Stochastic Processing Networks: What, Why and How? Ruth J. Williams University of California, San Diego

Transcription

1 Stochastic Processing Networks: What, Why and How? Ruth J. Williams University of California, San Diego 1

2 OUTLINE! What is a Stochastic Processing Network?! Applications! Questions! A Simple Example! Approximations! Perspective! Two Motivating Examples! Next Two Lectures 2

3 Stochastic Processing Networks (cf. Harrison 00) I buffers (classes) J activities K servers (resources) An activity consumes from certain classes, produces for certain (possibly different) classes, and uses certain servers. 3

4 Stochastic Processing Networks SPN Activities are Very General Queueing network Flexible servers, alternate routing Simultaneous actions 4

5 Semiconductor Wafer Fab: P. R. Kumar 5

6 Multiclass Queueing Network 6

7 Call Center: First Direct (branchless retail banking) Larreche et al., INSEAD 97 (see also Gans, Koole, Mandelbaum 93) 7

8 Differentiated Service Center (Parallel server system, alternate routing) 8

9 NxN Input Queued Packet Switch: Prabhakar L 11 (n) A 1 (n) A 11 (n) 1 1 D 1 (n) A 1N (n) A N (n) A N1 (n) D N (n) A NN (n) N N L NN (n) 9

10 2x2 Input Queued Packet Switch 10

11 Stochastic Processing Networks! APPLICATIONS Complex manufacturing, telecommunications, computer systems, service networks! FEATURES Multiclass, service discipline, alternate routing, complex feedback, heavily loaded! PERFORMANCE MEASURES Queuelength, workload and server idletime 11

12 QUESTIONS! STABILITY! PERFORMANCE ANALYSIS (when heavily loaded)! CONTROL (involves performance analysis for good controls) 12

13 A SIMPLE EXAMPLE: SINGLE SERVER QUEUE 13

14 M/M/1 Queue m! Poisson arrivals at rate (independent of service times)! i.i.d. exponential service times mean! FIFO order of service, infinite buffer m 14

15 M/M/1 Queue m! Poisson arrivals at rate (independent of service times)! i.i.d. exponential service times mean! FIFO order of service, infinite buffer m! Traffic intensity ρ = m 15

16 M/M/1 Queue! Poisson arrivals at rate (independent of service times)! i.i.d. exponential service times mean! FIFO order of service, infinite buffer! Traffic intensity! Queuelength is a birth-death process (Markov) m ρ = m m 16

17 M/M/1 Queue! Poisson arrivals at rate (independent of service times)! i.i.d. exponential service times mean! FIFO order of service, infinite buffer! Traffic intensity! Queuelength is a birth-death process (Markov) m ρ = m! Positive recurrent (stable) iff ρ <1 m 17

18 M/M/1 Queue m! Poisson arrivals at rate (independent of service times)! i.i.d. exponential service times mean! FIFO order of service, infinite buffer m! Traffic intensity ρ = m! Queuelength is a birth-death process (Markov)! Positive recurrent (stable) iff! Stationary distribution! Mean steady-state queuelength ρ <1 i πi = ρ (1 ρ), i = 0,1, 2, L = ρ /(1 ρ) 18

19 M/M/1 Queue m! Poisson arrivals at rate (independent of service times)! i.i.d. exponential service times mean! FIFO order of service, infinite buffer m! Traffic intensity ρ = m! Queuelength is a birth-death process (Markov)! Positive recurrent (stable) iff! Stationary distribution! Mean steady-state queuelength ρ <1 i πi = ρ (1 ρ), i = 0,1, 2, L= ρ /(1 ρ) = W 19

20 M/GI/1 Queue 2 m, σ s 20

21 M/GI/1 Queue 2 m, σ s! Mean steady-state queuelength L = ρ + ρ + σ s 2(1 ρ) (Pollaczek-Khintchine) 21

22 GI/GI/1 Queue (+mild reg. assumptions) 2, σ a 2 m, σ s 22

23 GI/GI/1 Queue (+mild reg. assumptions) 2, σ a 2 m, σ s ( a s) (1 ) L σ + ρ σ 2 for ρ " 1 (Smith 53, Kingman, 61) 23

24 M/M/1 Queue (Simulation of Dynamics) m ρ = =

25 M/M/1 Queue (Simulation of Dynamics) m ρ = =

26 M/M/1 Queue (Simulation of Dynamics) m ρ = =

27 GI/GI/1 Queue (Dynamics) 2, σ a 2 m, σ s Q(t) = queuelength at time t Theorem (Iglehart-Whitt 70): For 2 * (1 ρ) Q( /(1 ρ) ) Q ( ) where is a one-dimensional reflecting Brownian motion m 1 ρ 1, with drift and variance parameter Q * () σ + m σ a s 27

28 APPROXIMATE DYNAMIC MODELS! Most SPNs cannot be analyzed exactly! Consider approximate models (valid under some scaling limit, e.g., heavily loaded, many sources, many servers, large networks)! Two main classes of approximate models: Fluid models (functional law of large numbers) Diffusion models (functional central limit theorem) 28

29 ANSWERS (OPEN MULTICLASS HL QUEUEING NETWORKS) Last years: development of a theory for establishing stability and heavy traffic diffusion approximations for open multiclass queueing networks with HL (head-of-the-line) service disciplines. HL: service allocated to a buffer goes to the job at the head-of-the-line (jobs within buffers are in FIFO order). 29

30 PERSPECTIVE MQN SPN Sufficient conditions for e.g., parallel server system, HL stability and diffusion packet switch approximations Non- e.g., LIFO, Processor Sharing e.g., Internet congestion HL (single station, control / bandwidth sharing PS: network stability) model 30

31 MOTIVATING EXAMPLES Stability Performance Control 31

32 Two-Station Priority Queueing Network (Rybko-Stolyar 92) 3 = 1 m 4 m 3 1 = 1 m1 m2 32

33 Two-Station Priority Queueing Network (Rybko-Stolyar 92) 3 = 1 m 4 m 3 1 = 1 m1 m2 Poisson arrivals at rate 1 to buffers 1 and 3 m i Exponential service times: mean rate of service for buffer i Preemptive resume priority: * denotes high priority classes 33

34 Two-Station Priority Queueing Network (Rybko-Stolyar 92) 3 = 1 m 4 m 3 1 = 1 Poisson arrivals at rate 1 to buffers 1 and 3 Exponential service times: mean rate of service for buffer i Preemptive resume priority: * denotes high priority classes Simulation: Traffic intensities: m1 m2 m i m = m = 0.33, m = m = ρ = m + m = 0.99 ρ = m + m =

35 Two-Station Priority Queueing Network (Rybko-Stolyar 92) --- Server 1 (sum of queues 1 & 4) --- Server 2 (sum of queues 2 & 3) 35

36 Parallel Server System m2 m3 m1 m 4 = 0.05, = 1.2, = m = 0.5, m = 1, m = 1, m =

37 Parallel Server System Simulation with static priority discipline: server 1 gives priority to buffer 1, server 2 gives priority to buffer 2 Queuelengths for buffer 1 ---, buffer 2 ---, buffer versus time 37

38 Parallel Server System Simulation with dynamic priority discipline: server 1 gives priority to buffer 1, server 2 gives priority to buffer 2, except when queue 2 goes below threshold of size 10 Queuelengths for buffer 1 ---, buffer 2 ---, buffer versus time 38

39 NEXT TWO LECTURES! Open Multiclass HL Queueing Networks: Stability and Performance! Control of Stochastic Processing Networks: Some Theory and Examples 39

40 40