Graded Channel Reservation with Path Switching in Ultra High Capacity Networks Reuven Cohen, Niloofar Fazlollahi, David Starobinski ECE Dept., Boston University Gridnets Workshop 2006 San Jose, CA

Acknowledgements US Department of Energy Dr. N. Rao, ORNL

Outlines Advanced Channel Reservation Contributions Models and Algorithms Related work Variants Performance Evaluation Conclusion

Motivation Grid computation : Large Hadron Collider (LHC) at CERN Exabytes (10 18 bytes) → need for new protocol to support huge file transfers

Advanced Channel Reservation Emergence of new protocol stack complements TCP/IP Properties: 1 - Users request resources for connection in advance (bandwidth/duration, file size) 2 – Dedicated resources allocated by a scheduler (centralized or replicated) 3 – Implemented directly on top of layer 2

UltraScience Net

ACR Challenges Scheduling Routing Goal: maximum utilization of resources EarliestShortest A B

Graded Channel Reservation (GCR) Contributions:  path grading multi-criteria path optimization (shortest, earliest)  path switching connection can switch between paths  Complexity analysis (small polynomial)  Performance evaluation

Model Model: G (V,E) V: {A, B, C} E: {AB, AC, BC} requests: response: (time,path) Objective: Highest grade path A B C Source Destination Bandwidth Duration

Grading Example Primary grading criterion: earliest path Secondary grading criterion: Shortest Widest

Example (Cont.) Thm: GCR always returns the earliest time at which a path satisfying requested bandwidth B and duration T can be established between nodes s and d. Return path with highest grade (e.g., earliest-shortest)

GCR Algorithm:  Time slots: connection set up/ tear down  Steady state residual graph  Graph intersection  yes - highest grade path no – start from next slot  Reserve bandwidths BFS path search

Related Work Most closely related: Guerin & Orda, INFOCOM, 2000 Rao, Wing, Carter & Wu, IEEE ComSoc Mag., 2005 Focus on single criterion optimization No path switching Limited performance evaluation

Advantage of path switching A C Slot 1: [1,1:30] Slot 3: [4,8] Slot 2: [1:30,4] Request: (A,C) at time 2:00 pm, duration = 4 hours B A C B A C B A C B Slot 3: [2,4]Slot 2: [1:30,2] A C B Slot 5: [6,8]Slot 4: [4,6]

Variants to GCR GCR switch Switch to best grade path available at each slot

Reducing Path Switches GCR minimum Thm: GCR minimum returns the earliest path and minimizes number of path switches. GCR limitx Heuristic: limits up to x switches

Simulation measures & parameters Performance measures: 1- average delay 2- saturation throughput Requests: (s,d,B,T) Parameters:  uniform source  uniform or hot-spot destination  uniform or 80/20 bandwidth  exponential or heavy-tailed connection length

Topologies

Performance Evaluations

Conclusion Framework: grading & switching 1 st and 2 nd path optimization important Path switching widely improves performance

Future Work Time window: simulating blocking probability Cost of switching