Title: Scalable Methods for Provisioning and Restoration of QoS Paths and Trees
1Scalable Methods for Provisioning and Restoration
of QoS Paths and Trees
- Alex Sprintson
- Advisor Prof. Ariel Orda
2Motivation
- Networks grow in size at a rapid pace
- Emergence of new applications
- e.g., voice over IP, multimedia streaming
- Requirements
- Support for Quality of Service (QoS)
- Resilience to failures
- Network mechanisms must be scalable
- We investigate scalable network mechanisms for
the support of QoS and failure resilience
3Mechanisms
- We focus on two major network mechanisms
- Provisioning establishing a suitable routing
topology (paths or trees) - Restoration establishing a restoration
topology, i.e., a set of restoration paths or
trees, each protecting part of the primary path
4QoS constraints
- Imposed by applications
- e.g., voice over IP, requires
- bandwidth 16 - 64 Kbs
- delay 100 and 300 ms
- Can be divided into
- bottleneck constraints - e.g., bandwidth
- additive constraints - e.g., delay and jitter
- Both primary and restoration topologies must
satisfy QoS constraints
5Models
- Basic model
- Each link can provide a certain QoS guarantee at
certain cost - Extended model
- Each link can provide several QoS guarantees at
different costs - Link is associated with a cost function that
assigns a cost to each QoS value - The cost estimates the amount of resources
consumed in order to provide the QoS guarantee
6Extended Model
- How to allocate resources in order to provide
the required QoS?
- Each link is a provider's network
- Cost per delay is given by Service Level
Agreements (SLAs) - Less delay for more cost
7Extended Model
8Methods
- Scalability is achieved in two ways
- We establish algorithmic solutions that are
considerably less dependent on the size of
network - We employ a precomputation approach
- reduce the time required for computing a path by
performing computations in advance
9Precomputation
The main loop of a network element
Idea Reduce the time required to identify a
solution by precomputing solutions for each
possible delay value
10Problems considered
?-precomputation methods
We provide rigorous solutions, with proven
performance guarantees
11Practical Applications
- ATM PNNI recommendations
- Need to find paths and trees that satisfy QoS
constraints - Bandwidth is reserved for a long period of time
- IntServ/RSVP
- Unsuitable paths and trees incur major overhead
- DiffServ
- Identify paths and trees that satisfy SLA
(service level agreements - MPLS
- Level-switched paths
12Model
- Directed Graph G(V,E)
- Undirected for symmetric networks
- For each link l? E
- dl- the delay of link l
- cl - the cost of link l
- Path (P) Sequence of distinct nodes v0, v1, ,vn
- Tree (T) - A subgraph of G with a source node s
such that each node is reached from s by a unique
path
13Provisioning of QoS paths
- Example find a path that satisfies delay
constraint D6
(2,6)
u
v
1
2
(2,5)
(2,2)
(1,5)
t
s
(1,4)
(5,4)
(1,2)
(2,3)
(2,4)
(2,4)
v
u
1
2
14Provisioning of QoS paths
- Find a minimum cost path between s and t that
satisfies QoS constraint D
- Can be efficiently solved for bottleneck
constraints - For additive constraints, the problem is NP-hard
- Approximation schemes (PTAS) H92, LR01
- ?-approximate solutions
- O(E?V ?(1/?loglogV))
15Provisioning of QoS Trees
- Example find a tree that satisfies delay
constraint D6
16Provisioning of QoS trees
- Find a minimum cost tree T that connects source
s to each terminal tj?X and satisfies the delay
constraint D - Related problems
- Directed Steiner Tree (DST)
- Special case with no QoS constraints
- Extensively investigated for undirected networks
- Directed Networks Algorithm by Charikar et
al.,99 - Restricted Shortest Path (RST)
- Special case for unicast
17Provisioning of QoS trees
- Problem has attracted a large body of research
- Most studies proposed heuristic solutions
- often based on restricting assumptions such as a
symmetry of link delays - Probable solutions available only for special
cases - E.g., identical link delays
- Or incurred large violation of QoS constraint,
- e.g, an O(Log N, log N) solution
- No solution of provable performance has been
established for general networks and no violation
of QoS constraints.
18Our Results
19Solution methodology
20Solution to DST problem
- Procedure Ai(K,s,Y)
- T??
- while Kgt0
- Tbest?? Kbest?0
- for each link (r,v)?G and each 1?K'?K do
- T'?Ai-1(K,v,Y)?(r,v)
- if then
- Tbest?T' Kbest?K'
- Y'?tj tj ?Y ? tj ? Tbest
- T?T ? Tbest
- Y ?Y/Y'
- K ?Y
- return T
Approximation ratio
Complexity
21Preliminaries - Graph Unfolding
22T1-reduction
23T1-reduction
- Build a Layers Graph which allows to distinguish
between trees that satisfy QoS constraint D and
all other trees
24T1-reduction
- Any approximation scheme for problem DST can be
employed in order to obtain an approximate
solution for problem RST - E.g., Feldman Ruhl, 99 presents an optimal
solution for problem DST for a small number of
terminals - Computation complexity
- Depends on the delay constraint D
25Lemma Zelicovsky,97
- Let G' be a transitive closure of graph G. Then,
for each tree T?G' that connects source s with a
group X of terminals and for each i, 1 ? i? log
X there exists an i-level tree T' in G' that
connects s with X such that
We identify trees with small number of layers
26T2-reduction
- Efficient, but violates the delay constraint
- We look for a tree that includes at most i layers
- Delay scaling
27T2-reduction
28T2-reduction
29T2-reduction
- Computational complexity
- May violate the delay constraint by a factor of
- Returns a solution whose cost is at most
- higher than the optimum
30T3-reduction
- Employs several techniques
- Graph Unfolding
- Layers Graphs
- Cost Scaling
- Path Aggregation
31Graph Unfolding
32Layers Graph
33(No Transcript)
34Cost Scaling
B is an estimate of the optimum cost
35Path aggregation
- reduces the size of the auxiliary graph
- represent large number of paths by a small subset
S
36T3-reduction
- Construct a layers graph in which the outdegree
of each node is S
37T3-reduction
- Begin with initial Lower and Upper bounds on B,
which are iteratively improved - Computational complexity
- Does not violates the delay constraint
- The first solution
38Conclusions
- We focused on the fundamental problems in
provisioning and restoration of QoS paths and
trees - Considered generic models, that capture most
practical settings of networks - Practical applications IntServ, DiffServ, MPLS,
VPN, overlay networks
39What we did and what is tbd
40Future work
- Sharing of backup paths
- Topology aggregation
- Reducing complexity of network mechanisms
- End system multicast
- Investigate trade-off between complexity and
performance - Monitoring and enforcement of network protocols
- Similar to the enforcement of social laws