research & development mstp-te load balancing: some results benchmarking carrier ethernet...
Post on 15-Jan-2016
216 views
TRANSCRIPT
research & development
MSTP-TE Load Balancing:Some results
Benchmarking Carrier Ethernet TechnologiesSession AI.3
Krakow, Poland - April 30, 2008
Rémi Clavier - Orange Labs
COBRA Project
research & development
Goals Goals && Assumptions Assumptions
research & development
France Telecom Group
Input data Work done in collaboration with ALF
Topology Define a common "Reference Topology" Define "variations" from this topology
• To have a more exhaustive analysis• To take into account realistic FT aggregation topologies
Matrix Define a common (more or less) realistic traffic matrix including
• http,HSI, VoIP• Pear to Pear• IPTv• VPNs
Define variation from this Matrix • To try to catch P2P / P2MP influence
Residential NO-1 NO-2 NO-3 NO-4 NO-5 NO-6 NO-11 NO-12 NO-13 NO-14 NO-15 NO-16NO-1 359.1 359.1 359.1 623.7 623.7 276.48NO-2 359.1 623.7 623.7 359.1 359.1 276.48NO-3 359.1 623.7 623.7 359.1 359.1 276.48NO-4 359.1 623.7 623.7 359.1 359.1 276.48NO-5 623.7 359.1 359.1 359.1 623.7 276.48NO-6 623.7 359.1 359.1 359.1 623.7 276.48NO-11 18.5 18.5 18.5NO-12 18.5 18.5 18.5NO-13 18.5 18.5 18.5NO-14 18.5 18.5 18.5NO-15NO-16 1455.9 1455.92 1455.92 1455.92 1455.92 1455.92
iptv NO-1 NO-2 NO-3 NO-4 NO-5 NO-6NO-15 520 520 520 520 520 520
vpn1 NO-1 NO-2 NO-3 NO-4 NO-5 NO-6NO-1 137.2 137.2 137.2 137.2 137.2NO-2 137.2 137.2 137.2 137.2 137.2NO-3 137.2 137.2 137.2 137.2 137.2NO-4 137.2 137.2 137.2 137.2 137.2NO-5 137.2 137.2 137.2 137.2 137.2NO-6 137.2 137.2 137.2 137.2 137.2
vpn23 NO-1 NO-2 NO-3 NO-4 NO-5 NO-6NO-1 58.6 58.6NO-2 58.6 58.6NO-3 58.6 58.6NO-4 58.6 58.6NO-5 58.6 58.6NO-6 58.6 58.6
research & development
ExperimentsExperiments
research & development
France Telecom Group
Some definitions
For each "variation", different experiments were done
CL/CO CL (as Connection Less)
• The Control Plane is the 802.1Q one (MSTP with 1, 3 or 6 trees)• The Management plane set the MSTP's parameters of the Control Plane
CO (as Connection Oriented)• No Control Plane, the forwarding is positioned by the management plane
LB/SP (for CO only) LB (aka Load Balancing)
• The tool tries to optimize (maximize) the load balancing over the full network SP (aka Shortest Path)
• The tool tries to optimize (minimize) the sum of the hops for all flows over the full network
– A route is "acceptable" only if no link is overloaded over the full Network
MU/UN MU (aka Multicast) UN ( aka Unicast)
research & development
France Telecom Group
Variations of the topology for the experiment V0
Reference Topology, aggregation network fully meshed All links 10 G TV dispatcher directly connected to the aggregation Network
V0b Reference Topology , aggregation network fully meshed All DSLAM links 1G ; All other links (except one) of the Aggregation Network at 10G TV dispatcher directly connected to the aggregation Network
V1 Reference Topology, aggregation network fully meshed All links 10 G TV dispatcher outside the aggregation Network (core network)
V1b Reference Topology , aggregation network fully meshed All DSLAM links 1G ; All other links (except one) of the Aggregation Network at 10G TV dispatcher outside the aggregation Network (core network)
V3 Reference Topology but Aggregation network not meshed (Ring Aggregation topology) All DSLAM links 1G ; All other links (except one) of the Aggregation Network at 10G TV dispatcher outside the aggregation Network (core network)
research & development
Partial resultsPartial results
research & development
France Telecom Group
Results criteria and format Three major indicators chosen jointly with ALF (1 curve, two
values) PFD curve
• Probability Density Function– The probability that the load (in term of capacity of the link) is inside a given
interval– The CDF is the integral of the PDF and not used directly to compare results
The ME value• The average of the PDF function
– May give information about the fact that the less loaded links are preferentially chosen
The SD value• The root mean square of the PDF curve
– shows the dispersion of the load of links around the full network
REm (percent) 15.2716
SD (percent) 8614.98
Density 0 < 10% < 20% < 30% < 40% < 50% < 60% < 70% < 80% < 90% < 100%
CDF 1 19 27 32 35 40 40 40 40 40 40
PDF 1 19 7 5 3 5 0 0 0 0 0
V0-Multicast, PDF/ Nb Root
0
5
10
15
20
25
0
< 10%
< 20%
< 30%
< 40%
< 50%
< 60%
< 70%
< 80%
< 90%
< 100%
1 Root
3 Roots
6 Roots
research & development
France Telecom Group
V0
Analyze (to be discuss) Multicast give a non negligible gain against "multi unicast" For MU, no difference between CO and CL For UN, CO seems better than CL
FT Remarks (from detailed results) 3 trees are enough (no specific gain with 6 trees)
REm - V0 - Abstract Of Results
14.5
15
15.5
16
16.5
17
17.5
UNICO UNICL MULTICO MULTICL
SD - V0 - Abstract Of Results
7600
7800
8000
8200
8400
8600
8800
9000
UNICO UNICL MULTICO MULTICL
Fully Meshed Aggregation networkFully Meshed Aggregation networkAll Links 10GAll Links 10GTV inside aggregation NetworkTV inside aggregation Network
research & development
France Telecom Group
V1b
Analyze No "big" difference between CO and CL LB give a better REm and a well better
SD than SP in a CO context PDF curve shows that CO/SP doesn't
find a correct load balancing
Fully Meshed Aggregation networkFully Meshed Aggregation networkAll Links (except one) 10G for the aggregation Networks, All Links (except one) 10G for the aggregation Networks, DSLAM links 1GDSLAM links 1GTV dispatcher outside aggregation NetworkTV dispatcher outside aggregation Network
REm - V1b - Abstract Of Results
16
16.5
1717.5
18
18.5
19
19.520
20.5
21
UNICO UNICL MULTICO MULTICL
0
< 1
0%
< 2
0%
< 3
0%
< 4
0%
< 5
0%
< 6
0%
< 7
0%
< 8
0%
< 9
0%
< 1
00%
> 1
00%
V0COUNISP
V0COMUSP0
2
4
6
8
10
12
14
16
18
PDF for MU & UNI
V0COUNISP
V0COUNILB
V0COMUSP
V0COMULB
0
5
10
15
20
25
V0COUNISP V0COUNILB V0COMUSP V0COMULB
REm CO Uni/MU
0
5000
10000
15000
20000
25000
V0COUNISP V0COUNILB V0COMUSP V0COMULB
SD CO Uni/MU
research & development
Let us try to conclude …Let us try to conclude … ... keeping the door open for discussion... keeping the door open for discussion
research & development
France Telecom Group
Preliminary conclusions
With MSTP, paths are constrained to follow "trees" At first sight, this constraint could decrease load balancing performances
compared to ELS networks But, with a TE tool, MSTP gives the same results as CO networks in terms of
load balancing
For a very loaded network and/or a network with links with different capacities of links
A centralized optimization gives better results (load balancing) than optimization based solely on the calculation of a Shortest Path
The centralized optimization tool gives equivalent performances for CL or CO networks with LB routing ("TE") algorithm
Other "well known" properties of MSTP are not impacted by load balancing optimization
Low cost "Bad" convergence time Compliant to standard Natively multicast…
research & development
France Telecom Group