1. 1 Weighted Fair Queueing GPS PGPS SCFQ Implementation

2. 2 Round-robin zproblems ypackets lengths are not the same  not fair a eb cd a e b cd (a)round-robin (b) weight round-robin

3. 3 Generalized Processor Sharing (GPS): (A. Parekh, 1992) zassumptions ytraffic is infinitely divisible. yserve multiple sessions simultaneously based on the ratio of pre- defined weights. zproperties yfairness yminimum throughput guarantee yanalyzable with leaky bucket admission control

4. 4 znotations: zan example of GPS

5. 5 zminimum throughput guarantee ySumming over all session j, service rate=r:

6. 6 PGPS : (A. Parakh, 1992) zPGPS (packet-by-packet scheme) : an approximation to GPS for operating in real world yideally, PGPS wants to serve packets in the increasing departure order of GPS yan example of different between GPS and PGPS e.g. Fp=packet p's departure time under GPS session 1 session 2 packet i packet j FiFi FjFj ideally, serve packet j first. But at time t, there is only packet, i, in the system. Have to serve i first GPS t

7. 7 zproperties of PGPS ydeparture time of the two scheme Fp ^ ^ Lmax/rLmax(N-1)/r (GPS)(PGPS) F1=F2=F3 F1-F1=L max (3-1)/r ^ GPS PGPS F1F2F3 ^^^ L max /r exact low bound not exist

8. 8 zproperties (continued)

9. 9 zHard to keep track of departure time (GPS) zimplementation of PGPS (virtual time) yuse hypothetical ideal fluid flow model as reference yuse virtual time, v(t), to represent the progress of work in the reference system. yexample: F1 a1 a2

10. 10 real time, t virtual time, v(t) Real systemreference system the same departure order t v(t) 268 4 8 12 26848 s1s2V(s1)V(s2)

11. 11 zPGPS algorithm yvirtual time finishing times can be determined at the packet arrival time ypackets served in order of virtual time finishing time zproblems ynot feasible yway of assigning weight is not defined

12. 12 zpicking up

13. 13 Self Clocked Fair Queueing: J. Golestani(1994) zMotivation: eliminate the need for the hypothetical fluid-flow reference system. zApproach: different notion of virtual time, which depends on the progress of work in the actual packet-based system.

14. 14 zSCFQ algorithm yeach arriving packet is tagged with a service tag. yservice tags are iteratively computed as yvirtual time is defined equal to the service tag of the packet receiving service. ytransmit packets in the increasing order of service tags.

15. 15 zProperties of SCFQ yImplementable yfairness property yThe end-to-end session delay bounds are comparable to that of PGPS with leaky bucket admission policy. zProblems yWay of determining value is still not defined.

16. 16 Efficient FQ Architectures (info 96) zMotivation: ydevelop efficient scheduling algorithms for high-speed ATM network zContents: ysorting scheme yhierarchical approach for wide range of services

17. 17 zSCFQ yTo simplify implementation, calculate the service tag until the cell reaches the head of its connection queue.

18. 18 zObservation: yfor session k,

19. 19 zArchitecture yif the ith bin is receiving service, put the new head-of line cell to the th bin ylogic to transmit a cell and locate the next non-empty sorting bin is required. zproblem : ya large number of sorting bins are required to handle a wide range of bandwidth parameters... connection FIFOs sorting bins 0 1

20. 20 zHierarchical Fair Queueing (two-level scheduler) yfirst stage handle connections with similar rates ysecond stage serves a small number of groups yGroup weights SCFQ