Transcript BGP AS Number Exhaustion

IPv4 Address Lifetime Expectancy
Revisited - Revisited
Geoff Huston
November 2003
Presentation to the IEPG
Research activity
supported by APNIC
The Regional Internet Registries s do not make forecasts or
predictions about number resource lifetimes. The RIRs provide
statistics of what has been allocated. The following presentation
is a personal contribution based on extrapolation of RIR
allocation data.
IPv4 Address Lifetime
Expectancy

In July the IEPG presentation on address
lifetime expectancy used the rate of growth of
BGP advertised address space as the overall
address consumption driver

The presentation analysed the roles of the
IANA and the RIRs and created an overall
model of address consumption
Modelling the Process –
July 2003
IPv4 Model
IANA Pool Exhaustion 2022
RIR Pool Exhaustion 2024
200
IANA
RIR
BGP
IANA-P
RIR-P
BGP-P
RIR
LIR
150
100
50
0
Jan-00
Jan-05
Projections
Jan-10
Jan-15
Jan-20
Jan-25
Address Consumption
Models

The basic assumption was that continued
growth will remain at a constant proportion of
the total advertised address space
(compound growth), and that as a
consequence address exhaustion was
predicted to occur sometime around 2025

Does the advertised address data support
this view of the address growth model?
The Advertised Address
Space
Notes

It’s noisy data



There are 3 /8 prefixes that flap on a
multi-day cycle
There are shorter term flaps of smaller
prefixes
Reduce the noise by:



Removing large steps
Applying gradient filter
Apply averaging to smooth the data
3
ug
-0
A
n03
3
pr
-0
Ju
A
2
-0
b03
Fe
2
2
ct
-0
ec
D
O
ug
-0
A
n02
2
pr
-0
Ju
A
1
-0
b02
Fe
1
1
ct
-0
ec
D
O
ug
-0
A
n01
1
pr
-0
Ju
A
0
-0
b01
Fe
0
0
ct
-0
ec
D
O
ug
-0
A
n00
0
pr
-0
Ju
A
b00
Fe
Smoothed Data (1)
Filter to 18 /8 advertisements
75
73
71
69
67
65
63
61
59
A
3
3
ug
-0
n03
Ju
pr
-0
A
2
-0
2
b03
ec
Fe
D
ct
-0
2
ug
-0
O
A
2
n02
Ju
pr
-0
A
1
-0
1
b02
ec
Fe
D
ct
-0
1
ug
-0
O
A
1
n01
Ju
pr
-0
A
0
-0
0
b01
ec
Fe
D
ct
-0
0
ug
-0
O
A
0
n00
Ju
pr
-0
A
b00
Fe
Smoothed Data (2)
Gradient Filtered
75
73
71
69
67
65
63
61
59
0
Ju
n00
A
ug
-0
0
O
ct
-0
0
D
ec
-0
0
Fe
b01
A
pr
-0
1
Ju
n01
A
ug
-0
1
O
ct
-0
1
D
ec
-0
1
Fe
b02
A
pr
-0
2
Ju
n02
A
ug
-0
2
O
ct
-0
2
D
ec
-0
2
Fe
b03
A
pr
-0
3
Ju
n03
A
ug
-0
3
pr
-0
A
b00
Fe
Model Matching
Applying Models to the Data
77
75
73
71
69
Data
Linear
Polynomial
67
65
63
61
59
But Which Model?

A number of models can be applied
to this data:



Linear model, assuming a constant rate
of growth
Polynomial model, assuming a constant
rate of change of growth
Exponential model, assuming a
geometric growth with a constant
doubling period
First Order Differential of the
data
First Order Differential
6
5.5
5
4.5
4
3.5
3
2.5
2
1.5
1
0
-0
b
Fe
Ap
r-0
0
1
2
3
0
1
2
00
01
02
03
00
01
02
00
01
01
02
02
03
03
-0
-0
-0
-0
-0
-0
cccggggbbbnnnnct
ct
ct
pr
pr
pr
e
e
e
e
e
e
u
u
u
u
u
u
u
u
O
O
O
A
A
A
J
F
J
F
J
F
J
D
D
D
A
A
A
A
Aug-03
Jun-03
Apr-03
Feb-03
Dec-02
Oct-02
Aug-02
Jun-02
Apr-02
Feb-02
Dec-01
Oct-01
Aug-01
Jun-01
Apr-01
Feb-01
Dec-00
Oct-00
Aug-00
Jun-00
Apr-00
Feb-00
Linear Best Fit to Differential
Least Squares Best Fit
6
5.5
5
4.5
4
3.5
3
2.5
2
1.5
1
Growth Rate



The growth rate of 4 – 5 /8 blocks per
year in 99-00 is now appoximately half
that, at 2 – 3 /8 blocks per year
A constant growth model has a best fit
of 3.5 /8 blocks per year
The change in growth over the period is
a decline in growth rate by 0.4 /8 blocks
per year
3
ug
-0
A
3
n03
Ju
pr
-0
A
2
-0
b03
Fe
2
2
ct
-0
ec
D
O
ug
-0
A
2
n02
Ju
pr
-0
A
1
-0
b02
Fe
1
1
ct
-0
ec
D
O
ug
-0
A
1
n01
Ju
pr
-0
A
0
-0
b01
Fe
0
0
ct
-0
ec
D
O
ug
-0
A
0
n00
Ju
pr
-0
A
b00
Fe
Log of Data
Log of smoothed data
4.3
4.25
4.2
4.15
4.1
4.05
4
A
3
3
ug
-0
n03
Ju
pr
-0
A
2
-0
2
b03
Fe
ec
D
ct
-0
2
ug
-0
O
A
2
n02
Ju
pr
-0
A
1
-0
1
b02
Fe
ec
D
ct
-0
1
ug
-0
O
A
1
n01
Ju
pr
-0
A
0
-0
0
0
b01
Fe
ec
D
ct
-0
ug
-0
O
A
0
n00
Ju
pr
-0
A
b00
Fe
Best Fit to Log
Best Fit to Log Data
4.3
4.25
4.2
4.15
4.1
4.05
4
Exponential Model



The exponential model assumes a liner
best fit to the log of the data series
This linear fit is evident across 2000
More recent data shows a negative
declining rate in growth of the log of the
data.
1
-4
9
-3
ec
D
7
-3
ec
D
5
-3
ec
D
3
-3
ec
D
1
-3
ec
D
9
-2
ec
D
7
-2
ec
D
5
-2
ec
D
3
-2
ec
D
1
-2
ec
D
9
-1
ec
D
7
-1
ec
D
5
-1
ec
D
3
-1
ec
D
1
-1
ec
D
9
-0
ec
D
7
-0
ec
D
5
-0
ec
D
3
-0
ec
D
1
-0
ec
D
9
-9
ec
D
ec
D
Projections
Projections
200
150
100
Linear
X**2 Polynomial
Exponential
Average Data
50
0
Observations

Polynomial best fit sees a continuing
decline in growth until growth reaches
zero in 2010


Exponential best fit sees continuing
increase in growth until exhaustion occurs
in 2021


Matches a model of market saturation
Matches a model of uniform continued growth
in all parts of the network
Linear best fit sees constant growth until
exhaustion occurs in 2042

Matches a model of progressive saturation in
existing markets offset by demands in new
Modelling the Process



Assume that the RIR efficiency in allocation slowly
declines, then the amount of RIR-held space
increases over time
Assume that the Unannounced space shrinks at the
same rate as shown over the past3 years
Assume linear best fit model to the announced
address space projections and base RIR and IANA
pools from the announced address space projections
Modelling the Process
IANA Pool Exhaustion 2030
IPv4 Model
RIR Pool Exhaustion 2037
200
IANA
RIR
BGP
IANA-P
RIR-P
BGP-P
RIR
LIR
150
100
Unadvertised Address Pool
50
RIR Holding Pool
0
Jan-00
Jan-05
Jan-10
Projections
Jan-15
Jan-20
Jan-25
Jan-30
Jan-35
Jan-40
Jan-45
Observations



Extrapolation of current allocation practices and
current demand models using an exponential growth
model derived from a 2000 – 2003 data would see
RIR IPv4 space allocations being made for the next 2
decades, with the unallocated draw pool lasting until
2018 - 2020
The use linear growth model sees RIR IPv4 space
allocations being made for the next 3 decades, with
the unallocated draw pool lasting until 2030 – 2037
Re-introducing the held unannounced space into the
routing system over the coming years would extend
this point by a further decade, prolonging the useable
lifetime of the unallocated draw pool until 2038 –
2045
Questions

Externalities:



What are the underlying growth drivers and
how are these best modeled?
What forms of disruptive events would alter
this model?
What would be the extent of the disruption
(order of size of the disruptive address
demand)?