A Hybrid Approach to Effort- Limited Fair (ELF) Scheduling for By David Matsumoto June 20 th, 2003

Agenda Introduction Introduction Effort-Limited Fair (ELF) Scheduling Effort-Limited Fair (ELF) Scheduling The Thesis The Thesis Integrate ELF into Integrate ELF into Hybridize ELF Hybridize ELF Simulation setup & results Simulation setup & results Possible future research & recommendations Possible future research & recommendations Conclusion Conclusion

Introduction application transport network data link physical application transport network data link physical application transport network data link physical Applications on a network Applications on a network –Different requirements (bandwidth, timing, reliability) Network stack Network stack –Underlying protocols (TCP, UDP, IP, MAC, etc…) Providing quality of service (QoS) Providing quality of service (QoS) –Wired versus wireless Wireless: subject to capacity loss Wireless: subject to capacity loss “Effort” spent does not always equal “outcome” “Effort” spent does not always equal “outcome”

Earlier Work: Effort-Limited Fair (ELF) Scheduling Motivation Motivation –Applications have poor performance due to link errors –Link errors reduce useful link throughput Solutions introduced: Swapping time slots, Adaptive forward error control Solutions introduced: Swapping time slots, Adaptive forward error control Capacity loss is a fundamental reality for wireless schedulers. Capacity loss is a fundamental reality for wireless schedulers. Questions Questions –Can we support reservations amidst capacity loss? –How should we redistribute the remaining capacity among flows?

Example: 50% Packet Loss Wireless cell capacity: 800 kilobits/second Wireless cell capacity: 800 kilobits/second WFQ scheduler serves two guaranteed flows & two best-effort flows WFQ scheduler serves two guaranteed flows & two best-effort flows Client Expected Rate (kbit/s) Effort-fair rate (kbit/s) Preferable rate (kbit/s) Audio8 4 × 8√8√8√8√ Video × 350 √ FTP FTP

Challenges in QoS Capacity loss should result in throughput loss according to administrative controls Capacity loss should result in throughput loss according to administrative controls What about location dependent errors? What about location dependent errors? –How should we regulate air-time? Conclusion: Conclusion: –We must have a way to balance fidelity with efficiency.

ELF Scheduling – Design Principles Wireless scheduler should be equivalent to wireline scheduler in error-free environment Wireless scheduler should be equivalent to wireline scheduler in error-free environment Capacity loss suffered per flow should be administratively configurable Capacity loss suffered per flow should be administratively configurable Administratively bound capacity lost due to location- dependent errors Administratively bound capacity lost due to location- dependent errors Unless configured otherwise, flows with the same error rate should experience the same capacity loss Unless configured otherwise, flows with the same error rate should experience the same capacity loss Capacity unused by one flow should be distributed “fairly” among other flows Capacity unused by one flow should be distributed “fairly” among other flows

The ELF Scheduler Admission Control Application Protocols Packet Scheduler weights User Requests Administrative Controls “next packet” Link Layer Power factors Effort/outcome

ELF Power factor Main Idea: Main Idea: –Extend transmission time in a controlled fashion –Allot a flow extra effort to meet its reservations, up to some administrative bound  Power factor Adjusted Weight (A i ) Adjusted Weight (A i ) –A i = min (W i /1 – E i, P i X W i ) Effective throughput (T i ) Effective throughput (T i ) –T i = ((A i /∑ j A j ) X B) x (1 – E i ) This effectively balances fidelity with efficiency. This effectively balances fidelity with efficiency.

The Thesis It is possible to integrate an Effort-Limited Fair scheduler (ELF) into a standard implementation, and thus ensure sensible outcomes for flows in response to unrecoverable capacity loss. Moreover, it is possible to improve on the efficiency of ELF within an MAC by using a hybrid approach to regulate stations within both a DCF & a PCF. Moreover, it is possible to improve on the efficiency of ELF within an MAC by using a hybrid approach to regulate stations within both a DCF & a PCF.

IEEE Standard for wireless Local Area Networks (LAN) Standard for wireless Local Area Networks (LAN) Distributed Coordinator Function (DCF) Distributed Coordinator Function (DCF) –Uses Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) Point Coordinator Function (PCF) Point Coordinator Function (PCF) –Centralized, polling-based mechanism requiring a base station (Point Coordinator) DCF & PCF can co-exist DCF & PCF can co-exist Contention period (CP) under control of DCF Contention period (CP) under control of DCF Contention-free period (CFP) under control of PCF Contention-free period (CFP) under control of PCF Both methods use explicit Acks Both methods use explicit Acks –Useful for PC tracking outcome

PCF/DCF Alternation

Integration of ELF scheduling into Core ELF algorithm remains the same Core ELF algorithm remains the same –Choose “most deserving” flow –Flow is allocated effort as a function of its power factor Must have “effort” allocated to be chosen Must have “effort” allocated to be chosen –Separate best-effort flows from guaranteed flows Update each flow per clock “tick” Update each flow per clock “tick” ELF scheduling becomes the policy for the PCF ELF scheduling becomes the policy for the PCF –Departs from existing protocol specifications Records “outcome” via returned Acks (or NULL frames) Records “outcome” via returned Acks (or NULL frames) Charge a station for “effort” each time it is polled. Charge a station for “effort” each time it is polled. –Station charged even if packets are corrupted.

Extending ELF for DCF Motivation Motivation –DCF is more efficient in general, when traffic load isn’t high. –Ideal: PCF could interject only when necessary to bring flows into conformance with reservations. Design Principles Design Principles –DCF must remain completely distributed –ELF-DCF should adhere to original ELF design principles. –Integration of ELF-DCF/PCF should be seamless

ELF-DCF: Implementation Design Continue using both “deserve” & “effort” to (self) regulate flows Continue using both “deserve” & “effort” to (self) regulate flows “Deserve” carries over from CFP “Deserve” carries over from CFP “Effort” allocated based on past CP history “Effort” allocated based on past CP history –Cannot force station to expend effort –Make use of power factor –“Effort” is conserved (unused effort plugged into CFP) No distinction between best- effort versus guaranteed flows No distinction between best- effort versus guaranteed flows –Reduce overhead in beacon ELF Scheduler (PC) Mobile Nodes Send Beacon with ELF data Collect statistics /Start CFP

Implementation for ns-2 Build on top of CMU Monarch group’s work Build on top of CMU Monarch group’s work –Basic mobile node functionality with DCF Add semi-complete implementation of PCF. Add semi-complete implementation of PCF. –Beacon timer –Bi-directional polling –Poll list – integrated with ELF Implement Weighted Round Robin (WRR) instead of WFQ Implement Weighted Round Robin (WRR) instead of WFQ –Measure throughput in frame slots (done for simplicity) –Not concerned with complex based service characterizations (e.g. delay, jitter guarantees) Administratively configure breakdown of superframe Administratively configure breakdown of superframe –I.E. Percentage CFP/CP Assume admissions control module exists that can set appropriate per-flow power factors. Assume admissions control module exists that can set appropriate per-flow power factors.

Experimental setup One flow per station One flow per station –Two CBR flows & two FTP flows (each with 25% weight) Flows chosen so aggregate throughput consumes the entire link ~ 1.4 Mbit/s Flows chosen so aggregate throughput consumes the entire link ~ 1.4 Mbit/s –CBR flow ~ 450 Kbit/s –FTP flow ~ 250 Kbits/s (no errors) Relevant variables: Relevant variables: –Percentage CFP/CP –ELF-DCF Boolean –Power factor –Error rate Error models Error models –Uniform, Markov, real-world Traces WLAN

Uniform – 50% (CBR-1 Error prone) CBR-1 performs poorly while other stations operate normally No ELF intervention here.

Uniform – 50% (CBR-1 Error prone) ELF-PCF effectively restores throughput to CBR-1 at the expense of best-effort flows

Uniform – 50% (CBR-1 Error prone) ELF-PCF is only partially effective at restoring throughput

Uniform – 50% (CBR-1 Error prone) ELF-DCF effectively works with ELF-PCF to meet ELF design principles

Walls trace – (CBR-1 Error prone) No ELF intervention – other flows gain throughput as CBR-1 decreases

Walls trace – (CBR-1 Error prone) ELF-PCF/DCF work together Other flows share throughput loss, up to administrative bound Capping (“deserve”) mechanism needed

Future Work & Recommendations Future work Future work –Provide a capping mechanism for flows that accumulate large “deserve” values –Can we “encourage” stations to spend effort during a DCF (rather than just limiting the effort allotted)? Recommendations Recommendations –Incorporate a “policing” mechanism into beacons –Change specifications on how stations are polled –Allow polling of individual flows

Conclusions ELF scheduling can be implemented as the policy for a PCF in ELF scheduling can be implemented as the policy for a PCF in ELF-DCF/PCF provides a hybrid mechanism through we can achieve ELF’s design principles ELF-DCF/PCF provides a hybrid mechanism through we can achieve ELF’s design principles

Peter Steenkiste – Advisor Peter Steenkiste – Advisor David Eckhardt – Reader David Eckhardt – Reader INI Staff & Professors INI Staff & Professors Acknowledgements

Questions