1. Design of Packet-Fair Queuing Schedulers Using a RAM-Based Searching Engine IEEE JSAC, Vol.17, No 6, June 1999 H. Jonathan Chao 외 이 융 yiyung@mmlab.snu.ac.kr

2. Contents PFQ 기존 연구 및 일반적인 구현 기법 General Shaper-Scheduler –Slotted Updates –Implementation Architecture –Time-Stamp Aging Problem Conclusion

3. PFQ(Packet Fair Queuing) PFQ 알고리즘의 전개 방향 –GPS(General Processor Sharing) –WFQ(PGPS) GPS 보다 L max 이상 뒤쳐지지 않는다. –SCFQ, STFQ : Implementation Cost 의 decrease –WF2Q : WFQ 의 단점 보완. GPS 에 가장 가까운 모델. WFI Metric 이용 S(t) < = V(t) check GPS 보다 WF2Q 는 한 패킷 이상 앞서지 않는다는 점까지 추가 –WF2Q+ WF2Q 의 Implementation Cost 의 decrease 본 논문의 주요 고려 대상

4. Time Computation(1/2) Virtual Start Time Virtual Finish Time Virtual Time Implementation of PGPS Where : interval( ) 에 busy 인 session 집합 : weight for session i

5. Time Computation(2/2) WF2Q+ WF2Q –emulates the progress of GPS system WF2Q+ –be computed directly from the packet system

6. Packet Scheduler CPU Packet Search Engine Packet in Packet out Address Write/read Data Memory Packet Scheduler Packets Packet header Packets Packet Scheduler

7. PFQ 기존 연구 및 일반적인 구현 기법 General Shaper-Scheduler –Slotted Updates –Implementation Architecture –Time-Stamp Aging Problem Conclusion

8. Existing Implementation Researches PFQ’s Implementation depends on –system virtual time computation –relative ordering via finish time in a priority queue Efficient hardware-based priority queue –binary tree of comparators –sequencer chip(ASIC) Search-Based version –This paper –new ASIC(PCAM : priority content addressable memory) Author –off-chip memory, hierarchical searching –RSE(RAM-based Search Engine) –commercial memory and FPGA chips

9. Conceptual Framework(1/2) W = max {start time} N : Total number of sessions in the system Start Time V(t) AB

10. Conceptual Framework(2/2) Logical queue per session Head-pointer memory : 각 Queue 의 Header pointer 저장 Tail-pointer memory : 각 Queue 의 Tail pointer 저장 Idle queue : Data memory 의 Idle space 유지

11. Design Issue-I(1/3) Scheduler Queue –Search-based approach Vs. Sorting-based approach Data StructureImplementation of a scheduler queue using the RSE

12. Design Issue-I(2/3) Hierarchical searching with a tree structure –Data structure : Tree of priority encoders & decoders 0 0 0 1 0 0 1 1 0 0 0 1 1 0 01234567 1 0 1 5

13. Design Issue-I(3/3) Output of the m-bit time stamp F in hierarchical searching (L=3) 1 0 1 5 1 0 1

14. Design Issue-II Shaper Queue –Eligibility Test : S(t) <= V(t) –V(t)-> col addr and compare –move eligible packets to scheduler queue F S V(t)

15. Time-stamp Overflow Control Time-stamp overflow –Maximum value of F : M-1 maximum packet length over minimum allocated BW : F: 단조 증가, 한계치 (M-1) 에 도달 -> Overflow –CZ(Current zone bit) 현재 Service 되는 Packet Zone 을 가리킴

16. Time-stamp Overflow Control MSB of F does not change more than once when serving in the current zone -> packet out-of- sequence(x)

17. PFQ 기존 연구 및 일반적인 구현 기법 General Shaper-Scheduler –Slotted Updates –Implementation Architecture –Time-Stamp Aging Problem Conclusion

18. Slotted Updates(1/2) Scheme –Variable sized packet -> fixed-length segments –Packet Scheduler : slotted(synchronous) system –T : one slot time -> W(t,t+d) = dT : normalized to one –All times(start, finish, system time) -> integer –Ex) -> next page Discussion –Incomplete segment underutilized, non-workconserving. limited by one time slot. –Inaccuracy <- discrete time event(arrival, departure)

19. Slotted Updates(2/2)

20. Advantage using Slotted Update 일반적인 Shaper Queue –Maximum N packets are moved to Scheduler Queue Fixed Segment Using Slotted Update –Researches about ATM cells –Same effect as ATM cell(fixed size) –V(t)(=a) 의 값을 가지고 column search –only two packets at most need to be transmitted to the scheduler queue packet with the smallest finish time in column a packet with the smallest finish time in column b(b 는 현재 전송되고 있거나 막 전송한 패킷의 start time)

21. Implementation Architecture

22. Basic Operations At every time slot, CPU( using S min (start time queue)) -> virtual time update(=a) Shaper Queue –new HOL packet : 이전 패킷의 finish time -> CPU –start time, finish time 계산 -> Shaper Queue Scheduler Queue –packet(k) selection(minimum finish time) -> 전송 –F k is stored in the finish time queue( start time : b ) –a 와 b 는 2-D RSE 에서 eligibility test 를 위해서 사용됨

23. 2-D RSE W groups at level 0 to index a column

24. Time-Stamp Overflow(1/2) F and S’s overflow F = S + D –F is bounded by max packet length and min alloc bw –overflow information : 2 bit F < S F > S F < S F > S

25. Time-Stamp Aging Problem(1/2) Finish time 은 다음 패킷의 start time 계산을 위해서 값이 저장 Virtual system time 도 overflow 할 수 있다. – 그 때, finish time 은 obsolete –V(t) 는 한번에 W-1 까지 증가할 수 있다. Purging Algorithm –check each entry and purge all obsolete ones(should be fast) –perform many purging operations in a time slot(difficult) Solution –V(t) can overflow at most once in every time slot –have to see multiple time slot -> Periodic purging scheme

26. Time-Stamp Aging Problem(2/2) Check A entries in T time slots –A >= N-1 -> 2A memory accesses –2T memory accesses(read/wrtie) –T x slot time >= (2T +2A) x memory cycle EX) –64byte packet segment. 10Gbit(51.2ns) –memory cycle 10ns, N = 32K –A = N = 32768 -> T = 21006, A/T = 1.56 purging operations C v (t) : multibit counter, Time Zone Indicator for V(t) O i : obsolete bit counter C i : Time Zone Indicator for F i

27. Flow chart of purging scheme

28. Conclusion A novel RSE for PFQ –hierarchical searching –commercial memory chip, independent of # of session 2-D RSE for general shaper-scheduler –at most two packets to be transferred Time-Stamp overflow –2bit zone bit Time-Stamp Aging Problem –V(t)’s overflow –multibit counter variable within a fixed period