Self-organizing Wide-area Network Caches
Self-organizing Wide-area Network Caches
Samrat Bhattacharjee Ken Calvert Ellen Zegura
Networking and Telecommunications Group
College of Computing
Georgia Institute of Technology
Atlanta, Georgia, USA
http://www.cc.gatech.edu/projects/canes
Sponsors: DARPA, NSF
1
Self-organizing Wide-area Network Caches
Conclusions
Results
Self-organizing caching schemes
Caching approaches
Problem statement
Outline
2
Self-organizing Wide-area Network Caches
3
Problem Statement
Improve client access latencies by associating caches with routers
within the network
Network Cache
− caches reply
− serves request #1
server
request # 1
request # 0
Develop algorithms to co-ordinate network caches in order to
reduce latency
Assumptions
{ Request-response paradigm
{ Requests can be served by caches
Components
{ Location of caches
{ What to cache and where
{ Rules for forwarding requests
{ Protocols to communicate between caches
{ Cache consistency
Examples: Harvest, icp, Squid, Netcache
Wide-area Caching
Self-organizing Wide-area Network Caches
Develop algorithms to coordinate caches to reduce latency
4
Self-organizing Wide-area Network Caches
5
Active Networks and Network Caching
Active Networks provide a programmable network platform
an active network
active nodes
can execute
"user" algorithms
active routers could
execute new caching
algorithms
Use active networking to instantiate self-organizing cache
management algorithms within the network
Self-organizing Wide-area Network Caches
6
Caching Model and Topology Model
Clients, Servers, Popular Items
Transit-Stub Network Topologies
transit domain
T
T
T
T
T
T transit router
stub router
stub domain
hosts
Transit−Stub
Topology Model
Self-organizing Wide-area Network Caches
7
Approaches towards Caching
Caching by location
{ At transit nodes | Backbone routers
{ At border stub nodes | Access routers
server
T
T
T
T
T
T transit caches
stub connected
to transit caches
request
Self-organizing Wide-area Network Caches
8
Self-organizing Network Caches
Small caches without regard to location
Eective use of widely distributed caches is dicult
server
active caches,
run self−organizing
algorithms
request
Challenges: Where to cache, how to forward requests?
Self-organizing Wide-area Network Caches
9
Modulo Caching
Spread item over client-server path
Cache radius | modulo caching
server 1
2
1
3
2
3
1
2
3
modulo caching
radius = 3
1 request
Item is cached once on server-client path every radius hops
Self-organizing Wide-area Network Caches
10
Cache Lookaround
Observation: Item location size average item size
Use some item memory to store location of nearby items
Trade cache memory for location information
server
lookaround caching
1 level of lookaround
request
Request may be deected (even o the shortest path to server)
to caches within lookaround radius
Self-organizing Wide-area Network Caches
11
Simulation Methodology
Total cache size constant across methods
Example : total number of cache slots in network is 12.
S
C
T
C
S
S
S
self−organizing schemes
nominal cache size == cache size at each interface = 1
average cache size at each node ~ 1.72
S
C
T
transit−only caching
cache size at each interface = 6
average cache size at each node = 12
C
S
S
S
S
C
S
S
S
C
T
stub connected to transit caching
cache size at each interface = 3
average cache size at each node = 6
S stub node
T transit node
C stub node connected to transit node
caching node
Base Topology | 1500 nodes, average degree 3.71
{ Average transit-only cache 25 times larger than average
active cache, average SCT cache 4.25 times larger
20
25
30
35
Nominal Cache Size
40
Stub connected to Transit
Transit Only
Modulo Caching
Modulo w/ 2 hop Lookaround
Server Dist 0.8, Repeat Prob 0.1, Modulo Radius 3
15
No Cache RTL=13.23
10
45
10
9.5
9
8.5
8
7.5
5
10
15
40
Stub connected to Transit
Transit Only
Modulo Caching
Modulo w/ 2 hop Lookaround
20
25
30
35
Nominal Cache Size
No Cache RTL=13.24
Server Dist 0.8, Repeat Prob 0.5, Modulo Radius 3
Result: Varying Cache Size
Self-organizing Wide-area Network Caches
11.5
11
10.5
10
9.5
9
8.5
8
5
Round Trip Latency
45
SCT caches are not able to reduce latency as well
Self-organizing schemes perform better as cache size increases,
and as access correlations increase
Figure 1: Low Access Correlation Figure 2: High Access Correlation
Round Trip Latency
12
Self-organizing Wide-area Network Caches
13
Round Trip Latency (No Cache: 12.75-13.3
Result: Varying Server Distribution
Cache Size 16, Repeat Prob 0.5, Modulo Radius 3
9.5
9
8.5
8
7.5
Stub connected to Transit
Transit Only
Modulo Caching
Modulo w/ 2 hop Lookaround
7
6.5
0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Server Distribution (Fraction of Total Nodes)
Self-organizing schemes are robust against server distribution
16
Base
Cache Size 08, Server Dist. 0.8, Repeat Prob. 0.3
Higher Degree Less S-Domain Lower Degree More S-Domain
Topology
No Cache
No AN
15
Stub
connected
to Transit
Transit Only
Modulo Caching
14
Modulo w/ Lookaround
Modulo w/ 2 hop Lookaround
13
12
11
10
9
8
Figure 3: Dierent Topologies
13.5
13
12.5
12
11.5
11
10.5
10
9.5
9
2
6
8
Number of Preferred Servers
10
No Cache
No AN
Stub conn. to Transit
Transit Only
Modulo Caching
Modulo Caching w/ Lookaround
Modulo Caching w/ 2 Hop Lookaround
Cache Size 08, Sdist 0.8, Repeat Prob 0.3, Modulo Radius 3
4
12
Figure 4: Spatial Access Pattern
Round Trip Latency (Hops)
Result: Varying Topology and Access Patterns
Self-organizing Wide-area Network Caches
Round Trip Latency
14
Self-organizing Wide-area Network Caches
15
Analysis
Probabilistic analysis of cache performance
Expressions for latencies for various methods
14
Analysis, rtl=12
Analysis, rtl=14
Simulation, rtl=13.1
13
Round Trip Latency
12
11
10
9
8
7
6
5
4
0
100
200
300
400 500 600
Cache Size
700
800
900 1000
Comparison of analytic and simulation results | Lookaround Caches:
Base Topology, 50 location pointers per item, 2 levels of lookaround
Analysis of Lookaround Benet
Self-organizing Wide-area Network Caches
400
500
0.1
0.2
0.3
0.4
0.9
0.8
0.7
0.6
0.5
Local Obj. Cache Fract.
Partitioning of Cache between Local Objects and Location
Caches space is partitioned into data and location pointers
Question: What is the optimal amount to devote location
pointers?
Round Trip Latency
14
14.5
13
13.5
12
12.5
11.5
100
200
300
Cache Size
Valley in plot due to benets of lookaround caching
16
Increase activity of in-network processing:
{ Utilize per-application, per-user semantics
{ Increase per-item state in the network
Active Network implementation using ants
Self-organizing schemes reduce latency using far smaller
individual caches
Self-organizing schemes are robust against varying access
schemes, topologies, server distributions
Self-organizing algorithms are an active network application
that is independent of individual users
Conclusions and Future Work
Self-organizing Wide-area Network Caches
17