PR
EV
IE
W
PR
EV
IE
W
Algorithms for Self-Healing Networks
by
Amitabh Trehan
B.Sc., Biology, Punjab University, 1994
W
M.C.A., Indira Gandhi National Open University, 2000
M.Tech., Indian Institute of Technology Delhi, 2002
IE
DISSERTATION
EV
Submitted in Partial Fulfillment of the
Requirements for the Degree of
PR
Doctor of Philosophy
Computer Science
The University of New Mexico
Albuquerque, New Mexico
May, 2010
UMI Number: 3409363
All rights reserved
INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.
In the unlikely event that the author did not send a complete manuscript
and there are missing pages, these will be noted. Also, if material had to be removed,
W
a note will indicate the deletion.
IE
EV
UMI 3409363
PR
Copyright 2010 by ProQuest LLC.
All rights reserved. This edition of the work is protected against
unauthorized copying under Title 17, United States Code.
ProQuest LLC
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106-1346
c 2010, Amitabh Trehan
W
IE
EV
PR
iii
Dedication
W
IE To the sun, the moon
and the intrepid spirit,
EV
To my family
who made this journey possible.
PR
The question walks
the length of pages,
rain drops on roof.
iv
Acknowledgments
W
If this were an Oscar awards ceremony, my list of thank yous would have had the
music director going crazy trying to hound me off the stage. There are many many
IE
to thank for the journey responsible for this document. My foremost gratitude goes
towards my advisor, Professor Jared Saia, for his constant enthusiastic guidance.
He has patiently ironed out a multitude of rough edges that I, as a scientist, have
presented, and has taught the virtues of discipline and mathematical rigour. My
EV
academic collaborator and committee member Professor Thomas Hayes has been a
source of constant inspiration. I am thankful to my dissertation committee (Pro-
fessors Saia, Hayes, Cris Moore and Tanya-Beger Wolf) who have provided me with
much insight and guidance. I am thankful to all my close friends, who have been
PR
with me through good times and bad, especially Navin Rustagi and Vaibhav Mad-
hok (their lively discussions have lit up many evenings!). I am thankful to the US
educational system, for its support of quality graduate education and research. I owe
a debt of gratitude to all my teachers and friends in India, and to the Art of Living
foundation and it’s founder Sri Sri Ravi Shankar, for Sudershan Kriya, the medita-
tion and the satsangs . Finally, I have to thank my biggest inspiration: my mother,
and my family: my late father, my step-father, my brothers, my sister-in-laws, my
nieces and my nephew, without whose support and love I would never have been able
to pursue the path around the world and in my academic world that I have.
v
Algorithms for Self-Healing Networks
by
W
Amitabh Trehan
IE
ABSTRACT OF DISSERTATION
EV
Submitted in Partial Fulfillment of the
Requirements for the Degree of
PR
Doctor of Philosophy
Computer Science
The University of New Mexico
Albuquerque, New Mexico
May, 2010
Algorithms for Self-Healing Networks
by
Amitabh Trehan
B.Sc., Biology, Punjab University, 1994
M.C.A., Indira Gandhi National Open University, 2000
W
M.Tech., Indian Institute of Technology Delhi, 2002
Ph.D., Computer Science, University of New Mexico, 2010
IE
EV
Abstract
Many modern networks are reconfigurable, in the sense that the topology of the
PR
network can be changed by the nodes in the network. For example, peer-to-peer,
wireless and ad-hoc networks are reconfigurable. More generally, many social net-
works, such as a company’s organizational chart; infrastructure networks, such as an
airline’s transportation network; and biological networks, such as the human brain,
are also reconfigurable. Modern reconfigurable networks have a complexity unprece-
dented in the history of engineering, resembling more a dynamic and evolving living
animal rather than a structure of steel designed from a blueprint. Unfortunately, our
mathematical and algorithmic tools have not yet developed enough to handle this
complexity and fully exploit the flexibility of these networks.
We believe that it is no longer possible to build networks that are scalable and
never have node failures. Instead, these networks should be able to admit small, and
vii
maybe, periodic failures and still recover like skin heals from a cut. This process,
where the network can recover itself by maintaining key invariants in response to
attack by a powerful adversary is what we call self-healing.
Here, we present several fast and provably good distributed algorithms for self-
healing in reconfigurable dynamic networks. Each of these algorithms have different
properties, a different set of gaurantees and limitations. We also discuss future
directions and theoretical questions we would like to answer.
W
IE
EV
PR
viii
Contents
W
List of Figures xiii
List of Tables
IE xviii
EV
1 Introduction 1
1.1 Naive self-healing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Model of self-healing . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
PR
1.3 Healing by Reconstruction Trees . . . . . . . . . . . . . . . . . . . . . 8
1.4 Our Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5.1 Self-healing and Self-* properties . . . . . . . . . . . . . . . . 13
1.6 Structure of the document . . . . . . . . . . . . . . . . . . . . . . . . 14
2 DASH 16
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 DASH: An Algorithm for Self-Healing . . . . . . . . . . . . . . . . . 20
ix
Contents
2.2.1 DASH: Degree Assisted Self-Healing . . . . . . . . . . . . . . 20
2.2.2 Towards the proof of Theorem 2.1 . . . . . . . . . . . . . . . . 22
2.2.3 The Record Breaking Problem . . . . . . . . . . . . . . . . . . 32
2.2.4 Proof of Theorem 2.1 . . . . . . . . . . . . . . . . . . . . . . . 34
2.3 Lower bounds on Locality-aware algorithms . . . . . . . . . . . . . . 34
2.3.1 Necessity of Component tracking for healing strategies . . . . 34
W
2.3.2 A lower bound on healing by Degree-bounded locality-aware
healing algorithms . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.3
IE
A general lower bound on healing by locality-aware algorithms 40
EV
2.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.4.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.4.2 Attack Strategies . . . . . . . . . . . . . . . . . . . . . . . . . 46
PR
2.4.3 Healing strategies . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.4.4 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.5 Degree increase . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.6 Messages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.7 Heuristics and experiments involving Stretch . . . . . . . . . . 50
2.5 Conclusions and future work . . . . . . . . . . . . . . . . . . . . . . . 52
3 Forgiving Tree 55
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
x
Contents
3.2 Delete and Repair Model . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3 The Forgiving Tree algorithm . . . . . . . . . . . . . . . . . . . . . . 60
3.3.1 Distributed implementation . . . . . . . . . . . . . . . . . . . 63
3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.4.1 Upper Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.4.2 Lower Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
W
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4 Forgiving Graph
4.1
IE
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
88
89
EV
4.2 Node Insert, Delete and Network Repair Model . . . . . . . . . . . . . 91
4.3 The Forgiving Graph algorithm . . . . . . . . . . . . . . . . . . . . . 96
PR
4.4 Half-full Trees (“HAFTS”) . . . . . . . . . . . . . . . . . . . . . . . . 98
4.4.1 Operations on Hafts . . . . . . . . . . . . . . . . . . . . . . . 101
4.5 FG: Distributed implementation . . . . . . . . . . . . . . . . . . . . . 104
4.5.1 Representative mechanism . . . . . . . . . . . . . . . . . . . . 111
4.6 Real graph from the Forgiving Graph . . . . . . . . . . . . . . . . . . 114
4.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.7.1 Upper Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.7.2 Lower Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
xi
Contents
5 Future Directions 131
5.1 Empirical study of self-healing algorithms beyond assumptions . . . . 131
5.2 Routing in Self-healing structures . . . . . . . . . . . . . . . . . . . . 132
5.3 Load balanced Self-healing . . . . . . . . . . . . . . . . . . . . . . . . 132
5.4 Self-healing in Sensor Networks . . . . . . . . . . . . . . . . . . . . . 133
5.5 Self-healing/ Behavioral robustness in Social Networks . . . . . . . . 134
W
5.6 Self-* problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.7 Evolution of social and computer networks
IE
and study of group formation . . . . . . . . . . . . . . . . . . . . . . 135
5.8 Byzantine agreement: Distributed computing in presence of byzantine
EV
faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
References 139
PR
xii
List of Figures
W
1.1 A sequence of 3 deletions and healings using a naive algorithm. A
node marked red is deleted by the adversary. The neighbors of the
IE
deleted node reconnect (golden edges) to maintain connectivity. No-
tice node v increases its degree by 3. . . . . . . . . . . . . . . . . . . 5
EV
1.2 The general distributed Node Insert, Delete and Network Repair Model. 7
1.3 Graphs at time T. G0T : The graph of initial nodes and insertions over
time, GT : The actual healed graph. . . . . . . . . . . . . . . . . . . 8
PR
1.4 Deleted node x (in red, crossed) replaced by a Reconstruction Tree,
which is a structure formed by its neighbors (a, b, c, d, j). . . . . . . 8
1.5 A timeline of deletions and self healing in a network with 100 nodes.
The gray edges are the original edges and the red edges are the new
edges added by our self-healing algorithm. . . . . . . . . . . . . . . . 10
2.1 W (T (v, m)) ≥ rem(v). . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 node v is the root, with 2 children . . . . . . . . . . . . . . . . . . . 27
2.3 Internal node v with 1 child . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Internal node v with 2 children . . . . . . . . . . . . . . . . . . . . . 29
xiii
List of Figures
2.5 Steps in Prune(v,x). Leaf nodes are deleted at each step. . . . . . . 36
2.6 An internal node in a 3-node line reconnection suffers a degree increase. 36
2.7 M+2 -ary Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.8 Strategy-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.9 A timeline of deletions and self healing in a network with 100 nodes.
The gray edges are the original edges and the red edges are the new
edges added by our self-healing algorithm. . . . . . . . . . . . . . . . 49
W
2.10 Maximum Degree increase: DASH vs other algorithms . . . . . . . . 50
2.11
2.12
IE
ID changes for nodes . . . . . . . . . . . . . . . . . . . . . . . . . .
Number of messages exchanged for Component(ID) information main-
51
EV
tenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.13 Stretch for various algorithms . . . . . . . . . . . . . . . . . . . . . . 54
PR
3.1 Deleted node v replaced by its Reconstruction Tree. The nodes in the
oval are helper nodes. Regular helper nodes are depicted by circles
and the heir helper node by a rectangle. . . . . . . . . . . . . . . . . 62
3.2 The leftmost column shows a small segment of the network. The
RT(x) corresponding to this figure is shown. Every neighbor of node
x stores the portion of RT(x) relevant to it. Each rectangular box is
labelled with a neighbor and shows the portions and the value of the
corresponding fields . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.3 An illustrative sequence of deletions and healings. . . . . . . . . . . 66
3.4 The states of a node with respect to helper duties: Waiting, Ready
and Deployed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
xiv
List of Figures
3.5 Various cases of Leaf deletions . . . . . . . . . . . . . . . . . . . . . 72
3.6 Deletion of the central node v of a star leads to an increase in the
diameter. Here, the healing algorithm increases the degree of any
node by at most α. . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.1 The Node Insert, Delete and Network Repair Model – Distributed
View. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
W
4.2 Graphs at time T. G0T : The graph of initial nodes and insertions over
time, GT : The actual healed graph. . . . . . . . . . . . . . . . . . .
IE 95
4.3 Comparing degrees: In the figure the degree of node v in graph of
EV
only original and inserted nodes is 3, and in the actual healed network
it is 5. The nodes in red (dark gray in grayscale) were deleted by the
adversary and the golden (light shaded) edges were the ones added
PR
by the healing algorithm. . . . . . . . . . . . . . . . . . . . . . . . . 95
4.4 Comparing distances: In the figure nodes u and w have their distance
increased to 5 in the actual healed network compared to their distance
of 3 in the graph of only original and inserted nodes. The nodes in red
(darker in grayscale) were deleted by the adversary and the golden
edges (lighter shade) are the ones added by the healing algorithm . . 96
4.5 Deleted node v replaced by its Reconstruction Tree. The triangle
shaped nodes are ’virtual’ helper nodes simulated by the ’real’ nodes
which are in the leaf layer. . . . . . . . . . . . . . . . . . . . . . . . 96
4.6 haft (half-full tree) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
xv
List of Figures
4.7 Deletion of a node and its helper nodes lead to breakup of RT into
components. The Strip operation or a simple variant (for non-hafts)
returns a set of complete trees, which can then be merged. . . . . . . 103
4.8 Merging three hafts. The vertices in the square boxes are the new
isolated vertices used to join the complete The square shaped vertices
are the isolated vertices used to join the complete trees. Merging is
analogous to binary number addition, where the number of leaves are
W
represented as binary numbers. . . . . . . . . . . . . . . . . . . . . . 104
4.9 Effect of 3 deletions on a graph. The RT for each deleted node
IE
consists of the helper nodes, plus the neighbors of the deleted node
which form the leaves of the tree. In this example, the deleted nodes
EV
form an independent set, so the structure of the RTs does not depend
on the deletion order. . . . . . . . . . . . . . . . . . . . . . . . . . . 105
PR
4.10 Equivalent Representations of a RT. . . . . . . . . . . . . . . . . . . 106
4.11 On deletion of a node v, The RTfragments to be merged are con-
nected by a binary tree BTv . The leaf RTfragments merge with their
parents till a single RT is left. The solid circles are the primary roots.
The (red color) nodes in the square boxes are spine nodes removed
at each step. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.12 The underlined node d and corresponding helpers are deleted. This
leads to the graph breaking into components which are then merged
using BTd (the binary tree of anchors) and the primary roots in the
components. The dashed edges show the representative for that node. 109
xvi
List of Figures
4.13 Merging with representatives: Two singleton hafts of real nodes a and
b merge. Here a creates the parent helper node, and this helper node
inherits the representative of its right child (b) as its representative.
Notice b is the unique real node in a.helper’s subtree that is not
simulating a helper node. With regard to merging, the root nodes
representatives are ’active’ (shown in pink, dashed outline), while
others are ’dormant’ (shown in green, dotted outline). . . . . . . . . 112
4.14 Reusing representative information: RTs split into complete trees on
W
deletion of node a. A node always has a representative assigned to it
at birth and it never changes its representative. In the figure, node c0
IE
has d as its representative:- ’dormant’ before the split (green, dotted
outline), ’active’ afterwards (pink, dashed outline). . . . . . . . . . . 114
EV
4.15 The actual graph G (on the right) is a homomorphic image of the
Forgiving Graph FG (left) where the helper nodes are mapped to the
nodes simulating them. Note both the node degrees and distances
PR
between nodes in the real graph cannot be more than those in the
Forgiving Graph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.16 Proof by contradiction: Case 1. Two helper nodes in different RTs. . 117
4.17 Proof by contradiction: Case 2(a). Two helper nodes in same RT,
but in different subtrees. . . . . . . . . . . . . . . . . . . . . . . . . 117
4.18 Proof by contradiction: Case 2(b). Two helper nodes in the same
subtree. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.19 Deletion of the central node v of a star leads to an increase in the
stretch. Here, the healing algorithm can increase the degree of any
node by at most a factor of α. . . . . . . . . . . . . . . . . . . . . . 123
xvii
List of Tables
W
1.1 Comparison of our self-healing Algorithms. d is the degree of an
individual node, ∆ is the maximum degree of a node in the graph,
and δ is the degree of the deleted node. . . . . . . . . . . . . . . . .
IE 12
3.1 The fields maintained by a node v . . . . . . . . . . . . . . . . . . . 64
EV
4.1 The fields maintained by a processor v for edge(v, x), which is an
edge in G0 , the graph of only original nodes and insertions. Here RT
PR
refers to the reconstruction tree of which v : edge(v, x) is a part. . . 107
xviii
Chapter 1
Introduction
W
IE Begin at the beginning and go on till
you come to the end: then stop.
EV
The king of hearts
Alice in Wonderland
PR
Networks in the modern age have grown by leaps and bounds, both in size and
complexity. The size of some networks spans nations and even the globe. Networks
provide a multitude of services using a wide variety of protocols and components to
the extent that they have now begun to resemble self-governed living entities. The
Internet is the obvious example but there are others too like cellular phone networks.
There are networks which have always been around but which only now have been
scrutinized by tools of computer science, such as the social networks. Most networks
are dynamic since nodes can enter the network or be removed by choice, failure or
attack. We are also fortunate that we live in a time where we can observe and inuence
the evolution of a dynamic network like the Internet. Due to the scale and nature of
design of modern networks, it may simply not be practical to build robustness into
the individual nodes or into the structure of the initial network itself.
1
Chapter 1. Introduction
Many important networks are also reconfigurable in the sense that they can change
their topology. Often, individual nodes can initiate new connections or drop exist-
ing connections. For example, peer-to-peer, wireless and ad-hoc networks are re-
configurable. Looking beyond computer networks, many social networks, such as a
company’s organizational chart, or friendship networks on social networking sites are
reconfigurable. Infrastructure networks, such as an airline’s transportation network
are reconfigurable. Many biological networks, including the human brain, which
shows such capacity for learning and adaptability, are also reconfigurable. From
an engineering aspect, modern reconfigurable networks have a complexity unprece-
W
dented in history. We are approaching scales of billions of components. Such systems
are less akin to a traditional engineering enterprise built from a blueprint such as a
IE
bridge, and more akin to a dynamic and evolving living organism in terms of com-
plexity. A bridge must be designed so that key components never fail, since there
EV
is no way for the bridge to automatically recover from system failure. In contrast,
a living organism can not be designed so that no component ever fails: there are
simply too many components. For example, skin can be cut and still heal. Designing
PR
skin that can heal is much more practical than designing skin that is completely
impervious to attack. Unfortunately, current algorithms ensure robustness in com-
puter networks through hardening individual components or, at best, adding lots of
redundant components. Such an approach is increasingly unscalable.
Our mathematical and algorithmic tools have not yet developed enough to handle
the complexity and fully exploit the flexibility of modern networks. As an example,
on August 15, 2007 the Skype network crashed for about 48 hours, disrupting service
to approximately 200 million users [17, 42, 46, 51, 55]. Skype attributed this outage
to failures in their “self-healing mechanisms” [2]. We believe that this outage is
indicative of the much broader problems outlined earlier.
2
Chapter 1. Introduction
In the following chapters, we will propose some algorithms for self-healing. Infor-
mally, we define self-healing to be maintenance of certain properties within desirable
bounds by the nodes in a network suffering from failures or under attack. As the
name implies, self-healing has to be initiated and executed by the nodes themselves.
As such, the algorithms we have proposed here are fully distributed. Equivalenty we
can say that a self-healing system, when starting from a correct state, can only be
temporarily out of a correct state i.e. it recovers to a correct state, in presence of
attacks. Self-healing is one of the so called ‘Self-*’ properties which systems such as
autonomic systems may be required to have. Section 1.5.1 has a brief discussion on
W
these properties.
IE
One approach towards self-healing is to add additional capacity or rerouting in
anticipation of failures. There has been plenty of work which has followed this
EV
approach. However, there are obvious limitations including wastage of resources and
limitations on additional capacity. In this Dissertation, we have adopted a responsive
approach. Our approach is responsive in the sense that it responds to an attack
PR
(or component failure) by changing the topology of the network. This approach
works irrespective of the initial state of the network, and is thus orthogonal and
complementary to traditional non-responsive techniques.
Informally, the model we adopt in this work is as follows. We assume that the
network is initially a connected graph over n nodes. An adversary repeatedly attacks
the network. This adversary knows the network topology and our algorithm, and it
has the ability to delete arbitrary nodes from the network or insert a new node in
the system which it can connect to any subset of the nodes currently in the system.
However, we assume the adversary is constrained in that in any time step it can only
delete or insert a single node. Following that, the self-healing algorithm has a short
time to reconfigure and heal the network by adding edges between remaining nodes
before the next act of the adversary. Our model captures what can happen when a
3
Chapter 1. Introduction
worm or software error propagates through the population of nodes. This model is
described in more detail Section 1.2.
1.1 Naive self-healing
Even in a very simple setting, we need to be smart about reconfiguring. Suppose
we are trying to maintain a property such as connectivity of the network but our
algorithm is not very sophisticated. Then, it may be very easy for the adversary to
W
force the algorithm to cause high degree increase (which may lead to overload and
eventual network breakdown) or increase in distances between nodes (which may
IE
lead to poor communication). Figure 1.1 shows a naive algorithm attempting to heal
the network by using only a small number of edges at each timestep. However, node
EV
v in the figure ends up increasing its degree by 3 over a course of 3 deletions. Thus,
a naive algorithm could yield a degree increase as high as θ(n).
PR
1.2 Model of self-healing
Our general model of self-healing is shown in Figure 1.2. The specific models used
in our algorithms are special cases of this model, differing mainly in the way the
success metrics of the graph properties are presented. This model is very similar
to the model described in Figure 3.2.1. Let G = G0 be an arbitrary graph on
n nodes, which represent processors in a distributed network. In each step, the
adversary either deletes or adds a node. After each deletion, the algorithm gets to
add some new edges to the graph, as well as deleting old ones. At each insertion,
the processors follow a protocol to update their information. The algorithm’s goal
is to maintain the chosen graph properties within the desired bounds. At the same
time, the algorithm wants to minimize the resources spent on this task. Initially,
