Stable Load Control with Load Prediction in Multipath

Stable Load Control with Load Prediction in Multipath Packet Forwarding Il. Kyu Park, Youngseok Lee, and Yanghee Choi Proc. 15 th IEEE Int’l conf. on Information Networking

Introduction(1/2) Multi Protocol Label Switching(MPLS) n n n Explicit Routing(ER) Traffic Engineering(TE) Constrained-based Routing(CBR)

Introduction(2/2) Multipath forwarding n n n Multipath establishment Load distribution Link-state acquirement

Multipath forwarding with dynamic load control(1/2) Multipath establishment Forwarding mechanism with ER n ECMP w Equal Cost Multi. Path n MPLS w Arbitrary path forwarding w Fast path forwarding

Multipath forwarding with dynamic load control(2/2) Traffic Mapping n n Packet level v. s. flow level Hashing-based packet mapping Exchange of Link-state information n Load (received and dropped packets) Residual bandwidth QOSPF

Load Control with Dynamic Network State The goal is to minimize the number of congested links and idle links. Features of the distributed approach n n Lower control overhead Better scalability Inaccuracy of Link-state Instability of link load

Instability of Dynamic Load Control The exchange of link-state is required for load distribution. n Overhead v. s. Accuracy Global synchronization n Link oscillate between underutilized state and congested state.

Stable Load Control with Prediction Load prediction is required to eliminate load oscillation. n n n Observed increasing speed of link load is used to estimate present link load. Threshold T is deployed for deciding to increase/decrease load on links. Source traffic change can not be detected but diminish by extending link-state exchange period.

The Proposed Algorithm PL: Previous load CL: Current load EL: Estimated load

The Proposed Algorithm For each path with load higher than T, its resulting path load ratio is set to (current load ratio)*T/L. Total residual bandwidth is checked if it can accommodate the excess load. The excess load is added to the path with load lower than T by the factor of (T-CL)/(EL-CL).

Simulation Environment 6 CBR sources 4 LSPs α=0. 3 T=80%

Deviation of Load between Links Prediction -> deviation. Load+ -> deviation-

Dropping Rate Prediction -> loss rate. Load+ -> comparable

Discussions Issues n n n Management of multipath forwarding Accurate Prediction of link load Handling imprecision of state information Related Solutions n n Active negotiation Controlling load locally by routers’ own Exchanging link-state conditionally Two-level link-state aggregation