Pablorr Att Dtb 08 Dec

An Internet for Bulk Delivery … not communication

Pablo Rodriguez Telefonica Research, Barcelona

Is the Internet the preferred medium for distributing bulk (delay tolerant) digital content?

Not beyond a certain size… e.g. movies, home videos, data backups, scientific data

Currently, served by: -Dedicated networks -Parcel delivery

10M+ users 1.5 Million DVDs per day: 2.5 PB/day! All US P2P traffic: 14 PB/day(cisco) Postal still carries vast amount of multimedia traffic

Data Centres Replication

Akamai ships 1 Tbyte/day logs via postal mail

CERN: 20 TB/day scientific data. Private network

How well is the current Internet dealing with bulk transfers...?

Current bulk data demand is higher than what the Internet can handle

Existing solutions... Existing solutions 2. Congestion-based pricing

 Block all P2P

 Use capped “day-time” volume 9

Try moving 1 Tbyte data of non-P2P data on today’s internet... Happy hours in one place are unhappy in another place Postal network is a breeze

the REAL Problem Internet well suited for communications not delivery

bulk vs shot bursts instantaneous vs delay tolerant

Redesign the Internet as a Cargo network not a Telecommunications one...

Today’s Internet

Internet Postal Service Transit Storage Warehouses

Internet Post Offices

How many storage warehouses/post offices? where? how big? Routing and Scheduling, when? for how long? through where? Reliability Minimize Cost Maximize Speeds

So what is wrong with E2E transfers? Impact of cargo data on network economics load 95-percentile

Cargo Data

time day of week

Water filling

95-percentile

Cargo Data

Cargo Data today can be seen as creating Fat in the Internet Need to move Fat to times with more spare capacity

load

Cargo Data

time

Does simple scheduling at the sender solve the problem?

21

that is not enough…

Non-overlapping valleys

Sender in LAT Receiver in EU or China uvalley(LAT) 8am available rate

U

ε

sender

receiver

1pm

8pm

Minimize Cost or Maximize Cargo delivered Deadline : T

C(b)

b bandwidth price Sender Link

C(b)

b Data: D

load Sender Link

Dynamic Programming

bandwidth price Receiver link

load Receiver link

Data Collected… -Load in 440 interconnection points from large wholesale ISP -Peering with 140+ ISPs -Carries traffic from 12M+ users -3 month period -Pricing charges at each interconnection point

2 transfer policies SnF: Internet Postal E2E Service

26

Crossing the Atlantic from LAT to EU

E2E SnF

E2E SnF

B=100 Tbits, T=2 days

B=250 Tbits, T=2 days

B=250 Tbits, T=3 days

Cargo Data Cost Crossing the Atlantic from LAT to EU 2/2 D=100 T=2

Today E2E Intern et postal service

D=25 0

D=25 0

T=2

T=3

LAT

53

130

46

EU

8

14

9

Total

61

144

55

LAT

0

0

0

EU

0

7

0

Total

0

7

0

SnF 28

Cost (x $1000)

SnF vs E2E-Sched

SnF E2E-Sched

29

E2E-Sched vs. SnF 



E2E-Sched —

Simple

(+)

—

Suffers from non-coinciding valleys

(-)

SnF —

More complicated

(-)

—

Can handle non-coinciding valleys (+)

E2E-Sched vs. SnF

T=1 day

31

Cost of E2E-Sched to match SnF-free

32

Ongoing Work • Can optimize bulk delivery of point to point residential traffic which is being shaped at different times of the day? • What about point to multipoint?

33

Conclusions Communications and bulk delivery are very different beasts: Internet not yet ready Storage increases capacity or decreases cost Many opportunities to mix digital and physical delivery solutions: FedEx for bits

Thanks. questions…?

E2E-Sched vs. SnF vs. FedEx for CERN

37

Rushing cargo data is either very expensive or impossible -Volume based charging -Throttling (under flat rate)

Real FedEx? 250 Tbits; 31 disks ; 38 kilos Argentina => Spain Deliver within 2-3 days $600 (one shipment), $9000 (continuous) $9000 (fedex) €7000 (internet postal service)

E2E-Sched vs. SnF vs. FedEx for CERN

40

27 Tbytes per day (54 Tb/2 days) over one month

3D Printer

3D Printer

Brain back up retina:10 Mbits per second 28 Tbytes/day

Not all bits cost the same and not all bits are equally important

Today the network treats interactive and cargo data traffic indiscriminately

The cost of delivering data

ISPs dimension their networks based on peak hour load or manage traffic accordingly

And as more and more things become digital the demand will grow

Post mail is fine, but the Internet is a lot more convenient