W. K. Chai, D. He, I. Psaras and G. Pavlou "Cache "Less for More" in Information-Centric Networks (Extended Version)", Elsevier Computer Communications, Special Issue on Information-Centric Networking, vol. 36, no. 7, pp. 758-770, 1 April 2013, (DOI) 10.1016/j.comcom.2013.01.007.
W. K. Chai, D. He, I. Psaras, G. Pavlou, "Cache "Less for More" in Information-centric Networks", Proceedings of the 11th IFIP Networking, Prague, Czech Republic, 21-25 May 2012. BEST PAPER AWARD
How to Remove Document Management Hurdles with X-Docs?
Cache "Less for More" in Information-centric Networks (CL4M)
1. Cache “Less for More” in Information-
Centric Networks
W. K. Chai, D. He, I. Psaras and G. Pavlou (presenter)
Department of Electronic & Electrical Engineering
University College London
London WC1E 6EA
United Kingdom
2. Background: Caching in Information-Centric
Networking (ICN)
• In Information-Centric Networking (ICN), contents are named.
• Content requests can be served from any node having the content
identified by the content name.
• ICN features ubiquitous in-network caching:
– potentially every router caching indiscriminately all content fragments
that traverse it
– if a matching request is received while a fragment is still in its cache
store, it will be forwarded to the requester from that element, avoiding
going to the hosting server.
• Specifically, NNC* proposes:
– A router caches every content chunk that traverses it with the
assumption that routers are equipped with (large) cache stores.
– A least recently used (LRU) cache eviction policy is used.
2*V. Jacobson, et al., “Networking Named Content,” Proc. ACM CoNEXT, 2009, pp. 1-12.
3. Argument & Objective
• The Argument: indiscriminate universal caching strategy is
unnecessarily costly and sub-optimal
– High content replacement frequency may result in content
being replaced before getting a hit.
• The Objective: to study alternative in-network caching
strategies for enhancing the overall content delivery
performance.
– We address the central question of whether caching only at
specific sub-set of node(s) en route the delivery path can
achieve better gain.
– If yes, which are these nodes to cache and how can we
identify them?
3
4. “Cache All” vs “Random” Strategies
• Ubiquitous caching (NNC+LRU): caches every content in every
node along the delivery path and uses LRU eviction strategy
• Random caching strategy (Rdm+LRU): caches at one randomly
selected node along the delivery path and also uses LRU eviction
4
0 50 100 150 200
0.49
0.5
0.51
0.52
0.53
0.54
0.55
0.56
0.57
Time, t(s)
InstantaneousHopReductionRatio,β(t)
NNC+LRU
Rdm+LRU
0 50 100 150 200
0.42
0.44
0.46
0.48
0.5
0.52
0.54
0.56
0.58
Time, t(s)
InstantaneousServerHitReductionRatio,γ(t)
NNC+LRU
Rdm+LRU
Simple random caching outperforming ubiquitous caching in the number of hops to hit the
content (left) and reduced server hits (right).
5. Approach & Rationale
• The approach: to cache content only in selected node(s) for each
content request
• The rationale: some nodes have higher probability of getting a
cache hit than others.
– By strategically caching the content at “better” node(s), we can
decrease the cache eviction rate and increase the cache hit.
– Illustrative example:
5
– Events: t=0, client A requests a content from S1 and t=1, client B
requests a content from S1
– If NNC, ubiquitous caching at v1, v2, v3 and v4 but only v3 necessary
for the request from client B
– v3 being the “important” node in this case.
v2
v6
v5v1 v3s1
Client C
Client B
v4 Client A
6. (Ego Network) Betweenness Centrality
6
• The betweenness centrality (Betw+LRU): measures the number
of times a specific node lies on the content delivery path between
all pairs of nodes in a network topology.
– The basic idea is that if a node lies along a high number of content
delivery paths, then it is more likely to get a cache hit.
– Caching only at those “important” nodes reduces the cache
replacement rate while still caching content where a cache hit is most
probable to happen.
• The Ego Network Betweenness centrality (EgoBetw-LRU):
– The ego network consists of a node together with all its immediate
neighbours and all the links among those nodes.
– The idea is for each node, v to compute its CB(v) based on its ego
network rather than the entire network topology.
7. Operation
• Betw+LRU and EgoBetw+LRU caching:
– Operates at per request level.
– The CB value is pre-computed offline and configured to every
router by the network management system.
– The content request message records the highest centrality
value among all the intermediate nodes.
– This value is copied onto the content messages.
– On the way to the requesting user, each router matches its own
CB against the attached one
• The content is cached only if the two values match.
• If more nodes have the same highest centrality value, all of them
will cache the content.
7
8. Features
• Betw+LRU and EgoBetw+LRU:
– cache content at nodes which have high probability of
getting cache hits
– both approaches always spread content towards users
• If a content is popular at a specific region of the Internet, then this
content will eventually be cached in that region
– is lightweight - each node independently makes its caching
decision solely based on its own CB
• neither requiring information exchange with other nodes nor
inference of server location or of traffic patterns.
8
9. Metrics
9
• In-network caching aims to:
1. Reduce the content delivery latency whereby a cached content near
the client can be fetched faster than from the server
2. Reduce traffic and congestion since content traverses fewer links
when there is a cache hit
3. Alleviate server load as every cache hit means serving one less
request.
– where Hr(t) = the hop count from client to server requesting fr from
time t-1 to t and hr(t) = the hop count from the content client to the
first node where a cache hit occurs for fr from t-1 to t.
– where Wr(t) = the number of request for fr from t-1 to t and wr(t) = the
number of server hits for fr from t-1 to t.
10. Performance Evaluation on k-ary Trees
• Instantaneous performance for a 5-level binary tree (k=2,
D=4)
0 50 100 150 200
0.38
0.4
0.42
0.44
0.46
0.48
0.5
Time, t(s)
InstantaneousHopReductionRatio,β(t)
NNC+LRU Betw+LRU EgoBetw+LRU Rdm+LRU
0 20 40 60 80 100 120 140 160 180 200
0.18
0.2
0.22
0.24
0.26
0.28
0.3
0.32
0.34
0.36
Time, t(s)
InstantaneousServerHitReductionRatio,γ(t)
NNC+LRU Betw+LRU EgoBetw+LRU Rdm+LRU
Instantaneous behavior of the caching schemes for a binary tree; (left) β, (right) γ.
10
11. Performance Evaluation on k-ary Trees
• k-ary trees with varying k and D.
4 5 6
0.35
0.4
0.45
0.5
0.55
Binary tree depth, D
HopReductionRatio,β
NNC+LRU
Betw+LRU
EgoBetw+LRU
Rdm+LRU
2 3 4 5
0.32
0.34
0.36
0.38
0.4
0.42
0.44
0.46
0.48
0.5
Spread factor, k
HopReductionRatio,β
NNC+LRU
Betw+LRU
EgoBetw+LRU
Rdm+LRU
Betw+LRU consistently outperforms the rest over different D (left) and k (right).11
12. Performance Evaluation on Scale-free Graphs
• Scale-free graphs
– Barabasi-Albert (BA) graphs preferential attachment
– N = 100 and mean valence = 2
0 50 100 150 200
0.36
0.38
0.4
0.42
0.44
0.46
0.48
0.5
Time, t(s)
InstantaneousHopReductionRatio,β(t)
NNC+LRU
Betw+LRU
EgoBetw+LRU
Rdm+LRU
0 50 100 150 200
0.16
0.18
0.2
0.22
0.24
0.26
0.28
0.3
0.32
0.34
0.36
Time, t(s)
InstantaneousServerHitReductionRatio,γ(t)
NNC+LRU
Betw+LRU
EgoBetw+LRU
Rdm+LRU
Instantaneous behavior of the caching schemes in a B-A graph; (left) β, (right) γ.
12
13. Performance Evaluation on Scale-free Graphs
• 10 generations of B-A graphs with N = 100 and
mean valence = 2.
1 2 3 4 5 6 7 8 9 10
0.38
0.4
0.42
0.44
0.46
0.48
0.5
0.52
Graph ID
HopReductionRatio,β
NNC+LRU
Betw+LRU
EgoBetw+LRU
Rdm+LRU
0 20 40 60 80 100
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Betweenness
0 20 40 60 80 100
0
5
10
15
20
25
30
35
EgoNetworkBetweenness
Node ID
Betweenness
Ego Network Betweenness
Caching performance with different 100-node B-A graphs (left) and a sample ego network
betweenness and betweenness values of the nodes in a B-A graph (right).
13
14. Evaluation on Real Internet Topology
• Real-world Internet topology
– Large domain-level topology extracted from the CAIDA dataset
– Root at tier-1 ISP (AS7018)
– 6,804 domains and 10,205 links
Instantaneous behavior of the caching schemes in a large-scale real Internet topology; (left) β,
(right) γ. 14
0 20 40 60 80 100
0.5
0.55
0.6
0.65
0.7
0.75
Time, t (s)
InstantaneousHopReductionRatio,β(t)
NNC+LRU
Betw+LRU
EgoBetw+LRU
Random+LRU
0 20 40 60 80 100
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.55
Time, t (s)
InstantaneousServerHitReductionRatio,γ(t)
NNC+LRU
Betw+LRU
EgoBetw+LRU
Rdm+LRU
15. Summary and Conclusions
• We argue against the necessity of a ubiquitous caching
strategy in ICN and investigate the possibility of caching less to
achieve higher performance gain.
• We demonstrate that even a simple random caching strategy
(Rdm+LRU) can outperform (though inconsistently) the current
pervasive caching paradigm under the conditions
– that the network topology has low number of distinct delivery paths
– high average delivery path length.
• We propose a caching strategy based on the concept of
betweenness centrality (Betw+LRU)
– content is only cached at the nodes having the highest probability of
getting a cache hit along the content delivery path.
– also proposed an approximation of it (EgoBetw+LRU) for scalable and
distributed realization in dynamic network environments.
15
16. Summary and Conclusions (cont’d)
• We compare the performance of our proposals against the
ubiquitous caching of the NNC proposal (NNC+LRU).
• Observations:
– Betw+LRU consistently achieves the best hop and server
reduction ratios across topologies having different structural
properties without being restricted by the operating conditions
required by Rdm+LRU.
– EgoBetw+LRU approximates closely Betw+LRU in non-regular
topologies (e.g., B-A graphs, real Internet topology)
• a practical candidate for the deployment of this approach.
– Besides synthetic topologies (i.e., k-ary trees and B-A graphs), the
observations are further verified with a large-scale real Internet
topology.
• We conclude that indeed caching less can achieve more and
our proposed (Ego)Betw+LRU is a candidate for realizing this
promise. 16
17. • Work done in the context of the EU FP7 project COMET which we lead
technically - http://www.comet-project.org/
• W. K. Chai, I. Psaras, G. Pavlou, et. al., “CURLING: Content-
ubiquitous Resolution and Delivery Infrastructure for Next-generation
Services,” IEEE Commun. Mag., vol. 49, no. 3, pp. 112-120, 2011.
• I. Psaras, R. Clegg, R. Landa, W. K. Chai and G. Pavlou, “Modelling
and Evaluation of CCN-caching Trees,” Proc. of IFIP Networking,
Valencia, Spain, May 2011.
• I. Psaras, W. K. Chai and G. Pavlou, “To Cache or Not To Cache:
Probabilistic In-Network Caching for Information-Centric Networks,” to
appear in the ACM SIGCOMM ICN Workshop, August 2012.
17
Related Publications