greening the internet with nano data...
TRANSCRIPT
Greening the Internet with Nano Data Centers
V. Valancius (Georgia Institute of Technology), N. Laoutaris (Telefonica Research), L. Massoulie, C. Diot (Thomson), P. Rodriguez (Telefonica Research)
Presented by Sean Barker
2
Gateway vs. Set-top-box
Internet
Home gateway• Internet connection• Traffic management
Set-top-boxes• Content processing
3
Gateway as new center of media experience
The Internet
Content & ControlServers
Internet services
Extended home network
Homenetwork
4
The current model: Data centers
The Internet
Data
Centers
.
.
.DSLAM
Home boxes
5
Limitations
Expensive – High capital investment– Customer generally pays per byte
Location constraints in order to be “central” Requires a lot of redundancy to be robust– Electricity shortage– Content availability
Power, power, power New service deployment is slow– ISPs not encouraged to take risks, nor to deploy new services
6
The nano data center
Take advantage of always-on gateways
Add memory and stronger CPU to home gateways
Push content to gateways when bandwidth is cheap
Manage millions of gateways as a logical ‘single server’ using P2P infrastructure
7
The nano data center model
The Internet
Video server
Control server
.
.
.DSLAM
Home boxes
8
The nano data center
PROS– Multiple applications can take
advantage of the model (VoD, gaming, catchTV, UGC)
– ISP friendly– Reduces traffic volumes and
variability on backbones.– Highly scalable and robust by
design– Cheap for ISPs– Flexible for users– Localized & personalized
services
CONS– Uplink bandwidth often limited– Millions of boxes to manage
using P2P– Cost of gateway– Incentive?– Privacy?– Always on?
9
Gateway uptime (*)
(*) Courtesy of Krishna Gummadi
More than 60% of gateways are up more than 80% of the time
10
Push phase
Video content server
Control server
.
.
.DSLAM
RHGs
11
Pull phase
Control server
.
.
.DSLAM
RHGs
Video content server
12
Placement strategy
……
1
Data window
… 22 1M
…
… 2 M 1 M3 3 3 ……
Movie
11… 1 22 2 33 3 MM…M
1 22 1K
…
2 K 1 K3 3 3 …… ……
Erasure codes(e.g., LDPC)
gateways
13
Placement strategy
Replicate content according to popularity
Popular content served by gateways– slack bandwidth from original content servers
Number of replicas determined by solving optimization problem
– Constraints on available upload and storage, number of clients, request rates, etc.
14
Popularity aware placement
Partition content into hot / warm / cold categories• Hot: replicate on all gateways• Warm: use code-based placement• Cold: no proactive placement (stays on servers)
hotwarm
cold
b/wmemory
popularity
movies
15
Where should NaDa save energy?
The Internet
Data
Centers
.
.
.DSLAM
Home boxes
Data Centers: capital cost, redundancy,
cooling
Internet: less hops, less bandwidth
Free ride on always-on Gateways
16
Energy issues
Variables– Network topology– Hardware power consumption– Placement algorithms– Content popularity– User behavior
Data available– DSL gateways and VoD servers power (Thomson)– Routers power (Cisco data)– Telefonica Spain and Peru network topologies– Imagenio VoD platform (Telefonica Spain)– Telefonica IPTV– Netflix movie popularity
αs = 4%
17
VoD server power
αg = 1%
18
Gateway power
19
When would NaDa not work ?
# bytes transferred
power
Server
Gateways
Baselinepower
20
When does NaDa work ?
# bytes transferred
power Server
Gateway
Baselinepower s
N
21
Trace driven simulations
Traces from – Netflix, IPTV (Telefonica), YouTube
Content popularity from Netflix Topologies and workload from Telefonica Power numbers from Thomson’s gateway and IPTV
servers Popularity aware placement
22
Simulation parameters
Gateway Storage 100MB-10GBGateway Upstream 0.1-2MbpsContent characteristics from data set Users 10k-30kContent window 10s-120sReplicas for warm content 1 (20s windows)Simulation duration 1 day - 86400 sRouter energy/bit 150 10−9
Server energy/bit 40 10−9
Gateway energy/bit 18 10−9
Power Usage Effectiveness (PUE) 1.7Home electricity cost factor 1.1Hops to server 4Hops to peer 2
23
Total energy use (YouTube)
10k
20k
30k
40k
50k
60k
24
Gateway storage
2500
3000
3500
4000
4500
5000
500 1,000 1,500 2,000 2,500 3,000
25
Number of users
2500
3000
3500
4000
4500
5000
5500
15,000 20,000 25,000 30,000
Number of users
26
Upstream bandwidth
3000
3500
4000
4500
5000
0.2 0.4 0.6 0.8 1.0 1.2
Video streaming >
uplink rateVideo streaming
< uplink rate
27
Conclusions
Free-riding on existing infrastructure can significantly reduce load on conventional servers
Simulations demonstrated energy savings ranging from 20% to 60% versus data centers
Gateways can accomplish this with only modest resources (a few GB of storage, limited upload)
28
Questions?
Potential QoS issues moving control from content providers (YouTube) to ISPs
Effects of consumer line overprovisioning
Security of content serving from home gateways