CDN server & path selection - Usc - University of Southern California
Download
Report
Transcript CDN server & path selection - Usc - University of Southern California
Tradeoffs in CDN Designs for
Throughput Oriented Traffic
Minlan Yu
University of Southern California
Joint work with Wenjie Jiang, Haoyuan Li, and Ion Stoica
1
Throughput-Oriented Traffic
• Throughput-oriented traffic is growing in Internet
– Cisco report predicts that 90% of the consumer traffic
will be video by 2013 (E.g., NetFlix, Youtube)
– Software, game, movie downloads
– Most are delivered by content distribution networks
Revisit CDN design choices for throughputoriented traffic
2
Where is the throughput bottleneck?
Client:
Network:
Server:
Computer/access Congestions at peering Not enough resource
link too slow
and upstream links
(CPU, power, bw)
3
Understanding Throughput Bottleneck
• Network bottlenecks are common
Buffering ratio
– NetFlix sees reduced video rates due to low ISP capacity
– Akamai reported bottlenecks at peering links
4
3.5
3
2.5
2
1.5
1
0.5
Degraded video performance
caused by network congestion
0
2
4
6
8
10 12 14 16 18
Concurrent views (K)
4
Nature of Bottleneck is Changing
• More throughput-oriented applications
– Video traffic lasts longer and has higher volume
• More elephants step on each other in the future
– Decreases the benefits of statistical multiplexing
– Introduces more challenges in bandwidth provisioning
5
Improving Network Throughput
• ISP-CDNs: multiple paths and better path selections
– ISPs move up in the revenue chain to deliver content
• ISP-CDNs such as AT&T and Verizon
– Control both servers and the network
– Better traffic engineering for CDN traffic
• Existing CDNs: Deploy servers at more locations and
setting up more peering points
Peering
Question
points
1: What’s the throughput benefit of
more paths over more peering points?
……
6
Improving CDN Throughput
• Highly distributed approach (e.g., Akamai)
– Many server locations, more high-throughput paths
– Higher management, replication, bandwidth cost
• More centralized approach (e.g., Limelight)
– A few large data centers with more peering points
– Lower cost due to economy of scale
More centralized
Highly distributed
Question 2: How to compare more centralized vs.
more distributed CDNs on throughput and cost?
……
Modeling CDN Design Choices
• CDNs: Increase peering points at the edge
• ISPs: Improve path selection at the core
8
Increase Peering Points
• Modeling peering points (PPs)
– Increase #PPs to study throughput effect
– Pick PP locations from synthetic and real topologies
• Peering point selection
– Maximize aggregate throughput
– By assigning client locations to PPs
… and splitting traffic to different PPs
9
Improve Path Selection
• Today: No cooperation (1path)
– ISPs: Shortest path routing (e.g., OSPF)
– CDNs: Select peering points to maximize throughput
• Better contracts between ISPs and CDNs (n paths)
– ISPs: Expose multiple shortest paths to CDNs (e.g.,MPLS)
– CDNs: Select peering points and paths
10
Improving Path Selection
• ISP-CDNs: Optimal throughput (mcf)
– Joint traffic engineering and server selection
– Reduced to multi-commodity flow problem
• Optimization formulation
– Objectives: Max total throughput
– Subject to: Client demands & Link capacity constraints
– Variables: Peering point selection, traffic splitting on
each paths (Flow_{path, pp, client})
11
An Example
Min-cut size
– improving path selection only approximates the min-cut size
– increasing #peering points essentially increases min-cut size
Capacity =2
Capacity =1
Capacity =2
• With PP2 and PP3, the maximum throughput of multiple paths is 4
(min-cut size 4)
12
• Increase to 4 PPs, the min-cut size now is 8
Question 1:
What’s the benefit of path selection
over peering point selection?
13
Quantify the Benefits under Various Scenarios
• Network
– Topologies: power-law, random, hierarchy, different link
density, router-level ISP topo, AS-level Internet topo
– Link capacity distribution: uniform, exp., pareto, higher
inter-AS bandwidth
• CDN peering points
– Map Akamai and Limelight server IP addresses to ASes
(collected from PlanetLab measurement at Nov. 2010)
– Randomly pick peering points for synthetic topologies
• Client demands
– Session-level traces from Conviva collected between
14
Dec. 2011 and April. 2012
Multipath is better than Multiple Locations
– Power law graph (500 nodes, 997 links)
– Uniform link capacity distribution
– 200 clients at random locations
Multiple paths have little improvement
over increasing peering points
15
Effect of Network Topology
– Increasing peering points are better than multipath in
most topologies
– Except star-like topology with uniform link capacity
58
• The throughput from
1path to mcf increases by
110% - 584%
• The throughput from 10
PPs to 20 PPs increases by
337%
243
146
13
263
10
290229170136 28
200 38
222 158169 3
27 267 8
90 245 131
40
64
261133 175
201 67 193
71 91 60
223165
47
109230
19
62
65
21
227
283211 121
140
161
234
48117312122
79
32 51258
25352 318
196
159
188
83
300
23
302
25953
269
231 187
35149
153
255
110 41151
74 309 268
297
303
100
242 270 308120
20
235
147 220
72
185
11
138
244
66
59
319 106
293 160 167 197
42155
128
73
111108
226
208 81
317
198 246 95 93
191
189209 80150
69114
232 49 249299
135
156
145
37
115 118 190
296
311 284 216 31
301 294 154
82
164
184 102
162248 26
304
103
116
313
105 152 87
183 86
94 291
325
171
277
29 260173 288
17
6
104 264
233
326 239
139
292236143
26514
78 305163 289
112
101
144
98
213
207
217166 180
322 275
107 181
287
323
57
212 75
321
113
30 15
276 179 295
168
2572 55125
36
278 315
281 141 202
205
274
88
282
178
148
195
182 204 206
240 22
186
316 16
279237
224
33
320
157
174
314
247
298
34
215 24
256
89 127210 12
97
280
214
76 1306 68 61
134
192
126
1307
285 25 5119
45
272
177
70
43324 85
56
286
199
63
238
241
266
176
203
262225
132
251 46
99 123228
4 250
124
54
252
219 172
254
221 96 92
77
307142
84 310
9 137194 129 50
44
18
271
39
218
273
16
Path selection not useful under Flash Crowd
Thpt (Path + peering
point selection)
Thpt (Peering point
selection)
Relative scaling ratio
– Conviva traces during normal and flash crowd periods
– Path selection has little benefits under normal traffic
– Path selection is worse than only peering point
selection
1.4
1.2
1
0.8
0.6
0.4
0.2
0
flash crowd
normal
5min 10min30min 1hour 2hour
Path selection interval
17
More peering points always better than more paths
with long-tail Distribution of Contents
– Long-tail content distribution trace from Conviva
– With fewer replications, the throughput benefit of
multipath increases
Normalized Throughput
• Without replication the content delivery is closer to the singlesource traffic
8
7
6
5
4
3
2
1
0
100PP,1path
10PP,mcf
10PP,1path
0.1
1
2
10
20
Duplication Threshold (%)
18
Takeaway 1:
CDNs only need to control the edge of the Internet to
improve the throughput.
ISP-CDNs don’t get significant benefits from controlling
the network over CDNs
19
Question 2:
How to compare throughput and cost
between
more centralized vs more dist. CDNs?
20
Throughput Comparison of CDNs
– Assume a fixed aggregate peering bandwidth per CDN
– A more distributed CDN achieves better throughput than
more centralized one
Throughput (K)
200
peering bw
2-3
150
4-5
6-10
100
>10
Distributed
Centralized
50
0
0
50 100 150 200 250 300 350 400
#peering links
21
CDN Operation Cost
• Management cost
– At each location: electricity, cooling, equip maintenance,
and human resources
• Content replication cost
– Storage cost to replicate popular content
– Bandwidth cost to redirect traffic for rare content
• Bandwidth cost
– CDNs often pay ISPs for the bandwidth they use at the
peering points based on mutually-agreed billing model
22
Different Cost Functions
• Cost as a function of bandwidth at a location
Unit price per bandwidth
– Different functions: polynomial, linear, log, exp
– Model how fast the unit cost drops with throughput
– In practice: a linear combination of different functions
1
Polynomial
Linear
Log
Exponential
0.8
0.6
0.4
0.2
0
20
40
60
Throughput
80
100
23
Polynomial Cost
• Dist. CDN is more expensive than Centralized one
Unit price per bandwidth
– Limelight has larger throughput at each location and
thus better scalability gains
– Same observation holds across various operational cost
functions and their combinations
Distributed
0.5
0.45
0.4
2-3
4-5
6-10
>10
0.35
0.3
Centralized
0.25
0
20
40
60
80
100 120 140
Throughput (K)
24
Takeaway 2:
More distributed CDNs achieve higher throughput than
more centralized CDNs, but…
… are more expensive for same throughput
25
Conclusion
• A simple model to quantify CDN design choices
– Increasing the number of peering points
– Improving path selection
– More distributed vs more centralized design
• Optimizations at the edge is enough for CDNs
– Multipath has little benefit over increasing # locations
and choosing different peering links
– There’s a tradeoff of throughput and cost among CDNs
26
Thanks!
Questions?
27