Upload
anshuman-kalla
View
20
Download
1
Embed Size (px)
Citation preview
Selective Placement of Caches for Hash-Based Off-Path Caching in ICN*
Anshuman Kalla
1
*Indian Journal of Science and Technology, Vol 9 (37), October 2016
DOI: 10.17485/ijst/2016/v9i37/100323
For reading mode, Adobe Reader – use Ctrl + H and Foxit Reader – use F11
In-Network Caching in ICN
• Research based resurgence in the field of networking seems to have culminated in form of ICN– since it has the capability to conciliate in a unified manner all the
existing & anticipated issues
2/24/2017 2Anshuman Kalla
In-Network Caching in ICN
• Research based resurgence in the field of networking seems to have culminated in form of ICN– since it has the capability to conciliate in a unified manner all the
existing & anticipated issues
• In-network caching has evidently emerged as an indispensable core functionality of ICN– expected to play a vital role in ameliorating performance
2/24/2017 3Anshuman Kalla
In-Network Caching in ICN
• Research based resurgence in the field of networking seems to have culminated in form of ICN– since it has the capability to conciliate in a unified manner all the
existing & anticipated issues
• In-network caching has evidently emerged as an indispensable core functionality of ICN– expected to play a vital role in ameliorating performance
• Advantages of in-network caching, related issues, factors affecting, performance metrics etc. – Have been discussed in detail
2/24/2017 4Anshuman Kalla
In-Network Caching in ICN
• Research based resurgence in the field of networking seems to have culminated in form of ICN– since it has the capability to conciliate in a unified manner all the
existing & anticipated issues
• In-network caching has evidently emerged as an indispensable core functionality of ICN– expected to play a vital role in ameliorating performance
• Advantages of in-network caching, related issues, factors affecting, performance metrics etc. – Have been discussed in detail *
* A. Kalla, SK Sharma, “A Constructive Review of In-Network Caching: A Core Functionality of ICN,” Proc. IEEE International Conference (ICCCA), pp 567 – 574, 2016
2/24/2017 5Anshuman Kalla
Background & Motivation
• Off-Path caching has preliminarily shown better results – In terms of hit ratio, retrieval delay, cache diversity, server load etc.
– as compared on-path and edge caching [see references in the paper]
2/24/2017 6Anshuman Kalla
Background & Motivation
• Off-Path caching has preliminarily shown better results – In terms of hit ratio, retrieval delay, cache diversity, server load etc.
– as compared on-path and edge caching [see references in the paper]
• One of the practical & popular ways of implementing pervasive off-path caching is– Hash-based off-path caching that uses hash function
2/24/2017 7Anshuman Kalla
Background & Motivation
• Off-Path caching has preliminarily shown better results – In terms of hit ratio, retrieval delay, cache diversity, server load etc.
– as compared on-path and edge caching [see references in the paper]
• One of the practical & popular ways of implementing pervasive* off-path caching is– Hash-based off-path caching that uses hash function
* Pervasive means all the nodes are enabled with in-network caching capability
2/24/2017 8Anshuman Kalla
Background & Motivation
• Off-Path caching has preliminarily shown better results – In terms of hit ratio, retrieval delay, cache diversity, server load etc.
– as compared on-path and edge caching [see references in the paper]
• One of the practical & popular ways of implementing pervasive off-path caching is– Hash-based off-path caching that uses hash function
• Nevertheless, one downside of off-path caching is– Increase in intra-domain link load to achieve required gain
2/24/2017 9Anshuman Kalla
Background & Motivation
• Off-Path caching has preliminarily shown better results – In terms of hit ratio, retrieval delay, cache diversity, server load etc.
– as compared on-path and edge caching [see references in the paper]
• One of the practical & popular ways of implementing pervasive off-path caching is– Hash-based off-path caching that uses hash function
• Nevertheless, one downside of off-path caching is– Increase in intra-domain link load to achieve required gain
• Moreover, internal link stress and retrieval delay depends on – Network topology
2/24/2017 10Anshuman Kalla
Aim
• With this background the question raised is:
2/24/2017 11Anshuman Kalla
Does there exists a selective cache allocation map which could be conducive in achieving
• Smaller ‘average retrieval delay’ and simultaneous
• Lower ‘maximum intra-domain link-stress’
while keeping the other relevant metrics unaltered as compared to pervasive cache allocation?
Aim
• With this background the question raised is:
2/24/2017 12Anshuman Kalla
Does there exists a selective cache allocation map* which could be conducive in achieving
• Smaller ‘average retrieval delay’ and simultaneous
• Lower ‘maximum intra-domain link-stress’
while keeping the other relevant metrics unaltered as compared to pervasive cache allocation?
* Cache allocation or placement implies configuring the nodes with content storage capacity
Aim
• With this background the question raised is:
2/24/2017 13Anshuman Kalla
Does there exists a selective cache allocation map* which could be conducive in achieving
• Smaller ‘average retrieval delay’ and simultaneous
• Lower ‘maximum intra-domain link-stress’
while keeping the other relevant metrics unaltered as compared to pervasive cache allocation?
If so, what could be a simple way to find such a selective placement of caches* for hash-based off-path caching?
* Cache allocation or placement implies configuring the nodes with content storage capacity
Aim
• To ensure profitable, economic & selective placement of caches such that the placement– Tends to reduce simultaneously average retrieval delay &
maximum internal link-stress
2/24/2017 14Anshuman Kalla
Aim
• To ensure profitable, economic & selective placement of caches such that the placement– Tends to reduce simultaneously average retrieval delay &
maximum internal link-stress
• We propose simple yet elegant heuristic algorithm to find selective cache allocation map – by strategically identifying & excluding bad node-positions
2/24/2017 15Anshuman Kalla
Aim
• To ensure profitable, economic & selective placement of caches such that the placement– Tends to reduce simultaneously average retrieval delay &
maximum internal link-stress
• We propose simple yet elegant heuristic algorithm to find selective cache allocation map – by strategically identifying & excluding bad node-positions
• Rationale:– There are few nodes in network which if selected for caching
causes adverse effect
2/24/2017 16Anshuman Kalla
Aim
• To ensure profitable, economic & selective placement of caches such that the placement– Tends to reduce simultaneously average retrieval delay &
maximum internal link-stress
• We propose simple yet elegant heuristic algorithm to find selective cache allocation map – by strategically identifying & excluding bad node-positions
• Rationale:– There are few nodes in network which if selected for caching
causes adverse effect
• To the best of our knowledge, the work is first attempt – to explore selective placement of caches for hash-based off-path caching
2/24/2017 17Anshuman Kalla
Cost Computation of Nodes
• Cost for every node in the network is calculated
2/24/2017 18Anshuman Kalla
Cost Computation of Nodes
• Cost for every node in the network is calculated
2/24/2017 19Anshuman Kalla
• Since uniform rate of request arrival at ingress nodes is considered thus
Cost Computation of Nodes
• Cost for every node in the network is calculated
2/24/2017 20Anshuman Kalla
• Since uniform rate of request arrival at ingress nodes is considered thus
H = average hit ratio for the case of pervasive hash-based off-path caching
dxy = delay encountered to reach destination y from source x using the shortest path
I = set of ingress nodes, V = set of nodes in network topology
e ∈ E denotes egress node that falls over the shortest path & is connected to server
Cost Computation of Nodes
• Cost for every node in the network is calculated
2/24/2017 21Anshuman Kalla
• Since uniform rate of request arrival at ingress nodes is considered thus
H = average hit ratio for the case of pervasive hash-based off-path caching
dxy = delay encountered to reach destination y from source x using the shortest path
I = set of ingress nodes, V = set of nodes in network topology
e ∈ E denotes egress node that falls over the shortest path & is connected to server
• Using these values, a ordered list of nodes is prepared In the descending order of cost Cr i.e. is from most costly to least costly
Performance Metrics
• Average Retrieval Delay
– Delay encountered on an average to fetch requested content irrespective of source-of-request & point-of-retrieval
– Smaller the delay better is the user’s QoE
2/24/2017 22Anshuman Kalla
Performance Metrics
• Average Retrieval Delay
– Delay encountered on an average to fetch requested content irrespective of source-of-request & point-of-retrieval
– Smaller the delay better is the user’s QoE
• Link Stress
– For an intra-domain link, the metric indicates the volume of the traffic passed through it
– The value of link stress for the most heavily loaded internal link is termed as maximum internal link stress
2/24/2017 23Anshuman Kalla
Performance Metrics
• Hit Ratio
– Requests being satisfied by contents residing at in-network caches to the total requests received
– Higher is the hit ratio, lower is the traffic delegated over the expensive external links and lesser is the load on origin server
2/24/2017 24Anshuman Kalla
Performance Metrics
• Hit Ratio
– Requests being satisfied by contents residing at in-network caches to the total requests received
– Higher is the hit ratio, lower is the traffic delegated over the expensive external links and lesser is the load on origin server
• Unique Contents Cached
– Total number of distinct contents being cached by all the in-network caches collectively
– More is the unique contents cached better is the cache diversity
2/24/2017 25Anshuman Kalla
Environment Setup
• Exodux US, a PoP level topology from Rocketfuel[Reference # 22 in the paper] is used
2/24/2017 26Anshuman Kalla
Environment Setup
• Exodux US, a PoP level topology from Rocketfuel[Reference # 22 in the paper] is used
• About 8% of nodes were randomly appointed as egress nodes (connected to origin server)
2/24/2017 27Anshuman Kalla
Environment Setup
• Exodux US, a PoP level topology from Rocketfuel[Reference # 22 in the paper] is used
• About 8% of nodes were randomly appointed as egress nodes (connected to origin server)
• Approx. 55% of nodes were randomly selected as ingress nodes (connected to clients i.e. users)
2/24/2017 28Anshuman Kalla
Environment Setup
• Exodux US, a PoP level topology from Rocketfuel[Reference # 22 in the paper] is used
• About 8% of nodes were randomly appointed as egress nodes (connected to origin server)
• Approx. 55% of nodes were randomly selected as ingress nodes (connected to clients i.e. users)
• Aggregate network cache size is chosen to be 10% of content population
2/24/2017 29Anshuman Kalla
Environment Setup
• Exodux US, a PoP level topology from Rocketfuel[Reference # 22 in the paper] is used
• About 8% of nodes were randomly appointed as egress nodes (connected to origin server)
• Approx. 55% of nodes were randomly selected as ingress nodes (connected to clients i.e. users)
• Aggregate network cache size is chosen to be 10% of content population
• Rest of the parameters used are as per table, next
2/24/2017 30Anshuman Kalla
Parameters Used For Evaluation
2/24/2017 31Anshuman Kalla
Parameters Values
Total number of nodes 79
Number of server one
Number of egress nodes (connected to server) 6
Number of ingress nodes (connected to users) 44
Content population 79000
Aggregate network cache size 7900
Cache size per node 100
Popularity distribution of content requests Zipf (α = 0.8)
Delay to reach origin server from egress nodes 100 ms
Cache size per node Uniform
Rate of request arrival at ingress nodes Uniform
Content size Homogeneous
Cache replacement policy LRU
Number of iterations carried out for the selected topology 39
Number of runs per iteration of algorithm 1 20
Number of requests simulated per run (i.e. after launch of hash-based off-path
caching)500,000
Routing Shortest Path
Links between users and ingress nodes Bi-directional
Network regime Congestions Free
Points To Note
• First
– Evaluation takes into account only link propagation delay
– Factors like queuing delay, packet loss, cross traffic etc. have not been incorporated for preliminary evaluation
2/24/2017 32Anshuman Kalla
Points To Note
• First
– Evaluation takes into account only link propagation delay
– Factors like queuing delay, packet loss, cross traffic etc. have not been incorporated for preliminary evaluation
• Second
– Iteration denotes a repetitive process of debarring increasing number of nodes from caching, whereas
– Run indicates an act of deploying hash-based off-path caching with a selective placement of caches
– As per last table, 20 runs per iteration have been carried out during the simulations.
2/24/2017 33Anshuman Kalla
Points To Note
• Third
– Based on the hash function being used the mapping of contents to a designated off-path caches could be entirely different
– Thus for better evaluation content-to-cache mapping is being changed for every run
– This explains the reason behind multiple runs per iteration
2/24/2017 34Anshuman Kalla
Results and Discussion
2/24/2017 35Anshuman Kalla
Number of nodes debarred from caching
Ave
rage
Ret
riev
al D
ela
y
Results and Discussion
2/24/2017 36Anshuman Kalla
Number of nodes debarred from caching
Ave
rage
Ret
riev
al D
ela
y
Number of nodes debarred from caching
Max
imu
m L
ink
Stre
ss
Results and Discussion
2/24/2017 37Anshuman Kalla
Number of nodes debarred from caching
Ave
rage
Ret
riev
al D
ela
y
Number of nodes debarred from caching
Max
imu
m L
ink
Stre
ss77410
Selective Placement
Results and Discussion
2/24/2017 38Anshuman Kalla
Number of nodes debarred from caching
Ave
rage
Ret
riev
al D
ela
y
Number of nodes debarred from caching
Max
imu
m L
ink
Stre
ss77410
Selective Placement
Results and Discussion
2/24/2017 39Anshuman Kalla
Number of nodes debarred from caching
Ave
rage
Ret
riev
al D
ela
y
Number of nodes debarred from caching
Max
imu
m L
ink
Stre
ss
78.9465
77410
Selective Placement
Results and Discussion
2/24/2017 40Anshuman Kalla
Number of nodes debarred from caching
Ave
rage
Ret
riev
al D
ela
y
Number of nodes debarred from caching
Max
imu
m L
ink
Stre
ss
84.5661
81784
78.9465
77410
Selective PlacementPervasive Caching
Results and Discussion
2/24/2017 41Anshuman Kalla
Number of nodes debarred from caching
Ave
rage
Ret
riev
al D
ela
y
Number of nodes debarred from caching
Max
imu
m L
ink
Stre
ss
84.5661
81784
78.9465
77410
Selective PlacementPervasive Caching
Gain = 5.6196 i.e. 6.65%
Gain = 4374i.e. 5.35%
Results and Discussion
2/24/2017 42Anshuman Kalla
Content Descending in Popularity
Co
nte
nt
Wis
e A
vera
ge R
etr
ieva
l De
lay
Content Descending in Popularity
Co
nte
nt
Wis
e H
it R
atio
Results and Discussion
2/24/2017 43Anshuman Kalla
Placement of
CachesHit Ratio
Average
Retrieval Delay
Cache
Diversity
Maximum Link
Stress
Pervasive 0.4635 84.5661 7900 81784
Selective 0.4638 78.9465 7900 77410
Results and Discussion
2/24/2017 44Anshuman Kalla
Placement of
CachesHit Ratio
Average
Retrieval Delay
Cache
Diversity
Maximum Link
Stress
Pervasive 0.4635 84.5661 7900 81784
Selective 0.4638 78.9465 7900 77410
• The proposed algorithm is able to achieve
– 6.65% and 5.35% reduction in average retrieval delay and maximum internal link stress respectively
Results and Discussion
2/24/2017 45Anshuman Kalla
Placement of
CachesHit Ratio
Average
Retrieval Delay
Cache
Diversity
Maximum Link
Stress
Pervasive 0.4635 84.5661 7900 81784
Selective 0.4638 78.9465 7900 77410
• The proposed algorithm is able to achieve
– 6.65% and 5.35% reduction in average retrieval delay and maximum internal link stress respectively
• Quantitatively the gain might not be high but qualitatively the gain comes at no trade-off
– That is without sacrificing any other relevant performance metrics
Results and Discussion
2/24/2017 46Anshuman Kalla
Placement of
CachesHit Ratio
Average
Retrieval Delay
Cache
Diversity
Maximum Link
Stress
Pervasive 0.4635 84.5661 7900 81784
Selective 0.4638 78.9465 7900 77410
• The proposed algorithm is able to achieve
– 6.65% and 5.35% reduction in average retrieval delay and maximum internal link stress respectively
• Quantitatively the gain might not be high but qualitatively the gain comes at no trade-off
– That is without sacrificing any other relevant performance metrics
• The nodes debarred are not to be upgraded with ICN functionalities which implies lower cost
Thank You!
2/24/2017 Anshuman Kalla 47