2. Searching
pa,erns
in
SNMP
data
• Looked
at
two
busy
links:
– star-‐sdn1/interface/xe-‐7_3_0
– denv-‐cr2/interface/xe-‐1_1_0
-‐Generated
graphs
for
the
last
6
months
(including
graphs
per
month,
per
week
)
-‐
Visually
inspected
whether
there
is
any
Ime
related
specific
pa,ern
in
bandwidth
usage
3. Searching
pa,erns
in
SNMP
data
– star-‐sdn1/interface/xe-‐7_3_0
star-‐>wash
h6ps://sdm.lbl.gov/~balman/temp/1-‐a/
– denv-‐cr2/interface/xe-‐1_1_0
sunn-‐>denv
h6ps://sdm.lbl.gov/~balman/temp/1/
4. Searching
pa,erns
in
SNMP
data
• Collected
data
for
those
two
links
(one
year
long)
and
tried
to
analyze
the
data
with
a
machine
learning
soMware
• Converted
data
into
arff
format
• Used
Weka
• Evaluated
the
bandwidth
vs.
Ime
data
(Ime
series
analysis)
to
see
whether
day
of
the
week,
PM
or
AM,
day
of
the
year,
etc.
have
any
visible
effect
on
bandwidth
usage
8. Searching
pa,erns
in
SNMP
data
• Our
iniIal
results
on
Ime
series
predicIon
gave
40-‐50%
error
rate.
• By
using
some
other
techniques,
we
were
able
to
achieve
30-‐40
%
error
rate.
• At
this
moment,
taking
average
link
usage
may
be
a
reasonable
way
to
start
with.
• Further
study
is
required
to
make
useful
predicIons
– Gretl
is
also
another
alternaIve
– Using
R
instead
of
Weka