Paolo Lucente, author of the software pmacct.
Traffic matrices can greatly benefit key IP Service Provider activities like capacity planning, traffic engineering, better understand their traffic patterns and take meaningful peering decisions. This talk wants to present a way to build traffic matrices using telemetry data and BGP, leveraging along the way some case-studies with a technical cut. pmacct (http://www.pmacct.net) is a commonly used, free, open-source IPv4/IPv6a ccounting package which integrates a NetFlow/sFlow and a multi-RIB BGP collector in a single piece of software and is authored by the presenter.
1 of 35
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
More Related Content
Traffic Matrices and its measurement
1. Universitat Politecnica de Catalunya, Castelldefels – Jun, 5th 2013
Traffic
Matrices
and
its
measurement
Paolo Lucente
pmacct
2. Agenda
o Introduction
o The tool: pmacct
o Setting the pitch
o Case study: peering at AS286
Traffic
Matrices
and
its
measurement
Universitat Politecnica de Catalunya, Castelldefels – Jun, 5th 2013
3. Why
speaking
of
traffic
matrices?
– A;er
focusing
on
traffic
matrices
during
your
course
you
can
find
this
ques>on
silly.
But
..
§ In
the
industry,
finding
knowledgeable
people
on
the
subject
is
hard
§ Quite
unspoken
subject
at
technical
conferences
§ Presenta>ons
(nega>vely)
stop
to
the
math
4. Why
speaking
of
traffic
matrices?
– So,
a;er
all,
are
traffic
matrices
useful
to
a
network
operator
in
the
first
place?
Yes
…
§ Capacity
planning
(build
capacity
where
needed)
§ Traffic
Engineering
(steer
traffic
where
capacity
is
available)
§ BeQer
understand
traffic
paQerns
(what
to
expect,
without
a
crystal
ball)
§ Support
peering
decisions
(traffic
insight,
traffic
engineering
at
the
border,
support
what
if
scenarios)
7. Let’s
focus
on
an
external
traffic
matrix
to
support
peering
decisions
– Analysis
of
exis>ng
peers
and
interconnects:
§ Support
policy
and
rou>ng
changes
§ Fine-‐grained
accoun>ng
of
traffic
volumes
and
ra>os
§ Determine
backbone
costs
associated
to
peering
§ Determine
revenue
leaks
– Planning
of
new
peers
and
interconnects:
§ Who
to
peer
next
§ Where
to
place
next
interconnect
§ Modeling
and
forecas>ng
8. Our
traffic
matrix
visualized
P
P
P
PE
A
PE
D
PE
C
PE
B
P3
P1
P2
P4
BZ
A
=
{
src_as,
in_iface,
peer_src_ip,
peer_dst_ip,
as_path,
tstamp,
bytes
}
{
CZ,
X
(-‐>
CY),
PE
C,
PE
A,
AZ_AY,
<>me>,
<bytes>
}
{
BX,
Y
(-‐>
BY),
PE
B,
PE
C,
CY_CX,
<>me>,
<bytes>
}
{
AX,
Z
(-‐>
AZ),
PE
A,
PE
B,
BY_BZ,
<>me>,
<bytes>
}
9. Our
traffic
matrix
shopping
cart
– What
is
needed:
§ BGP
§ (IGP)
§ Telemetry
data:
NetFlow,
sFlow
§ Collector
infrastructure:
tool,
system(s)
§ Storage:
either
of
RDBMS,
RRD,
noSQL
§ Maintenance
and
post-‐processing
scripts
– Risks:
§ 800
pound
gorilla
project
10. Agenda
o Introduction
o The tool: pmacct
o Setting the pitch
o Case study: peering at AS286
Traffic
Matrices
and
its
measurement
Universitat Politecnica de Catalunya, Castelldefels – Jun, 5th 2013
11. pmacct
is
free
open-‐source
GPL’ed
so;ware
pmacct
libpcap
BGP
maps
IS-IS
MySQL
PostgreSQL
SQLite
3
MongoDB
BerkeleyDB
flat-‐files
memory
tables
NetFlow
IPFIX
sFlow
sFlow
NetFlow
IPFIX
tee
12. Why
pmacct
was
started
anyway?
– I
was
23
– I
had
just
joined
CNR
(Na>onal
Council
for
Research)
– I
thought
you
can
manage
IP
networks
like
that
13. Why
pmacct
was
started
anyway?
– Before
turning
24
I
realized
I
was
wrong:
§ Too
much
data
(ie.
flow-‐tools)
and
canned
reports
(ie.
ntop)
were
causing
lack
of
visibilty
in
my
backbone.
§ Not
everything
was
invented
yet!
15. Key
pmacct
non-‐technical
facts
§ A
name
you
can’t
spell
a;er
the
second
drink
§ 10
years
old
project
§ Free,
open-‐source,
independent
§ Under
ac>ve
development
§ Innova>on
being
introduced
§ Well
deployed
around,
also
large
SPs
§ Aims
to
be
the
traffic
accoun>ng
tool
closer
to
the
SP
community
needs
16. Some
technical
facts
§ Pluggable
architecture
• Straighnorward
to
add
support
for
new
collec>on
methods
or
backends
§ An
abstrac>on
layer
allows
out-‐of-‐the-‐box
any
collec>on
method
to
interact
with
any
backend
§ Both
mul>-‐process
and
(coarse)
mul>-‐threading
• Mul>ple
plugins
(of
same
or
different
type)
can
be
instan>ated
at
run>me,
each
with
own
config
17. Introducing
BGP
na>vely
into
a
NetFlow/sFlow
collector
– pmacct
introduced
a
Quagga-‐based
BGP
daemon
§ Implemented
as
a
parallel
thread
within
the
collector
§ Doesn’t
send
UPDATEs
and
WITHDRAWs
whatsoever
§ Behaves
as
a
passive
BGP
neighbor
§ Maintains
per-‐peer
BGP
RIBs
§ Supports
32-‐bit
ASNs;
IPv4
and
IPv6
families
– Why
BGP
at
the
collector?
§ Telemetry
reports
on
forwarding-‐plane
§ Telemetry
should
not
move
control-‐plane
informa>on
over
and
over
18. Agenda
o Introduction
o The tool: pmacct
o Setting the pitch
o Case study: peering at AS286
Traffic
Matrices
and
its
measurement
Universitat Politecnica de Catalunya, Castelldefels – Jun, 5th 2013
19. Getng
BGP
to
the
collector
– Let
the
collector
BGP
peer
with
all
PE
devices:
facing
peers,
transit
and
customers.
– Determine
memory
footprint
(below
in
MB/peer)
44.03
22.73
19.97
18.59
18.12
17.89
17.76
17.57
17.48
17.39
50
0
10
20
30
40
50
60
0
200
400
600
800
1000
1200
1400
MB/peer
>=
0.12.4
MB/peer
<
0.12.4
Number
of
BGP
peers
MB/peer
500K
IPv4
routes,
50K
IPv6
routes,
64-‐bit
executable
~
9GB
total
memory
@
500
peers
20. – Set
the
collector
as
iBGP
peer
at
the
PE
devices:
§ Configure
it
as
a
RR
client
for
best
results
§ Collector
acts
as
iBGP
peer
across
(sub-‐)ASes
– BGP
next-‐hop
has
to
represent
the
remote
edge
of
the
network
model:
§ Typical
scenario
for
MPLS
networks
§ Can
be
followed
up
to
cover
specific
scenarios
like:
• BGP
confedera>ons
– Op>onally
polish
the
AS-‐Path
up
from
sub-‐ASNs
• default
gateway
defined
due
to
par>al
or
default-‐only
rou>ng
tables
Getng
BGP
to
the
collector
(cont.d)
21. Getng
telemetry
to
the
collector
– Export
ingress-‐only
measurements
at
all
PE
devices:
facing
peers,
transit
and
customers.
§ Traffic
is
routed
to
des>na>on,
so
plenty
of
informa>on
on
where
it’s
going
to
§ It’s
crucial
instead
to
get
as
much
as
possible
about
where
traffic
is
coming
from
– Leverage
data
reduc>on
techniques
at
the
PE:
§ Sampling
§ Aggrega>on
(but
be
sure
to
carry
IP
prefixes!)
23. Storing
data
persistently
– Data
need
to
be
aggregated
both
in
spa>al
and
temporal
dimensions
before
being
wriQen
down:
§ Op>mal
usage
of
system
resources
§ Avoids
expensive
consolida>on
of
micro-‐flows
§ Suitable
for
project-‐driven
data-‐sets
– Open-‐source
RDBMS
appear
a
natural
choice
§ Able
to
handle
large
data-‐sets
§ Flexible
and
standardized
query
language
§ Solid
and
scalable
(clustering,
par>>oning)
– noSQL
DBs
picking
up:
between
myth
and
bust
..
24. Briefly
on
scalability
– A
single
collector
might
not
fit
it
all:
§ Memory:
can’t
store
all
BGP
full
rou>ng
tables
§ CPU:
can’t
cope
with
the
pace
of
telemetry
export
§ Divide-‐et-‐impera
approach
is
valid:
• Assign
PEs
(both
telemetry
and
BGP)
to
collectors
• Assign
collectors
to
RDBMSs;
or
cluster
the
RDBMS.
– The
matrix
can
get
big,
but
can
be
reduced:
§ Keep
smaller
routers
out
of
the
equa>on
§ Filter
out
specific
services/customers
on
dense
routers
§ Focus
on
relevant
traffic
direc>on
(ie.
upstream
if
CDN,
downstream
if
ISP)
§ Sample
or
put
thresholds
on
traffic
relevance
25. Agenda
o Introduction
o The tool: pmacct
o Setting the pitch
o Case study: peering at AS286
Traffic
Matrices
and
its
measurement
Universitat Politecnica de Catalunya, Castelldefels – Jun, 5th 2013
26. Case-‐study:
peering
at
AS286
Nego>a>on
Engineering
Produc>on
Op>miza>on
– Peering
as
a
cycle
– NetFlow
+
BGP
traffic
matrix
steers
peering
op>miza>on:
§ Iden>fy
new
and
“old”
peers
§ Traffic
analysis:
backbone
costs,
95th
percen>les,
ra>os
§ Analysis
of
interconnec>on
density
and
traffic
dispersion
§ Forecas>ng
and
trending
§ Ad-‐hoc
queries
from
Design
&
Engineering
and
indeed
…
the
IPT
Product
Manager
27. Case-‐study:
peering
at
AS286
– 250+
Gbps
rou>ng-‐domain
– 100+
high-‐end
routers
around
the
globe:
§ Export
sampled
NetFlow
§ Adver>se
full
rou>ng
table
§ Mix
of
Juniper
and
Cisco
– Collector
environment:
§ Runs
on
2
Solaris/SPARC
zones
§ pmacct:
dual-‐core,
4GB
RAM
§ MySQL:
quad-‐core,
24GB
RAM,
500
GB
disk
– Data
reten>on:
6
months
AS286
rou>ng
domain
pmacct
MySQL
Internal
users
28. – AS286
backbone
routers
are
first
configured
from
templates:
§ NetFlow
+
BGP
collector
IP
address
defined
over
there
§ Enabler
for
auto-‐discovery
of
new
devices
– Edge
interfaces
are
provisioned
following
service
delivery
manuals:
§ Relevant
manuals
and
TSDs
include
NetFlow
ac>va>on
§ Periodic
checks
NetFlow
is
ac>ve
where
it
should
– Maps,
ie.
source
peer-‐AS,
are
re-‐built
periodically
Case-‐study:
peering
at
AS286
29. Thanks for your attention!
Questions? Now or later ..
Paolo Lucente
pmacct
<paolo at pmacct dot net>
Keep in touch via LinkedIn
Traffic
Matrices
and
its
measurement
Universitat Politecnica de Catalunya, Castelldefels – Jun, 5th 2013
30. Backup slides
Traffic
Matrices
and
its
measurement
Universitat Politecnica de Catalunya, Castelldefels – Jun, 5th 2013
31. IGP
(IS-‐IS)
integra>on
– A
Quagga-‐based
IS-‐IS
daemon
was
introduced:
§ Implemented
as
a
parallel
thread
within
the
collector
§ IS-‐IS
neighborship
over
a
GRE
tunnel
§ Currently
limited
to
single
neighborhip,
single
level,
single
topology
§ Useful
to
look
up
non
BGP-‐routed
networks
§ Promising,
many
plans:
• Implement
a
non
real-‐>me,
IGP-‐protocol
agnos>c
version
• LFA
kind
of
computa>ons
• Collec>ng
feedback
on
interest
about
TRILL
32. Storing
data
persisently:
RDBMS
create table acct_bgp (
agent_id INT(4) UNSIGNED NOT NULL,
as_src INT(4) UNSIGNED NOT NULL,
as_dst INT(4) UNSIGNED NOT NULL,
peer_as_src INT(4) UNSIGNED NOT NULL,
peer_as_dst INT(4) UNSIGNED NOT NULL,
peer_ip_src CHAR(15) NOT NULL,
peer_ip_dst CHAR(15) NOT NULL,
comms CHAR(24) NOT NULL,
as_path CHAR(21) NOT NULL,
local_pref INT(4) UNSIGNED NOT NULL,
med INT(4) UNSIGNED NOT NULL,
packets INT UNSIGNED NOT NULL,
bytes BIGINT UNSIGNED NOT NULL,
stamp_inserted DATETIME NOT NULL,
stamp_updated DATETIME,
PRIMARY KEY (…)
);
BGP
Fields
Counters
Time
Tag
shell> cat peers.map
id=65534 ip=X in=A
id=65533 ip=Y in=B src_mac=J
id=65532 ip=Z in=C bgp_nexthop=W
[ … ]
shell> cat pretag.map
id=100 peer_src_as=<customer>
id=80 peer_src_as=<peer>
id=50 peer_src_as=<IP transit>
[ … ]
– In
any
schema
(a
subset
of)
BGP
primi>ves
can
be
freely
mixed
with
(a
subset
of)
L1-‐L7
primi>ves
33. Storing
data
persisently:
RDBMS
– Traffic
delivered
to
a
BGP
peer,
per
loca>on:
mysql> SELECT peer_as_dst, peer_ip_dst, SUM(bytes), stamp_inserted
FROM acct_bgp
WHERE peer_as_dst = <peer | customer | IP transit> AND
stamp_inserted = < today | last hour | last 5 mins >
GROUP BY peer_as_dst, peer_ip_dst;
– Aggregate
AS
PATHs
to
the
second
hop:
mysql> SELECT SUBSTRING_INDEX(as_path, ‘.’, 2) AS as_path, bytes
FROM acct_bgp
WHERE local_pref = < IP transit pref> AND
stamp_inserted = < today | yesterday | last week >
GROUP BY SUBSTRING_INDEX(as_path, ‘.’, 2)
ORDER BY SUM(bytes);
– Focus
peak
hour
(say,
8pm)
data:
mysql> SELECT … FROM … WHERE stamp_inserted LIKE ‘2010-02-% 20:00:00’
…
34. Storing
data
persisently:
RDBMS
– Traffic
breakdown,
ie.
top
N
grouping
BGP
peers
of
the
same
kind
(ie.
peers,
customers,
transit):
mysql> SELECT … FROM … WHERE …
local_pref = <<peer | customer | IP transit> pref>
…
– Download
traffic
matrix
(or
a
subset
of
it)
to
3rd
party
backbone
planning/traffic
engineering
applica>on
(ie.
Cariden,
Wandl,
etc.):
mysql> SELECT peer_ip_src, peer_ip_dst, bytes, stamp_inserted
FROM acct_bgp
WHERE [ peer_ip_src = <location A> AND
peer_ip_dst = <location Z> AND … ]
stamp_inserted = < today | last hour | last 5 mins >
GROUP BY peer_ip_src, peer_ip_dst;
35. Storing
data
persisently:
MongoDB
– pmacct
opening
to
noSQL
databases
– noSQL
landscape
difficult
to
move
through,
ie.
fragmented
and
lacks
of
standardiza>on
– MongoDB
seemed
interes>ng
for:
§ Its
na>ve
grouping
opera>on
(more
performing
and
less
complex
than
map/reduce)
§ Its
horizontal
scaling
concept
(called
sharding)
– Currently
collec>ng
feedbck:
some
preliminar
tes>ng
at
one
large
operator
looks
promising
..