1 / 17

Revisiting the Optimal Scheduling Problem

Revisiting the Optimal Scheduling Problem. Sastry Kompella 1 , Jeffrey E. Wieselthier 2 , Anthony Ephremides 3 1 Information Technology Division, Naval Research Laboratory, Washington DC 2 Wieselthier Research, Silver Spring, MD

louis
Télécharger la présentation

Revisiting the Optimal Scheduling Problem

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Revisiting the Optimal Scheduling Problem Sastry Kompella1, Jeffrey E. Wieselthier2, Anthony Ephremides3 1 Information Technology Division, Naval Research Laboratory, Washington DC 2 Wieselthier Research, Silver Spring, MD 3 ECE Dept. and Institute for Systems Research, University of Maryland, College Park, MD CISS 2008 – Princeton University, NJ March 2008 ______________________________________________ This work was supported by the Office of Naval Research.

  2. = transmission rate (or “capacity”) Elementary Scheduling Demand: bits (volume) 2 i 1 M Minimize Schedule Length for given demand bits/sec (rate) CISS 2008 2 Princeton University, NJ

  3. Rate: bits/sec Elementary Scheduling (cont…) Volume: bits per frame Maximize total delivery (rate or volume) for given schedule length (sec) LP problems !! CISS 2008 3 Princeton University, NJ

  4. = # of subsets of the set of links ( ) = set of links activated in slot (duration ) Schedule Feasibility of = rate on link i when set is activated. More generally Also an LP !! Past work: Truong, Ephremides Hajek, Sasaki Borbash, Ephremides etc CISS 2008 4 Princeton University, NJ

  5. = channel gain from to link = Transmit Power at More Complicated • Incorporation of the physical layer (through SINR) • Still an LP problem for given ‘s and ‘s • Feasibility criterion on the ‘s • But, may also choose either or or both. CISS 2008 5 Princeton University, NJ

  6. Our Approach: Column Generation • Idea: Selective enumeration • Include only link sets that are part of the optimal solution • Add new link sets at each iteration • Only if it results in performance improvement • Implementation details • Decompose the problem: Master problem and sub-problem • Master problem is LP • Sub-problem is MILP • Optimality • Depends on termination criterion • Finite number of link sets • Complexity: worst case is exponential • Typically much faster CISS 2008 6 Princeton University, NJ

  7. Column Generation • Master Problem: start with a subset of feasible link sets • Sub-problem: generate new feasible link sets • Steps • Initialize Master problem with a feasible solution • Master problem generates cost coefficients (dual multipliers) • Sub-problem uses cost coefficients to generate new link sets • Master problem receives new link sets and updates cost coefficients • Algorithm terminates if can’t find a link set that enables shorter schedule MASTER PROBLEM dual multipliers new link set SUB-PROBLEM (Column Generator) CISS 2008 7 Princeton University, NJ

  8. Master Problem • Restricted form of the original problem • Subset of link sets used; Initialized with a feasible schedule • e.g. TDMA schedule • Schedule updated during every iteration • Solution provides upper bound (UB) to optimal schedule length • Yields cost coefficients for use in sub-problem • Solution to dual of master problem CISS 2008 8 Princeton University, NJ

  9. Sub-problem (1) • How to generate new columns? • Idea based on revised simplex algorithm • Sub-problem receives dual variables from master problem • Sub-problem can compute “reduced costs” based on use of any link set • Sub-problem • Find the matching that provides the most improvement CISS 2008 9 Princeton University, NJ

  10. Sub-problem (2) • Mixed-integer linear programming (MILP) problem • Algorithm Termination • If solution to “MAX” problem provides improved performance • Add this column to master problem • Will improve the objective function • Otherwise, current UB is optimal • If lower bound and upper bound are within a pre-specified value CISS 2008 10 Princeton University, NJ

  11. Extend to “variable transmit power” scenario • Nodes allowed to vary transmit power • Sub-problem generates better matchings by reducing cumulative interference • More links can be active simultaneously • Still a mixed-integer linear programming problem • No additional complexity Sub-problem Constraints Transmission Constraints SINR Constraints CISS 2008 11 Princeton University, NJ

  12. An Example • 6-node network, 8 links • Fixed transmit power: 22% reduction in schedule length compared to TDMA • Variable transmit power: 32% reduction in schedule length compared to TDMA Fixed transmit Power: schedule length = 124.9 s 1 6 3 5 2 4 TDMA schedule = 159.2 s Variable transmit power: schedule length = 108.6 s CISS 2008 12 Princeton University, NJ

  13. 15-node network Schedule length for different instances (sec) Spatial reuse ( = Avg. number of links per matching) CISS 2008 13 Princeton University, NJ

  14. = # of sessions = set of links that originate with node = source node for session = set of links that end with node = destination node for session Introducing Routing Flow Equations: For each session and for each node Written concisely, CISS 2008 14 Princeton University, NJ

  15. Formulation • Multi-path routing between and for each session • Still an LP problem • Column generation still applies CISS 2008 15 Princeton University, NJ

  16. 15-node network Variable transmit Power Fixed transmit Power CISS 2008 16 Princeton University, NJ

  17. Summary & Conclusions • Physical Layer-aware scheduling • LP problem but complex • Solution approach based on column generation works • Decompose the problem into two easier-to-solve problems • Worst-case exponential complexity but much faster in practice • Enumeration of feasible link sets a priori is average-case exponential • Incorporation of Routing • Possibility of Power and Rate control Makes the MAC issue irrelevant !! CISS 2008 17 Princeton University, NJ

More Related