Mixed Stochastic and Event Flows

Transcription

1 Simulation Mixed and Event Flows Modeling for Simulation Dynamics Robert G. Cole 1, George Riley 2, Derya Cansever 3 and William Yurcick 4 1 Johns Hopkins University 2 Georgia Institue of Technology 3 SI International, Inc. 4 University of Texas - Dallas 04 June 2008 / ACM PADS shop 2008

2 Outline Simulation 1 s 2 3 4

3 Scalable Network s Simulation Future network design and modeling requires large scale, high fidelity simulations capability. Training requires real-time speedup of network simulations. Parallelization of network simulations not always useful due to lack of topological communities of interest. analytic/event simulations appear to be an attractive alternative.

4 Simulation Types Simulation Network models: Partitioned models, e.g., packet edge and analytic core Mixed node models, i.e., packet and analytic traffic mixed at each network queue. Analytic models: Deterministic models, e.g., dynamics described by deterministic differential equations. models, e.g., Brownian motion models of queue dynamics.

5 to, Mixed Node External Input Split Network Departures Simulation Loss Merge Simulation of network of queues. Mixing the two fundamentally different traffic types at a single, finite queue. Time dependent models of finite-sized queue dynamics. Splitting and merging mixed-traffic flows within network simulation.

6 Approach to Packet Loss Event Maximum w i (+) w i+1 ( ) =w i+1 (+) w i+2 (+) Simulation Minimum w ( ) i Trajectory t i t i+1 Time t i+2 w i+2 ( ) Packet Loss Event Maximum Minimum w (+) i w (+) i+1 w ( ) i+2 =w (+) i+2 Trajectory w ( ) i Density Sampled Mean Drift w ( ) i+1 t i t i+1 t i+2 Time Contrasting deterministic versus stochastic fluid mixing at queue.

7 Simulation Cumulative Distribution Function for a GI/G/1/K queueing system, F(w, t w 0, t = 0) satisfies the Fokker-Planck Equation [Kobayashi, 1974a], [Kobayashi, 1974b] and [Heyman, 1975] F F = m t w σ2 2 F w 2 (1) where m = λ s µ, and σ 2 = λ s C 2 v + µ C 2 v,µ. The Fokker-Planck equation has an analytic solution F (w, t w 0, t = 0) = α Φ( w w 0 mt σ ) t + β e 2mw/σ2 Φ( w w 0 mt σ ) (2) t Φ( ±w w 0 mt σ ) = 1 Z ±w w0 mt σ t e x2 /2 dx (3) t 2π

8 Boundary Conditions Simulation Initial Conditions: F (w, t = 0) = H(w w 0 ); H(x) is the Heavy-side Function. Upper and lower limits on work in system: Common BCs - Our alternative BCs - lim F (w, t w 0, t = 0) = 0 (4) w 0 lim F (w, t w 0, t = 0) = 1 (5) w w max lim F (w, t w 0, t = 0) = 0 (6) w max[0,w 0 µt] lim w w max F (w, t w 0, t = 0) = 1 (7)

9 Rationale for Our Alternative BCs Simulation Maximum Minimum Initial Condition Maximum Drain Rate t 0 t 1 t 2 t 3 Time Boundary Points Mean Drain Rate Alternative BCs and their impact of density functions versus time. Buffer size sets maximum and minimum limits to work in queue. Maximum drain rate (assuming no arrivals) sets short term time-dependent limit on the minimum work in queue.

10 Simple Packet Foreground Router Interface Simulation Merge Hyper Exponential Background C1 100 Mbps R1 Loss Buffer 500 Kbps Simple simulation model in GTNets. Background is hyper-exponential arrival with deterministic service, modeling UDP packets Foreground is exponential arrival with deterministic service, modeling UDP packets Foreground is later modeled as TCP stream Investigate foreground packet delay (UDP and TCP), loss (UDP and TCP), and goodput (TCP) S1 Sink

11 UDP Delay and Loss Results Simulation Delay (msec) Simulation Deterministic Stoc_mvBC Background rate (Kbps) Loss probability Simulation Deterministic Stoc_mvBC Background rate (Kbps) UDP delay and loss versus rate in mixed traffic-type queue. Deterministic fluid model has no mechanism to allow foreground traffic buffer access at high utilization. model allows foreground traffic access even in overload situations.

12 UDP Delay Results versus C v Simulation Delay (msec) Base(300Kbps) Det(300Kbps) Stoc_mvBC(300Kbps) Delay (msec) Base(400Kbps) Det(400Kbps) Stoc_mvBC(400Kbps) Coefficient of Variation Coefficient of Variation Delay (msec) Delay (msec) Base(430Kbps) 150 Det(430Kbps) Stoc_mvBC(430Kbps) Coefficient of Variation 200 Base(450Kbps) 150 Det(450Kbps) Stoc_mvBC(450Kbps) Coefficient of Variation UDP delay versus C v in mixed traffic-type queue.

13 TCP Goodput Results Simulation Throughput (Bps) Simulation Deterministic Background rate (Kbps) TCP goodput versus rate in mixed traffic-type queue. model matches well with simulation results for TCP dynamics.

14 TCP Goodput versus C v Simulation Goodput (Bps) Goodput (Bps) Base(300Kbps) Det(300Kbps) Stoc_mvBC(300Kbps) Stoc(300Kbps) Base(400Kbps) Det(400Kbps) Stoc_mvBC(400Kbps) Stoc(400Kbps) Base(450Kbps) Det(450Kbps) Stoc_mvBC(450Kbps) Stoc(450Kbps) Coefficient of Variation Base(500Kbps) Det(500Kbps) Stoc_mvBC(500Kbps) Stoc(500Kbps) Coefficient of Variation Goodput (Bps) Goodput (Bps) Coefficient of Variation Coefficient of Variation TCP goodput versus C v in mixed traffic-type queue.

15 Research Simulation Improve Mixed Queue : Not happy with C v 1 results. Possible solution: discretize distribution [Kobayashi, 1974b] or investigate scaling laws for diffusional drift and variance. Develop the background loss models. Possible approach is based upon applications of Bayes Theorem (see below). Develop Network Flow : Only investigated mixing at single node models to date. Leverage literature of network queueing, e.g., [Whitt, 1995], [Kobayashi and Mark, 2002], others.

16 Loss of Fluid Application of Bayes Theorem Simulation Maximum Minimum w i (+) w i ( ) p(w A,t w B =w i+1 ( ),t i+1 ;w i (+),t i ) (w A,t ) t i t t i+1 Time Rt Fluid loss (t t ) = λ p(w max, x)dx t Bayes Equation: P(A B) = P(B A)P(A)/P(B) p(w A, t w B = w ( ) i+1, t i+1; w (+) i w i+1 (+) w B =w i+1 ( ), t i ) = p(w B = w ( ) i+1, t i+1 w A, t )p(w A, t w (+) i, t i ) p(w B = w ( ) i+1, t i+1 w (+) i, t i ) (8)

17 Simulation Initial investigations into mixed, hybrid stochastic simulation models Much work to be done: Improve C v 1 results Develop time-dependent fluid loss models Develop network flow models, i.e. time-dependent network calculus Outlook Initial simulation results are encouraging Need much more development and simulation studies results of time-dependent dynamics

18 References Simulation Cole, R.G., Riley, G., Cansever, D. and W. Yurcik, Process for Packet/Analytic-Based Network, IEEE/ACM Principles of Distributed Simulation (PADS), Rome, Italy, June Heyman, D., A Diffusion Model Approximation for the GI/G/1 Queue in Heavy Traffic, Bell Labs Technical Journal, Vol. 54, No. 9, November Kobayashi, H., Applications of the Diffusion Approximation to Queueing Networks, I. Equilibrium Queue Distributions, Journal of the Association for Computing Machinery, Vol. 21, No. 2, Kobayashi, H., Applications of the Diffusion Approximation to Queueing Networks, I. Nonequilibrium Distributions and Computer Modeling, Journal of the Association for Computing Machinery, Vol. 21, No. 3, H. Kobayashi and B. L. Mark, A Unified Theory for Queuing and Loss-Queueing Networks, IFORS 99 Conference, Beijing, August 16-20, H. Kobayashi and B. L. Mark, Generalized Loss and Queueing-Loss Networks, Int. Trans. in Operational Research, Vol. 9, No. 1, pp , January Whitt, W., Variability Functions for Parametric-Decomposition Approximations of Queueing Networks. Management Science, vol. 41, No. 10, pp , 1995.