to solve routing problems
- Traffic modeling and simulation Simplify
the traffic model (such as M/M/1 model), simulate
the routing protocols and analyze results by
using queuing theory
- Simulation graph on a tree Combine a
tree solution of an online problem and tree
representation of the network
62.3 Oblivious online routing algorithms
- Oblivious routing algorithm path selection for
the i-th request si does not depend on routing
decisions made for other requests sj - Oblivious adversary The request sequence s is
not allowed to depend on the selection of online
algorithms
72.4 Assumption and target
- Assumption There is a ct-competitive online
algorithm for the tree TG(Vt, Et) associated
with a graph G(V,E) -
(1)
- Target For the same algorithm, find a small
factor c for the graph G(V,E) , satisfying -
(2) -
82.5 Three steps to achieve it
- 1st step Find a method to construct an
associated tree which satisfies the following
conditions - 2nd step A tree TG can simulate the network G,
i.e. for any request sequence s, an algorithm
which produces congestion C when it is processed
on graph G,, also produces congestion
when it is processed on the tree TG. - 3rd step Prove that for any request sequence s,
an online algorithm which produces congestion Ct
when it is processed on TG,, also produces
congestion when it is processed on
G.
93.1 Construct a tree
?
A graph G(V,E)
Associated tree TG(Vt, Et)
103.2 Definitions
- Vt a node in TG
- SVt the cluster in G corresponding to Vt
- Bandwidth between two sets
-
- Bandwidth of edges leaving set X
- The height of TG h(TG)
- Set of all level nodes
-
113.2 Definitions (cont)
- Weight function
-
- For a subset X, the bandwidth of all edges that
are adjacent to nodes in X and that do not
connect nodes of the same cluster to . - One important property Additive
- Example
-
124.1How to Simulate G on TG
- A node is simulated by a node vt in
TG corresponding to cluster Svt v. - So, a message sent between node u and v in G is
sent along one unique path connecting the
respective counterparts in TG. - Example
- Our goal is to states that this simulation does
not increase the congestion.
134.2 Theorem 1
- Theorem 1 For any request s for an routing
problem on network G that can be processed with
congestion C, its simulation on TG yields
congestion
145.1How to Simulate TG on G
- A level node vt of TG is simulated by a
random node of the corresponding cluster Svt with
the probability
155.2 Theorem 2
- Theorem 2 The expectation of the relative load
L(e) of an edge in graph G, due to the simulation
of a tree strategy on G is bounded by
where Ct is the congestion on TG, hh(TG),
165.2.1 CMCF Problem
- Concurrent Multi-commodity Flow Problem (CMCF)
Each commodity fu,v has a flow size q.du,v,
where q is the maximized minimal throughput
fraction over all commodities, and -
175.2.2 absolute/relative load on G
- Expected number of messages have to be routed
between u and v is -
- The minimum capacity of edge is q.du,v, , so the
expected relative load at level l is at most
Ct/q, - Its expected relative load at all levels is
185.2.3 capacity ratio
- Therefore, the expected relative load on G has a
upper bound -
195.2.4 Next target
- So far, we show that a path of tree can be
simulated by a path in graph such that the
expected relative load of this path on the graph
has a upper bound. - Our goal is to show that the congestion in graph
also has a upper bound compared to that in tree,
i.e, to satisfy the lemma 2. So we should extend
the expected value to true value, that means
show - L(e)O(L(e))
205.3 Theorem 3
- Theorem 3 Give a graph G and an associated tree,
there exists an oblivious online routing
algorithm, which is -
-competitive with respect to the congestion.
215.3.1 Proof and Chernoff bound
- Let X1, X2,Xn be independent 0-1random
variables, i.e. - Pr (Xi1)pi,
- mp1p2pn,
- SX1X2Xn, then
-
225.3.2 Improvement to Theorem 3
- Theorem 4 Give a graph G, there exists a
associated tree that has height h(TG)O(log n),
maximum bandwidth ratio ?(TG) O(log n), and
maximum weight ratio d(TG)O(log n). - The online routing algorithm is
-competitive.
236. Conclusion and Application
- The paper proposes a method to construct a
associated tree regarding to a general network
and proves that the congestion on the network is
only a small factor c
larger than the congestion on the tree. - Since the tree topology is much simpler than
graph, we can study the routing algorithm on a
tree and also can get a good competitive
algorithm on the general network. It is a very
useful tool for research on routing problems on
general networks.
247. Reference
- Full paper The paper published in the
conference IEEE FOCS02 skips some proofs due to
space limitation. I contacted the author and got
the full paper. It is available to everyone If
you are interested. - Improved paper The author improve the results in
the following paper - A Practical Algorithm for Constructing
Oblivious Routing Schemes, published at
fifteenth annual ACM symposium on Parallel
algorithms and architectures.