Weighted_ECMP_Slides
Weighted ECMP for IP Network
1
Motivation
•
Equal-Cost Multipath (ECMP) routing equally split traffic among the multiple paths [4]. However, evenly splitting traffic among the paths does not achieve optimal load balancing
0.5
0.5
S1
0.5
D1
0.5
3
0.5
1
4
0.5
0.5
0.5
5
•
2
0.5
0.5
S2
1
D2
6
Link between node 3 and node 4 becomes a bottleneck link.
2
Proposal
• Weighted ECMP
– distributes traffic among available equal cost paths based on a set of predetermined ratios.
– While most existing work tries to minimize the traffic load on the most utilized link, we develop a model to optimize the end-to-end delay
– present a heuristic algorithm to obtain the nearoptimal weight configuration.
3
Problem Formulation
•
The network is represented by a connected graph G(V,E) with node set V and directed edge set E. Let’s denote the notations as follows:
•
Link delay
•
Path delay
4
Problem Formulation
•
Since optimizing the delay of path with higher traffic load would give us more performance gain. We assign path load hdαdp as a weight to each path. Total weighted end-to-end delay for demand d.
•
Our objective is to minimize overall end-to-end delay in the network.
•
Since
•
After transformation, our objective turns out to be minimizing average packet delay.
5
Problem Formulation
•
Let Fe(ye) denote ye /(ce − ye). The problem can be formulated as follows:
6
Heuristic Algorithm Description
•
•
•
•
State S: determined by split ratios for multi-path of each source and destination pair in the network.
Energy function E: used for state evaluation.
Temperature T: the remaining iterations before termination. T0 is the initial value of T.
Acceptance probability P: probability of transition from current state S with energy E = E(S) to neighbor state Sn with energy En = E(Sn).
7
Heuristic Algorithm Description
•
1
Motivation
•
Equal-Cost Multipath (ECMP) routing equally split traffic among the multiple paths [4]. However, evenly splitting traffic among the paths does not achieve optimal load balancing
0.5
0.5
S1
0.5
D1
0.5
3
0.5
1
4
0.5
0.5
0.5
5
•
2
0.5
0.5
S2
1
D2
6
Link between node 3 and node 4 becomes a bottleneck link.
2
Proposal
• Weighted ECMP
– distributes traffic among available equal cost paths based on a set of predetermined ratios.
– While most existing work tries to minimize the traffic load on the most utilized link, we develop a model to optimize the end-to-end delay
– present a heuristic algorithm to obtain the nearoptimal weight configuration.
3
Problem Formulation
•
The network is represented by a connected graph G(V,E) with node set V and directed edge set E. Let’s denote the notations as follows:
•
Link delay
•
Path delay
4
Problem Formulation
•
Since optimizing the delay of path with higher traffic load would give us more performance gain. We assign path load hdαdp as a weight to each path. Total weighted end-to-end delay for demand d.
•
Our objective is to minimize overall end-to-end delay in the network.
•
Since
•
After transformation, our objective turns out to be minimizing average packet delay.
5
Problem Formulation
•
Let Fe(ye) denote ye /(ce − ye). The problem can be formulated as follows:
6
Heuristic Algorithm Description
•
•
•
•
State S: determined by split ratios for multi-path of each source and destination pair in the network.
Energy function E: used for state evaluation.
Temperature T: the remaining iterations before termination. T0 is the initial value of T.
Acceptance probability P: probability of transition from current state S with energy E = E(S) to neighbor state Sn with energy En = E(Sn).
7
Heuristic Algorithm Description
•