version: { Do cument reo rganized fo r b etter readabilit y { Mino r changes to pa rameters from simulation exp erience { Option to relax SPF b est path criteria (Dave W a rd) { Pseudco de included in app endices to aid implemento rs
Up
coming changes: { Optimization
f
pa rtial paths
Multiple
vendo rs considering implementation. { Not sure if any have sta rted co ding.
SLIDE 3OMP Algo rithm Highlights
Flo
d
Loading Info rmation via OSPF
r
IS-IS { Router samples
wn
SNMP counters every 15 seconds { Filters and
ds
dep ending
n
load level and change
F
rw
a rding (OSPF-OMP , ISIS-OMP , MPLS-OMP at ingress) { Compute Hash
n
IP source/destination { Select from available paths based
n
hash value { 14-16 bit hash p rovides ne adjustment granula rit y
Load
Adjustment (OSPF-OMP , ISIS-OMP , MPLS-OMP) { load adjustment through change in hash b
unda
ry { small initial adjustment { exp
nential
increase in adjustment increment { increment is halved when adjustment reverses
P
ath Setup (MPLS-OMP
nly)
{ Setup new paths after p ersistent high utilization { Remove extra paths after p ersistent lo w utilization { TE do es not dep end
n
ca reful conguration
f
IGP link metrics
SLIDE 4Flo
d
Loading Info rmation via OSPF
r
IS-IS
Router
samples
wn
SNMP counters every 15 seconds { Counters a re iffIn,OutgfOctets,P ack et,Disca rdg
Filter
using a few compa re, shift, and add
p
erations
Compute
\equivalent load" as describ ed in OSPF-OMP
Check
fo r reo
ding
based
n:
{ Time elapsed since last
ding
{ Greater
f
current load and last
ded
value { P ercent change since last
ded
value
When
needed, reo
d
and reco rd time and value
SLIDE 5F
rw
a rding (OSPF-OMP , ISIS-OMP , MPLS-OMP at ingress)
After
SPF create \next hop" data structures
Compute
Hash
n
IP source/destination p er pack et { 14-16 bit hash p rovides ne adjustment granula rit y { CRC16 seems to w
rk
ne. Others ma y b e used.
Select
from available paths based
n
hash value { Compa re hash value to b
unda
ries in \next hop" struct { F
rw
rding
is lik e ECMP except load split is unequal { Adjustments to b
unda
ries will adjust load split
Cha
racteristics: { No transient routing lo
ps
r
drops { No microo w pack et reo rdering except during adjustment { Adjustments a re every few minutes (minimal reo rdering)
adjustment reverses, adjustment increment is halved { Halving the rate
n
reversal insures stabilit y { Stabilit y has not b een mathematically p roven, but simu- lation results strongly indicate stabilit y { When load stabilizes,
ding
rate also drops
SLIDE 7OSPF-OMP when used with MPLS-OMP
Besides
doing what OSPF no rmally do es, an interio r router in an MPLS-OMP domain do es the follo wing: { Sample its
wn
SNMP counters every 15 seconds. { Apply simple lter to SNMP sampled data. { Determine when to
d
itered result acco rding to guidelines in OSPF-OMP
Ingress
routers must also do the follo wing: { Setup MPLS LSP path sets acco rding to MPLS-OMP { Adjust loading
n
path sets acco rding to OSPF-OMP { Hash
n
src/dst and fo rw a rd acco rding to OSPF-OMP
SLIDE 8A Simple Example
F B D G E C A
2 2 2 2 2 2 3 6 2
Majo
r ingress and egress a re A, F and E, G
Majo
r
ws
a re A-E = F-G = F-E = 0.5, A-G = 1
Link
E-G is double capacit y
f
thers
Link
costs a re as sho wn in the circles
Utilizations:
Red = 1, Green = 0.5
SLIDE 9First Opp
rtunit
y fo r Load Adjustment
F B D G E C A
No
de B can move load from B-D-E to B-C-E
Utilizations
f
B-C, C-E, and D-E app roach 0.75
Utilizations
f
B-D drops to 0.25
SLIDE 10Second Opp
rtunit
y fo r Load Adjustment
F B D G E C A
No
de F can move load from F-G to F-D-E
No
de B will continue to move load from B-D-E to B-C-E
Utilizations
f
B-C, C-E, and D-E, F-G app roach 0.83
Utilizations
f
F-G and E-G app roach 0.67
No
de F will actually not w ait until D-E loading has reached 0.75, it will sta rt moving load when D-E loading is noticed to b e lo w er than F-G
SLIDE 11Second Opp
rtunit
y fo r Load Adjustment
F B D G E C A
No
de A can move load from A-B-fCDg-E-G to A-F-G
No
de F will continue to move load fro F-G to F-D-E
No
de B will continue to move load from B-D-E to B-C-E
Utilization
f
A-B will app roach 0.83
Utilizations
f
F-G and E-G also app roach 0.83
No
de A will sta rt moving load when F-G loading is noticed to b e lo w er than A-B
A-F
go es to 0.67, B-D app roaches zero
SLIDE 12The Need fo r P a rtial P ath Optimization
B D G E C A F
2 2 2 2 2 3 2 6 2
Consider
trac in the reverse direction
W
rst
loading
n
the E-C-B-A path load
f
1.0
n
D-A
W
rst
loading
n
the E-D-B-A path is
n
D-A and E-D
Moving
load from E-D-B-A to E-C-B-A do es not reduce load
n
the link D-A so it do es not reduce the load
n
E-D-B-A.
No
de E will not move load from E-D-B-A to E-C-B-A
SLIDE 13V alidating the Algo rithms
Simulations
a re at http://engr.ans.net/ospf-omp { tuto rial directo ry has simple examples { simulations directo ry has la rger top
logies
{ simulations directo ry has adverse conditions cases
link
failure
fast
rise in
erred
load
high
noise in
erred
load
la
rge drift
ver
time
Simulations
coverage: { OSPF-OMP is completely covered. { ISIS-OMP is not implemented at all. { MPLS-OMP LSP deletion is not implemented. { MPLS-OMP link failure is not implemented. { MPLS simulations a re not y et
n
the w eb page. { If UUNET simulations cannot b e made available, simula- tions using a complex hyp
thetical
top
logy
a re needed.
SLIDE 14Summa ry
Algo
rithms a re b eing validated through simulations. { Most a re available at http://engr.ans.net/ospf-omp
Status:
draft-ietf-ospf-omp-01.txt,ps { Little change to the algo rithms since
00
version. { Substantial imp rovement to the do cument since
00.
{ Optimization
f
pa rtial paths needs to b e added. { Another iteration
