Dynamic Window-Constrained Scheduling for Multimedia Applications - PowerPoint PPT Presentation

Dynamic Window-Constrained Scheduling for Multimedia Applications Richard West and Karsten Schwan Georgia Institute of Technology Rich West (1999)

Introduction ■ Real-Time media servers need to support 100s (even 1000s) of clients with individual RT (QoS) constraints. ■ Need fast/efficient scheduling on such servers. ■ We describe Dynamic Window-Constrained Scheduling (DWCS): ■ DWCS limits the number of late packets over finite windows of arrivals requiring service. ■ DWCS can be both fair and unfair when necessary - Performs Fair-Queueing, SP and EDF. ■ We demonstrate DWCS using a streaming video application. Rich West (1999)

DWCS Packet Scheduling ■ Two attributes per packet: ■ Deadline (max inter-packet gap). ■ Loss-tolerance, x/y . ■ x late/lost packets every y arrivals for service from same stream. ■ At any time, all packets in the same stream: ■ Have the same current loss-tolerance. ■ Have deadlines offset by a fixed amount from predecessors. Rich West (1999)

DWCS - Conceptual View Network Pipe Higher Priority . = . Lower . Loss-Tolerance EDF-ordered queues Rich West (1999)

Heterogeneous Scheduling 0 loss-tolerance queue EDF-ordered queues Network Pipe Higher Priority . = . Lower . Loss-Tolerance (Time-constrained traffic) Compare first packet with packet at head of 0 loss-tolerance queue to see if schedulable. Lowest loss-tolerance first (Non-time-constrained traffic) Rich West (1999)

Pairwise Packet Ordering Table Precedence amongs t pairs of packets • Lowest loss-tolerance first • Same non-zero loss-tolerance, order EDF • Same non-zero loss-tolerance & deadlines, order lowest loss-numerator first • Zero loss-tolerance and denominators, order EDF • Zero loss-tolerance, order highest loss- denominator first • All other cases: first-come-first-serve Rich West (1999)

Example: L1=1/2, L2=3/4, L3=6/8 D=1, Service Time (C)=1 s1 s2 s1 s3 s s2 s1 s3 s1 s2 s1 s3 s1 s2 s s3 1 1 time, t 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 s1 1/2(0), 1/1(1),1/2(2),1/1(3),1/2(4)... s 3/4(0),2/3(1),2/2(2),1/1(3),3/4(4),2/3(5),2/2(6),1/1(7),3/4(8)... 2 s3 6/8(0),5/7(1),4/6(2),3/5(3),3/4(4),2/3(5),1/2(6),0/1(7),6/8(8)... Rich West (1999)

Example: L1=1/2, L2=1/2, C1=5, C2=3, D1=5, D2=3 s1 s2 s2 s1 time, t 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 s2 s2 s1 s2 time, t 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 s1 1/2(0),1/1(5),1/2(10),0/1(15),1/2(20),0/1(25),1/2(30)... s2 1/2(0),0/1(3),1/4(6),1/3(9),1/2(12),0/1(15),1/4(18), 1/3(21),1/2(24),0/1(27),1/2(30)... Rich West (1999)

Loss-Tolerance Adjustment (A) ■ For stream i whose head packet is serviced before its deadline: ■ if ( y i ’ > x i ’ ) then y i ’ = y i ’ -1; ■ if ( x i ’ = y i ’ =0) then x i ’ = x i ; y i ’ = y i ; ■ Where: ■ x i =Original loss-numerator for stream i ■ y i =Original loss-denominator for stream i ■ x i ’ =Current loss-numerator for stream i ■ y i ’ =Current loss-denominator for stream i Rich West (1999)

Loss-Tolerance Adjustment (B) ■ For stream j whose head packet misses its deadline: ■ if ( x j ’ > 0) then ■ x j ’ = x j ’ -1; y j ’ = y j ’ -1; ■ if ( x j ’ = y j ’ =0) then x j ’ = x j ; y j ’ = y j ; ■ else if ( x j ’ =0) and ( y j > 0) then ■ x j ’=2 x j -1; y j ’ =2 y j +( y j ’ -1); ( method 1 ) ■ x j ’ = x j ; y j ’ = y j ; ( method 2 ) ■ if ( x j > 0) then y j ’ = y j ’ + ( y j - x j )/ x j  ; ( method 3 ) ■ if ( x j =0) then y j ’ = y j ’ + y j ; Rich West (1999)

DWCS Algorithm Outline ■ While TRUE : ■ Find stream i with highest priority (see Table) ■ Service packet at head of stream i ■ Adjust loss-tolerance for i according to (A) ■ Deadline( i ) = Deadline( i ) + Inter-Pkt Gap( i ) ■ For each stream j missing its deadline: ■ While deadline is missed: ■ Adjust loss-tolerance for j according to (B) ■ Drop head packet of stream j if droppable ■ Deadline( j ) = Deadline( j ) + Inter-Pkt Gap( j ) Rich West (1999)

Video Server MPEG-1 active stream s1 c 1 s2 ATM c Scheduler 2 Network c sn n Exponential idle time VIDEO SERVER Rich West (1999)

Fair Scheduling: W=1,2 L=2/3,1/3 (b) 1.1e+06 DWCS (s1) DWCS (s2) 1e+06 SFQ (s1) SFQ (s2) 900000 800000 Bandwidth (bps) 700000 600000 DWCS (s2) & SFQ (s2) 500000 400000 DWCS (s1) & SFQ (s1) 300000 200000 0 20000 40000 60000 80000 100000 120000 140000 160000 Time (mS) Rich West (1999)

Fair Scheduling: W=1,1,2,4 L=7/8,14/16,6/8,4/8 (a) 1.4e+06 DWCS (s1) DWCS (s2) DWCS (s3) 1.2e+06 DWCS (s4) SFQ (s1) SFQ (s2) 1e+06 SFQ (s3) SFQ (s4) Bandwidth (bps) DWCS (s4) & SFQ (s4) 800000 600000 DWCS (s1,s2) & SFQ (s1,s2) 400000 DWCS (s3) & SFQ (s3) 200000 0 0 5000 10000 15000 20000 25000 30000 35000 40000 45000 50000 Time (mS) Rich West (1999)

Mixed Traffic: L1=1/3,L2=2/3, L3=0/100,D1=1,D2=1,D3= ∞ (a) 1.4e+06 s1 s2 s3 1.2e+06 1e+06 Bandwidth (bps) 800000 600000 400000 200000 0 0 50000 100000 150000 200000 250000 Time (mS) Rich West (1999)

Mixed Traffic: L1=1/3,L2=2/3, L3=0/1500,D1=1,D2=1,D3= ∞ (b) 1.2e+06 s1 s2 s3 1e+06 800000 Bandwidth (bps) 600000 400000 200000 0 0 50000 100000 150000 200000 250000 Time (mS) Rich West (1999)

Mixed Traffic: L1=1/3,L2=2/3, L3=0/100,D1=1,D2=1,D3= ∞ (a) 160000 s1 s2 140000 s3 120000 100000 Delay (mS) 80000 60000 40000 20000 0 0 200 400 600 800 1000 1200 1400 1600 Number of Packets Sent Rich West (1999)

Mixed Traffic: L1=1/3,L2=2/3, L3=0/1500,D1=1,D2=1,D3= ∞ (b) 120000 s1 s2 100000 s3 80000 Delay (mS) 60000 40000 20000 0 0 200 400 600 800 1000 1200 1400 1600 Number of Packets Sent Rich West (1999)

Missed Deadlines: W=1,2 L=2/3,1/3 (a) 1600 DWCS (s1) DWCS (s2) 1400 SFQ (s1) SFQ (s2) 1200 Deadlines Missed 1000 DWCS (s1) & SFQ (s1) 800 600 400 DWCS (s2) & SFQ (s2) 200 0 0 20000 40000 60000 80000 100000 120000 140000 160000 Time (mS) Rich West (1999)

Mean Packet Delay: W=1,2 L=2/3,1/3 (b) 60000 DWCS (s1) DWCS (s2) SFQ (s1) 50000 SFQ (s2) 40000 Delay (mS) 30000 DWCS (s1) & SFQ (s1) 20000 SFQ (s2) 10000 DWCS (s2) 0 0 200 400 600 800 1000 1200 1400 1600 Number of Packets Sent Rich West (1999)

DWCS Summary ■ Aimed at servicing packets with delay and loss- constraints. ■ Attempts to service each stream so that at most x packets are lost/late for every y packets requiring service. ■ DWCS minimizes the number of consecutive late packets over any finite window of packets in a given stream. ■ DWCS can perform fair scheduling, SP, and EDF scheduling. ( It can be unfair when necessary ). Rich West (1999)

DWCS - Current Work ■ DWCS is currently being adapted for use as a CPU scheduler (using Linux), for hard real-time threads, so that (y-x) out of y deadlines can be met. ■ Leads to bounded service delay, and guaranteed service in any finite window of service time. ■ Aim is to support coordinated thread/packet scheduling. Rich West (1999)

Scheduling Related Work ■ Fair Scheduling : WFQ/WF 2 Q (Shenker, Keshav, Bennett, Zhang etc), SFQ (Goyal et al), EEVDF/Proportional Share (Stoica, Jeffay et al). ■ (m,k) Deadline Scheduling : Distance-Based Priority (Hamdaoui & Ramanathan), Dual-Priority Scheduling (Bernat & Burns). Rich West (1999)