BGP table fragmentation

BGP table fragmentation: what & who?
Julien Gamba
Romain Fontugne, Cristel Pelsser, Randy Bush, Emile Aben
University of Strasbourg
1
Why looking at fragmentation?
647 000
2
Why looking at fragmentation?
647 000
Amount of IP prefixes announced on the Internet (and counting)
2
Why looking at fragmentation?
647 000
Amount of IP prefixes announced on the Internet (and counting)
• Large routing tables consume memory
• Routers memory (TCAM) is expensive, so they do not have a lot of it
(especially the older ones)
• Too many routes means routers slowing down or shutting down
2
Why looking at fragmentation?
• Some of these prefixes are fragmented and could be aggregated
3
Why looking at fragmentation?
• Some of these prefixes are fragmented and could be aggregated
• Some of these prefixes are redundant and could be removed from the
routing table
3
Why looking at fragmentation?
• Some of these prefixes are fragmented and could be aggregated
• Some of these prefixes are redundant and could be removed from the
routing table
• How bad is the fragmentation?
• How many prefixes are redundant?
• What are the causes of all this?
3
IP prefixes classification
• Top: is covering some
smaller prefix(es);
• Deaggregated: is covered
by another prefix which is
originated by the same AS;
• Delegated: is covered by
another prefix which is
originated by another AS;
• Lonely: does not overlap
with any other prefix.
Luca Cittadini et al. (2010). “Evolution of Internet Address Space Deaggregation:
Myths and Reality”. In: IEEE journal on selected areas in communications
4
IP prefixes classification
• Top: is covering some
smaller prefix(es);
• Deaggregated: is covered
by another prefix which is
originated by the same AS;
• Delegated: is covered by
another prefix which is
originated by another AS;
• Lonely: does not overlap
with any other prefix.
Luca Cittadini et al. (2010). “Evolution of Internet Address Space Deaggregation:
Myths and Reality”. In: IEEE journal on selected areas in communications
5
IP prefixes classification
• Top: is covering some
smaller prefix(es);
• Deaggregated: is covered
by another prefix which is
originated by the same AS;
• Delegated: is covered by
another prefix which is
originated by another AS;
• Lonely: does not overlap
with any other prefix.
Luca Cittadini et al. (2010). “Evolution of Internet Address Space Deaggregation:
Myths and Reality”. In: IEEE journal on selected areas in communications
6
IP prefixes classification
• Top: is covering some
smaller prefix(es);
• Deaggregated: is covered
by another prefix which is
originated by the same AS;
• Delegated: is covered by
another prefix which is
originated by another AS;
• Lonely: does not overlap
with any other prefix.
Luca Cittadini et al. (2010). “Evolution of Internet Address Space Deaggregation:
Myths and Reality”. In: IEEE journal on selected areas in communications
7
IP prefixes classification
• Top: is covering some
smaller prefix(es);
• Deaggregated: is covered
by another prefix which is
originated by the same AS;
• Delegated: is covered by
another prefix which is
originated by another AS;
• Lonely: does not overlap
with any other prefix.
Luca Cittadini et al. (2010). “Evolution of Internet Address Space Deaggregation:
Myths and Reality”. In: IEEE journal on selected areas in communications
8
Deaggregation evolution — prefixes
• routeviews data, as seen
by AS 3356 (Level3),
counting prefixes
Evolution of prefixes classification - IPv4
Percentage of prefixes
50
40
30
• Proportion of deaggregated
is increasing over time
20
10
0
2002
2004
2006
Deaggregated
Delegated
2008 2010
Year
2012
Lonely
Top
2014
2016
• Combined fraction of
deaggregated and delegated
prefixes is constant
AS10
AS10
AS20
AS30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
Top
Deaggregated
Delegated
Lonely
9
Deaggregation evolution — addresses
Evolution of IPv4 addresses classification
Percentage of addresses
60
• Deaggregated addresses are
still increasing, top addresses
are increasing too
50
40
30
20
• Lonely addresses are
decreasing
10
0
2002
2004
2006
Deaggregated
Delegated
2008 2010
Year
2012
2014
2016
Lonely
Top
routeviews data, as seen by AS 3356 (Level3)
AS10
AS10
AS20
AS30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
Top
Deaggregated
Delegated
Lonely
10
Deaggregation evolution
In other words
• There are more and more deaggregated prefixes
• Combined fraction of deaggregated and delegated prefixes is constant
• The proportion of lonely addresses is decreasing, while top and
deaggregated increase
11
Deaggregation evolution
In other words
• There are more and more deaggregated prefixes
• Combined fraction of deaggregated and delegated prefixes is constant
• The proportion of lonely addresses is decreasing, while top and
deaggregated increase
What is going on?
• Are lonely prefixes becoming top? Deaggregated?
• Do they disappear? Is there just less new lonely prefixes?
• What are the movements between the categories? What is the
stability of the prefixes?
11
Prefixes dynamics
Movements between categories
del
2001
del
2002
del
2003
del
2004
del
2005
del
2006
del
2007
del
2008
del
2009
del
2010
del
2011
del
2012
del
2013
del
2014
del
2015
del
2016
dea
2001
dea
2002
dea
2003
dea
2004
dea
2005
dea
2006
dea
2007
dea
2008
dea
2009
dea
2010
dea
2011
dea
2012
dea
2013
dea
2014
dea
2015
dea
2016
lon
2001
lon
2002
lon
2003
lon
2004
lon
2005
lon
2006
lon
2007
lon
2008
lon
2009
lon
2010
lon
2011
lon
2012
lon
2013
lon
2014
lon
2015
lon
2016
top
2001
top
2002
top
2003
top
2004
top
2005
top
2006
top
2007
top
2008
top
2009
top
2010
top
2011
top
2012
top
2013
top
2014
top
2015
top
2016
Most movements appear between lonely and deaggregated prefixes, both ways
Lonely prefixes become deaggregated when the covering prefix is announced
12
Prefixes disappearances
• Counting how many prefixes
disappear from the RIB
between two months
Prefix disappearing over time
Percentage of prefixes
3
2.5
2
• Graph show average per year
1.5
1
0.5
0
2002
2004
2006
Deaggregated
Delegated
2008 2010
Year
2012
2014
2016
Lonely
Top
routeviews data, as seen by AS 3356 (Level3)
• Same trend across
categories, although
deaggregated prefixes
disappear more
AS10
AS10
AS20
AS30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
Top
Deaggregated
Delegated
Lonely
13
Stability of prefixes
Deaggregated prefixes seems to be the most volatile kind:
• Why is that?
• Is it traffic engineering?
• Is it something else?
14
Detection of traffic engineering
AS path prepending
• Graph shows average per
year
Proportion of prefixes using path prepending
Percentage by category
40
a
35
30
25
20
15
• Showing the proportion of
prefixes announced using
path prepending in each
category
10
5
0
2002
2004
2006
Deaggregated
Delegated
2008 2010
Year
2012
2014
2016
Lonely
Top
routeviews data, as seen by AS 3356 (Level3)
• Path prepending usage is
slowly increasing
• No clear trend between
categories
AS10
AS10
AS20
AS30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
Top
Deaggregated
Delegated
Lonely
15
Selective announcements
• Graph shows average per
year
Use of selective announcements
Percentage of prefixes
25
• Showing the proportion of
prefixes seen by less than
half the peers by category
20
15
10
5
0
2002
2004
2006
Deaggregated
Delegated
2008 2010
Time
2012
Lonely
Top
2014
2016
• Lonely and deaggregated
prefixes are the categories
that use selective
announcements the most
AS10
AS10
AS20
AS30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
Top
Deaggregated
Delegated
Lonely
16
Who is engineering traffic?
• Graph shows proportion of
deaggregated prefixes by AS
business type
Cumulative distribution fonction
1
0.8
0.6
• Large transit AS list comes
from CAIDA’s classification
0.4
0.2
Large transit (30)
Content (137)
Others (53564)
0
0
20
40
60
80
Percentage of deaggregated prefixes per AS
100
• Large transit providers may
split their address space to
do traffic engineering
AS10
AS10
AS20
AS30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
Top
Deaggregated
Delegated
Lonely
17
Aggregability of lonely prefixes
Aggregables lonely prefixes
300000
Two prefixes are aggregables if:
Number of prefixes
250000
200000
• they have the same AS origin
150000
• they are consecutive
100000
50000
0
2002
2004
2006
All lonely prefixes
With aggregation
2008 2010
Year
2012
2014
2016
• the aggregate falls on a
power of two boundary
Difference
routeviews data, as seen by AS 3356 (Level3)
AS10
AS10
AS20
AS30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
Top
Deaggregated
Delegated
Lonely
18
In conclusion
• Deaggregation has been increasing for the last fifteen years
• Combined fraction of deaggregated and delegated prefixes is constant
• Some large transit AS split heavily their address space
19
In conclusion
• Deaggregation has been increasing for the last fifteen years
• Combined fraction of deaggregated and delegated prefixes is constant
• Some large transit AS split heavily their address space
We still don’t know:
• why the heavy splitting?
• for traffic engineering?
• for security?
19
Questions?