Transcript 4) Distributed Cache Management

Vasilis Sourlas – PhD defense – 23/7/2013
Electrical and Computer Engineering,
University of Thessaly
“Replication Management and Cache aware
Routing in Information-Centric Networking”
Vasilis Sourlas
Dissertation Committee:
Leandros Tassiulas (UTH,GR), Supervisor
Spyros Lalis (UTH, GR)
George Pavlou (UCL, UK)
1
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
2
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
3
Internet-based Content
Vasilis Sourlas – PhD defense – 23/7/2013

The vast majority of interactions relate  New approaches are required to cater
to content access:
for the explosion of video-based
content and for creating novel use
 P2P overlays (BitTorrent)
experiences.
 Media aggregators (YouTube)
 Continue throwing more capacity
 Content Delivery Networks
cannot work anymore!
(Akamai)
 Social Networks (Facebook)
 Photo sharing sites (Picasa)
4
Expected IP Traffic Growth 2012-2017
Vasilis Sourlas – PhD defense – 23/7/2013

According to the Cisco Visual Networking Index:
Global IP traffic will reach 1.3 zettabytes per year.
3 networked devices per capita in 2016 vs 1 per capita in
2011.
15 GBytes per capita IP traffic in 2016 vs 4 GBytes in 2011.
Approx. 55% of the overall Internet traffic will be video by
2016, without counting P2P video file sharing (~ 86%
including P2P).

It will take over 5 years to watch the amount of video
that will cross global IP networks every second in
2015!!

What is exchanged is becoming more important
than who are exchanging it.
5
Vasilis Sourlas – PhD defense – 23/7/2013
Information-Centric Networking
Paradigm shift from the host-to-host
Internet to a host-to-content one.
 Information-Centric Networking (ICN)
targets a general infrastructure that
provides in-network caching and multicast
communication so that content is
distributed in a scalable, cost-efficient &
secure manner.

6
ICN Architectural Models
Vasilis Sourlas – PhD defense – 23/7/2013

Information-centric (Content-Centric) networks
Content is explicitly named.
Subscriptions/Interests act on the name of each
packet.
One-time fetch and ongoing subscribe operation.
DONA, PURSUIT, NDN/CCN, SAIL, ...

Content-Based Publish/Subscribe (CBPS)
networks
Overlay event notification services.
Broader request semantics (attribute/value scheme).
One-time fetch only operation.
No content servers assumed.
IBM Gryphon, Siena, REDS, Elvin, …
7
CCN Operation
Vasilis Sourlas – PhD defense – 23/7/2013
Check Pending
Interests Table
Interest
Data
Check Content Store
Check Content Store
Check Pending
Interests Table
S1: /spiegel.com/crisisingreece/news.pdf/page34/....
S1: /spiegel.com/crisisingreece/news.pdf/page34/....
Check Forwarding
Information Base
8
Vasilis Sourlas – PhD defense – 23/7/2013
CBPS Operation
Publish(
)
S1: [type,=,movie/english], [artist,=,Bruce Lee],[year,=,*]
S2: [type,=,music/mp3], [artist, =, madonna], [album, =, *], [year, >, 1990]
P: [type,=,movie/english], [artist, =, Bruce Lee, Chuck Norris], [title, =, BLvsCN.avi]
9
Research Challenges
Vasilis Sourlas – PhD defense – 23/7/2013

Information-Centric Networking Research
Group (ICNRG)
Cache management
 Traffic engineering
Scalable Routing
QoS approaches
Novel caching strategies
…….
10
Work Synopsis
CDN-like replication management
framework.
 In-network opportunistic caching
framework.
 Cache aware routing scheme.
Vasilis Sourlas – PhD defense – 23/7/2013

11
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
12
Introduction
CDN-like replication distributes a site’s
content across multiple mirror servers.
 A request is redirected to the “closest”
server.
 Replication is used to increase availability
and fault tolerance.
 Side effects: load balancing and enhanced
publisher/subscriber proximity.
Vasilis Sourlas – PhD defense – 23/7/2013

13
Contributions
Vasilis Sourlas – PhD defense – 23/7/2013

A three phase replication management
framework for ICN
I.
Planning phase

Decides the placement of the replication points.
II. Off-line Assignment phase


Assignment of information items to the replication
points based on the observed popularity.
Generalized assignment problem (reduced to NPcomplete multiple knapsack problem).
III. On-line Replacement phase

Replacement of information items in real-time, based
on the changing demand pattern.
14
Replication Framework (off-line)
Vasilis Sourlas – PhD defense – 23/7/2013
Storage
Planning
Network
Topology
Replica
Assignment
Monitoring
each node
Long-term
forecast
Mediumlong term
forecast
Subscription
Forecast
Sub
Data
Monitor
subscriptions
Configure (subscribe item t2, publish item t2)
Configure (subscribe item t1, publish item t1)
Storage
device
Storage
device
Forwarding
Nodes
Forwarding
Nodes
Subscribers
…
…
…
…
…
Subscribers
15
Replication Framework (dynamic)
Vasilis Sourlas – PhD defense – 23/7/2013
Replace
item i with
item j?
Cache
Managers
Cache
enabled ICN
node
Cache Replacement
Substrate
client request rates,
topology, cache configs
coordinate
decisions
Clients
request for
items
16
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
17
Introduction
Replication is thoroughly investigated in the
area of CDNs (approximate solutions and
optimal algorithms for tree topologies).
 2-aproximation algorithms have been
proposed also in the area of distributed
replication groups.
 In ICN only approaches based on distributed
databases.
 Less attention has been given to network
constraints (limited storage capacity).
Vasilis Sourlas – PhD defense – 23/7/2013

18
Contributions
Vasilis Sourlas – PhD defense – 23/7/2013




Enhanced the CBPS with an advertisement and a
request/response mechanism.
Modified known Greedy algorithm (CDN context).
Used the modified greedy for the proposed placement
algorithm.
Proposed a new algorithm for the selection of R storage
points among the V network nodes (R < V) based on:
a) the locality and the popularity of the interests for each item
b) the targeted “replication degree km” of each item m
c) the storage capacity “L” of each replication device
Proposed two alternative assignment mechanisms.
 Target - Minimize client’s response latency subject to
installing the minimum number (or any given number)
of replicas in the network.

19
Greedy Algorithm
1st round: evaluates each of the V nodes to
determine its suitability to become a
storage. Computes the Gain (traffic served
by replica and does not need to access
original server) associated with each node
and selects the one that maximizes the Gain.
 2nd round: searches for a second storage
which, in conjunction with the storage
already picked, yields the highest Gain.
 Completes: iterates until the requested
number of storages have been chosen for
the replication of the specific server.
Vasilis Sourlas – PhD defense – 23/7/2013

20
Modified Greedy Algorithm
No knowledge of the location of the
server, differently there is no server at all.
 Repeat Greedy alg V times (server j is a
different node of the network).
 V vectors of possible storages.
 Choose as our storages those nodes that
appeared more times in the per element
summation of the V vectors.
Vasilis Sourlas – PhD defense – 23/7/2013

21
Planning and Assignment
Vasilis Sourlas – PhD defense – 23/7/2013
Planning
Steps:
1. For each item m we execute the modified greedy algorithm and
we get M vectors of possible storages.
2. Each vector is weighted by each item’s weight (significance
regarding the traffic of each item in the network).
3. Select as storages those M nodes that appeared more times in the
per element weighted summation of the M vectors.
4. For each item m starting from the most significant (based on the
weight) assign km storages following the procedure below:

For each entry in the vector of item m calculated in step 1 assign a storage if that
entry also appears in the final storage nodes calculated in step 3 and only if in that
storage has been assigned less than L items until we get km storages.

A similar weighted round robin-like mechanism based on the weight of each
item has also been proposed.
Assignment
22
Vasilis Sourlas – PhD defense – 23/7/2013
Evaluation
Compare to:
 “grd_opt”: each item m is assigned to the
km storages produced by the first step of the
placement algorithm
 “rnd”: no differentiation among items,
random assignment after the selection of
the storages
Metrics:
Mean hop distance between the requesting
client and the storage (indicative of the
response latency)
23
Vasilis Sourlas – PhD defense – 23/7/2013
Predefined Minimum Replication Degree
Off-line Assignment Phase Evaluation
24
Results
The proposed planning and the two offline placement algorithms perform only
1%-5% worse than greedy, using 50%-80%
less storages.
 Appropriate solution for real world
scenarios where a storage provider has
limitations in the number of replicas that
can install.
Vasilis Sourlas – PhD defense – 23/7/2013

25
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
26
Contributions
Proposed a distributed cache management
architecture that dynamically (re-)assigns
information items to caches, based on items’
demand patterns in order to minimize the overall
network traffic.
 Presented four distributed on-line cache
management algorithms, categorized them based
on the level of cooperation needed between the
managers and compared them against their
performance, complexity, message overhead and
convergence time.
 Derived a lower bound of the overall network
traffic for regular network topologies.
Vasilis Sourlas – PhD defense – 23/7/2013

27
Distributed Cache Management
Architecture
Distributed Cache Managers (CM) decide
in a coordinated manner whether to
cache an item and replace an already
cached.
 Every CM should have a holistic networkwide view of all the cache configurations
and the demand patterns.
 Upon a change in a cache configuration
the CM should inform (event-based
manner) every other CM in the network.
Vasilis Sourlas – PhD defense – 23/7/2013

28
Distributed On-Line Cache Management Algorithms
Vasilis Sourlas – PhD defense – 23/7/2013

Known global demand patterns and global
replica placement (global cache configuration), minimize overall network traffic
1. Cooperative Cache Management Algorithm
2. Holistic Cache Management Algorithm
3. Holistic-all Cache Management Algorithm

Known local demand patterns and global
replica placement, minimize local traffic
(local clients)
4. Myopic Cache Management Algorithm
29
Cooperative Algorithm
1.
Each CM computes:
Vasilis Sourlas – PhD defense – 23/7/2013





2.
3.
For each item m in the cache the performance loss lm if item m is removed
from the cache.
For each item m not in the cache of the performance gain gm if item m is
cached.
Candidate for insertion the item of maximum performance gain.
Candidate for replacement the items of minimum performance loss.
Maximum local relative gain r= gm - lm and report it to the rest CMs.
CMs calculate the most network-wide beneficial replacement and
updated their configuration matrix.
Steps 1-2 are repeated until no further replacements are
beneficial for the network.
Each replacement decreases the overall network traffic converges to an equilibrium point (local minimum given the initial
cache configuration).
Vasilis Sourlas – PhD defense – 23/7/2013
Holistic Algorithm
Holistic-all Algorithm
Only one CM runs the algorithm at a time
e.g. token based decision making.
 In holistic only one replacement at each
node per iteration.
 In holistic-all all possible replacements at
each node per iteration.

31
Myopic Algorithm
In highly dynamic environments each CM may
don’t have info about the demand pattern in
the network.
 Decision based on local info only w.r.t to local
requests but every CM is aware of the global
cache configurations.
 Each CM calculates its replacements in order to
minimize the traffic cost for the demand it
serves.
 Same decision making as the holistic.
Vasilis Sourlas – PhD defense – 23/7/2013

32
Network Traffic Lower Bound
Vasilis Sourlas – PhD defense – 23/7/2013

Assumptions
Uniform request pattern.
Unit size information items.
Regular network topologies (distance regular
graphs, n-dim torus).
Theorem:
33
Vasilis Sourlas – PhD defense – 23/7/2013
Evaluation
Metrics:
Overall network traffic, ONT (reqs*hops/sec) at
equilibrium.
Total number of replacements per node, RE.
Total number of iterations per node, IT (indicative of
the running time).
Two sets of experiments
1.
2.
Uniform demand pattern
Synthetic workload & Zoo Topologies
34
Vasilis Sourlas – PhD defense – 23/7/2013
Uniform Demand Pattern
35
Vasilis Sourlas – PhD defense – 23/7/2013
Synthetic Workload & Zoo Topologies
36
Vasilis Sourlas – PhD defense – 23/7/2013
Convergence & Mean Cache Hit
Distance
37
Results
Vasilis Sourlas – PhD defense – 23/7/2013


The algorithms that use network-wide
information are near optimal since the
corresponding difference from the lower
bound varies between 0,5% and 3,6%
regardless of the topology, the size of the
network and the storing capacity of each
cache and the initial cache assignment.
Network wide knowledge and cooperation
give significant performance benefits and
reduce the time to convergence at the cost
of additional message exchanges and
computational effort.
38
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
39
Introduction
In-network opportunistic caching is a salient
characteristic of ICNs.
 Caching in ICN takes as granted the presence
of a hosting server (caches are used to
improve delivery of popular items).
 In CBPS implementations (or in future P2P
ICN implementations) servers do not exist.
Vasilis Sourlas – PhD defense – 23/7/2013

Caching to preserve information over time
instead of making information available in
nearer space is missing.
40
Contributions
Vasilis Sourlas – PhD defense – 23/7/2013






Enhanced CBPS with a req/resp scheme (subscribers
can retrieve already published items).
Decomposed caching mechanism is a set of basic policies/strategies (proposed ICN oriented at each set).
Proposed two duplicate dropping mechanisms
(proactive & reactive).
Proposed a stochastic model that captures the
dynamics of the new ICN oriented policies.
Described a prototype implementation of the
proposed caching mechanisms (Planetlab).
Modified the proposed caching scheme to support
mobility of subscribers.
41
Policies
Vasilis Sourlas – PhD defense – 23/7/2013

Caching – selects a number of nodes and assigns them as
caching points.
 Selective caching (SEL)
 En-route caching (NRT)

Placement/Replacement - decides a position in the cache
where a new message will be cached and which message will
be discarded in case of an overflow.
 Least Recently Used policy (LRU)
 Least Frequently Used policy (LFU)
 Priority policy (PRT)

Request – dictates how requests (interests) are propagated
in the network.
 Subscription-based request policy (SUB)
 Flooding request policy (FLD)
42
Handling Multiple Responses
Vasilis Sourlas – PhD defense – 23/7/2013


Reactive mechanism, nodes check passing
responses whether the item is in its Cache. If
true, discards the response packet
(Responses follow backwards the same path
with Requests).
Proactive mechanism, responded
node/cache appends to the Request’s APID
(Aggregated Publication Ids) the pub-id of
the responded item. Recipients of the
request respond with cached items which
pub-id are not in the Request.
43
Stochastic Cache Modeling
Use Absorbing Markov Processes to
compute the Mean Absorption Time (AT).
of an item in the caches of the network.
 Present analytical results for a single node
network (~ multi-node scenario without
item copying).
 Reduce the state space with an approximation.
 Use the reduced space for the multi-node
scenario.
Vasilis Sourlas – PhD defense – 23/7/2013

44
Mobility Support
A mechanism that uses a portion of a
proxy’s buffer.
 Manages subscriptions and publications
on behalf of the Mobile Node (MN).
Vasilis Sourlas – PhD defense – 23/7/2013

When the MN is disconnected, stores items
matching MN’s interests.
During the switch-over phase (reconnection
phase) delivers stored items to the MN.
45
Evaluation
Vasilis Sourlas – PhD defense – 23/7/2013



Implemented the
framework in a Javabased overlay
framework/REDS and in
a discrete event
simulator using MATLAB.
Compared the analytical
model with discrete
event simulations.
Planetlab and simulation
experimentation of
various combinations of
opportunistic caching
schemes.
Metrics:
 Mean Absorption time,
AT – caching capability of
the network
 Minimum hop distance –
delay, perceived QoS
 Traffic Overhead –
replication and overhead
 Satisfaction – perceived
QoS
46
Vasilis Sourlas – PhD defense – 23/7/2013
Planetlab experimentation
47
Results
The newly proposed ICN oriented policies
outperform traditional ones.
 The two duplicate dropping mechanisms
minimizes the traffic overhead significantly
even when used with the flooding request
policy.
 The Markov model is accurate enough, but
looses accuracy when the number of nodes
increases.
 Prototype implementation results are inline
with discrete-event simulator outcome.
Vasilis Sourlas – PhD defense – 23/7/2013

48
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
49
Introduction
Vasilis Sourlas – PhD defense – 23/7/2013


Performance management and traffic
engineering approaches are required in ICN
to control routing, configure cache
replacement policies, etc.
Routing functionalities is completely missing
from the current ICN design.
Only flooding or OSPF-like shortest path
mechanisms have been proposed.
Recently hash-routing (similar to datacenters) has
been proposed to maximize cache hit within a
domain regardless of the traffic.
50
Contributions
Proposed a novel cache aware intra-domain routing
scheme that dynamically computes the paths
followed by each subscription/interest for each item
and from each node in the network.
 Presented a Dynamic Programming (DP) approach for
the computation of the “cheapest” transportation
paths based on the observed item request patterns,
in order to minimize the overall transportation cost
imposed by the user requests.
 Proposed an iterative algorithm for those scenarios
where the routing decisions interact with the caching
strategy.
 Presented a resource management system
architecture for the cache aware routing in ICN.
Vasilis Sourlas – PhD defense – 23/7/2013

51
Vasilis Sourlas – PhD defense – 23/7/2013
Problem Formulation
52
Vasilis Sourlas – PhD defense – 23/7/2013
Problem Formulation
53
Vasilis Sourlas – PhD defense – 23/7/2013
Problem Formulation
54
Motivation Example
Shortest path
Vasilis Sourlas – PhD defense – 23/7/2013
P = 0.2
Publisher of
Sid/Rid
E
P = 0.5
C
P = 0.5
Caching
probabilities of
item
A
P = 0.8
D
P = 0.8
B
Subscriber
requests
Sid/Rid
Our “shortest” path
1∙PA + 2(1-PA)PC+ 3(1-PA)(1-PC)PE+ 4 (1-PA)(1-PC)(1-PE)∙1 = 2.1 hops
In-network cache hit probability = PA + (1-PA)PC+ (1-PA)(1-PC)PE = 0.8
1∙PA + 2(1-PA)PB+ 3(1-PA)(1-PB)PD+ 4 (1-PA)(1-PB)(1-PD) ∙ PE + 5 (1-PA)(1-PB)(1-PD) (1PE ) ∙ 1 = 1.63 hops
In-network cache hit probability = PA + (1-PA)PB+ (1-PA)(1-PB)PD+ (1-PA)(1-PB)(1-PD) ∙
PE = 0.99
55
Vasilis Sourlas – PhD defense – 23/7/2013
Dynamic Programming approach
56
Iterative Algorithm
Cache hit ratio of each item is not independent from node to node.
 Iteratively execute the DP algorithm and
observe the network performance until
the cache hit ratios of each item
converge.
 At convergence the DP algorithm
computes the same paths as long as the
demand pattern remains stable.
Vasilis Sourlas – PhD defense – 23/7/2013

57
Evaluation
Compare the cache aware routing
scheme (CAWR) to the shortest path
routing scheme (SHPT).
Metrics:
Vasilis Sourlas – PhD defense – 23/7/2013

Total Transportation Cost, TTC
(resps*hops/sec)
Server Hit Ratio, SHR (reqs/sec)
58
Vasilis Sourlas – PhD defense – 23/7/2013
Synthetic Workload & Zoo Topologies
59
Results
CAWR outperforms SHPT 15%-20% improvement
regarding TTC and 35%-45% regarding SHR.
 The SHR improvement is almost twice as much as
the improvement in the TTC. Even if we cannot
alleviate the TTC, we can at least achieve
significantly better utilization of the network
resources and reduce the load at the hosting
servers and the need of replication devices.
 CAWR is robust enough and requires a
recalculations of the paths only after extreme
changes in the demand pattern.
Vasilis Sourlas – PhD defense – 23/7/2013

60
Vasilis Sourlas – PhD defense – 23/7/2013
Outline
1) Introduction
2) Replication Management Framework
3) Storage Planning and Off-line Replica
Assignment
4) Distributed Cache Management
5) Opportunistic Caching
6) Cache Aware Routing
7) Conclusions and Future Work
61
Replication Management

Presented a three phase
replication framework:
In-network opportunistic
caching

 Planning phase
Vasilis Sourlas – PhD defense – 23/7/2013
 A modified ICN oriented
greedy algorithm.
 Offline assignment
phase
 Two replica assignment
algorithms.
 On-line replacement
phase
 Four distributed on-line
cache management algs.
 A lower bound (overall
network traffic) for
regular network
topologies.




Proposed a new
opportunistic caching
mechanism.
Decompose it in a set of
basic policies.
Proposed a Markov
stochastic model.
Described a prototype
implementation and
evaluate it in Planetlab.
Modification of the
mechanism to enable
mobility of the
subscribers.
62
Cache aware routing

Vasilis Sourlas – PhD defense – 23/7/2013


Presented a new cache
aware intra-domain
routing scheme.
Proposed DP approach
for the computation of
the minimum
transportation cost
paths.
Proposed an iterative
algorithm when routing
interacts with caching
schemes.
Future Work
- Core work
Different objectives and
SLAs among the storage
provider and the content
providers.
 Combine opportunistic
caching with replication
nodes.
 Enhance routing scheme
with multiple servers and
traffic engineering schemes.

- ICN area
Security/anomaly detection.
 Seamless mobility.
 Energy Efficient usage of ICN
resources.
 Pricing schemes for the new
ICN paradigm.

63
Related Publications

Book chapter
[B.01] Vasilis Sourlas, Paris Flegkas, Dimitrios Katsaros and Leandros Tassiulas, “Content Replication and Delivery in
Information-Centric Networks,” to appear in Advanced Content Delivery and Streaming in the Cloud by Wiley Publishers,
USA.
Vasilis Sourlas – PhD defense – 23/7/2013

Journal publications
[J.04] Vasilis Sourlas, Paris Flegkas and Leandros Tassiulas, “A Novel Cache Aware Routing Scheme for Information-Centric
Networks,” submitted in Computer Networks Elsevier.
[J.03] Vasilis Sourlas, Lazaros Gkatzikis, Paris Flegkas and Leandros Tassiulas, “Distributed Cache Management in InformationCentric Networks,” to appear in IEEE Transaction on Network and Service Management (TNSM), 2013.
[J.02] Mohamed Diallo, Vasilis Sourlas, Paris Flegkas, Serge Fdida, and Leandros Tassiulas, “A Content-Based Publish/Subscribe
framework for Large-scale Content Delivery,” in Computer Networks Elsevier, Volume 57, Issue 4, pp. 924-943, March
2013.
[J.01] Vasilis Sourlas, Paris Flegkas, Georgios S. Paschos, Dimitrios Katsaros, and Leandros Tassiulas, “Storage Planning and
Replica Assignment in Content-Centric Publish/Subscribe Networks,” in S.I. on Internet-based Content Delivery, Computer
Networks Elsevier, Volume 55, Issue 18, pp. 4021-4032, December 2011.

Conference publications
[C.13] Vasilis Sourlas, Paris Flegkas and Leandros Tassiulas, “Cache-Aware Routing in Information- Centric Networks,” in
IFIP/IEEE International Symposium on Integrated Network Management (IM 2013), pp. 582-588, Ghent, Belgium, 2013.
[C.12] Vasilis Sourlas and Leandros Tassiulas, “Effective Cache Management and Performance Limits in ICN,” in International
Conference on Computing, Networking and Communications (ICNC 2013), pp. 955-960, San Diego, USA, 2013.
[C.11] Paris Flegkas, Vasilis Sourlas, George Parisis and Dirk Trossen, “Storage Replication in Information-Centric Networking,”
in International Conference on Computing, Networking and Communications (ICNC 2013), pp. 850-855, San Diego, USA,
2013.
[C.10] Vasilis Sourlas, Paris Flegkas, Lazaros Gkatzikis and Leandros Tassiulas, “Autonomic Cache Management in InformationCentric Networks,” in 13th IEEE/IFIP Network Operations and Management Symposium (NOMS 2012), pp. 121-129,
Hawaii, USA, April 2012.
64

Conference publications (cont’d)
[C.09] Dirk Trossen, Xenofon Vasilakos, Paris Flegkas, Vasilis Sourlas and George Parisis, “Mobility Work Re-Visited Not
Considered Harmful,” in IEEE WMCNT 2011, pp. 1-8, Budapest, Hungary, October 2011.
Vasilis Sourlas – PhD defense – 23/7/2013
[C.08] Vasilis Sourlas, Lazaros Gkatzikis and Leandros Tassiulas, “On-Line Storage Management with Distributed Decision
Making for Content-Centric Networks,” in 7th Conference on Next Generation Internet (NGI) 2011, pp. 1-8, Kaiseslautern,
Germany, June 2011.
[C.07] Mohamed Diallo, Serge Fdida, Vasilis Sourlas, Paris Flegkas and Leandros Tassiulas, “Leveraging caching for Internetscale content-based publish/subscribe networks,” in IEEE ICC 2011, pp. 1-5, Kyoto, Japan, June 2011.
[C.06] Vasilis Sourlas, Georgios S. Paschos, Petteri Mannersalo, Paris Flegkas and Leandros Tassiulas, “Modeling the dynamics
of caching in content-based publish/subscribe systems,” in 26th ACM Symposium On Applied Computing (SAC), Taiwan,
March 2011.
[C.05] Vasilis Sourlas, Paris Flegkas, Georgios S. Paschos, Dimitrios Katsaros and Leandros Tassiulas, “Storing and Replication in
Topic-Based Publish/Subscribe Networks,” in IEEE Globecom 2010 Next-Generation Networking and Internet
Symposium,Miami, USA, December 2010.
[C.04] Vasilis Sourlas, Georgios S. Paschos, Paris Flegkas and Leandros Tassiulas, “Mobility support through caching in contentbased publish/subscribe networks,” in 5th International Workshop on Content Delivery Networks (CDN 2010) in
conjuction with 10th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid 2010), pp. 715720, Melbourne, Australia, May 2010.
[C.03] Vasilis Sourlas, Georgios S. Paschos, Paris Flegkas and Leandros Tassiulas, “Caching in content based publish/subscribe
systems,” in IEEE Globecom 2009 Next-Generation Networking and Internet Symposium, pp. 1-6, Hawaii, USA, December
2009.
[C.02] Vasilis Sourlas, Paris Flegkas, Georgios S. Paschos and Leandros Tassiulas, “Distribute, Store and Retrieve Management
Policies on Wireless Ad-Hoc Networks using the Content Delivery Publish/Subscribe Paradigm,” in proc. of 3rd IEEE
Workshop on Autonomic Communications and Network Management - IM 2009 / ACNM 2009, pp 169-176, NY, USA, June
2009.
[C.01] Vasilis Sourlas, Paris Flegkas and Leandros Tassiulas, “Policy Distribution using the Publish-Subscribe Paradigm for
Managing MANETs,” in proc. of 11th IFIP/IEEE International Conference on Management of Multimedia and Mobile
Networks and Services (MMNS 2008) held as part of Manweek 2008,pp 14-19, Samos, Greece, August 2008.
65
Vasilis Sourlas – PhD defense – 23/7/2013
Thank you!!!
66