predicting network performance characteristics
play

Predicting Network Performance Characteristics - Introducing Flow - PowerPoint PPT Presentation

Predicting Network Performance Characteristics - Introducing Flow Field Forecasting - Michael Frey Kyle Caudle Bucknell University South Dakota School mfrey@bucknell.edu of Mines & Technology e-Weather Center Project DOE ASCR Program


  1. Predicting Network Performance Characteristics - Introducing Flow Field Forecasting - Michael Frey Kyle Caudle Bucknell University South Dakota School mfrey@bucknell.edu of Mines & Technology e-Weather Center Project DOE ASCR Program - Phil DeMar - Thomas Ndousse-Fetter - Brian Tierney Summer er 201 201 E ESCC/Inter ernet et2 J 2 Joint T Tec echs 1 . .

  2. Outline • The NPC forecasting problem • Standard forecasting tools • Flow field forecasting • Demonstrations • Present status Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 2 . . . .

  3. The problem Given: A sequence of observations of an NPC along a network path with given observation times NPC ?  ?  t 1 t 2 t 3 t 4 t 5 t 6 t 7 t N  1 t N t F t F Goal: Predict the NPC reliably at time F in the near future t Extrapolate plausibly at F t beyond the near future Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 3 . . . .

  4. Beyond near future Stable underlying mechanism Non-stable underlying mechanism Reliable prediction beyond near future Unreliable prediction beyond near future David Keeling CO2 Data - Mauna Loa, Hawaii Euro - US Dollar Exchange Rate 400 $1.60 Sole Atmospheric CO2 (ppm) 390 Measured monthly currency $1.50 of EU 380 $1.40 370 Euro price $1.30 360 350 $1.20 340 $1.10 330 $1.00 320 $0.90 310 Recorded daily, at close 1960 1970 1980 1990 2000 2010 $0.80 Year 1/1/01 1/1/03 1/1/05 1/1/07 1/1/09 1/1/11 Channel Untilization - Los Angeles/Houston Backbone Sunspot cycle 200 U.S. National Oceanic and Los Angeles Inbound (Gbits/sec) Atmospheric Administration 4 150 3 Count 100 2 50 1 rt r.losa.int ernet 2.edu--ge-6/ 1/ 0.0 BACKBONE: LOSA-HOUS 1 | 0 0 12-HOUS-LOSA-10GE-05581 noon 20:00 04:00 noon 20:00 04:00 noon 1700 1750 1800 1850 1900 1950 2000 Fri 24 Jun 2011 12:00:00 EDT - Sun 26 Jun 2011 12:00:00 EDT Year Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joi oint nt T Techs hs 4 . . . .

  5. Desiderata Error estimate – reliably estimates near-future prediction error Plausible – plausibly extrapolates beyond near future Robust – accepts non-uniformly spaced observation times Autonomous – no human guidance Fast – computationally efficient; e.g., no multi- dimensional numerical optimizations Accommodative – capable of exploiting “parallel” data Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 5 . . . .

  6. Available Tools Gaussian process regression Semiparametric regression Spectral/wavelet methods Traditional regression ARIMA forecasting Moving averages Neural networks ??? Error estimate X √ √ X √ √ X Plausible X X X X √ X √ Robust X √ X √ √ √ X Autonomous √ X X √ X √ √ Fast √ √ X X X √ √ Accommodative X √ X √ X X X Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 6 . . . .

  7. New Forecasting Method Gaussian process regression Semiparametric regression Spectral/wavelet methods Traditional regression Flow field forecasting ARIMA forecasting Moving averages Neural networks Error estimate X √ √ X √ √ X √ Plausible X X X X √ X √ √ Robust X √ X √ √ √ X √ Autonomous √ X X √ X √ √ √ Fast √ √ X X X √ √ √ Accommodative X √ X √ X X X √ Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 7 . . . .

  8. Flow Field Forecasting - 1 Step 1: Extract skeleton = + ε Data Y S n n n κ κ κ κ    0   1 2 3 K = Skeleton  s s s s s   0 1 2 3 K   δ δ δ δ δ    0 1 2 3 K = Noise • Semi-parametric regression • Use only skeleton for forecast • Data reduction        + Skeleton      • Original time spacing not             relevant   Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 8 . . . .

  9. Flow Field Forecasting - 2 Step 2: Interpolate flow field Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 9 .

  10. Flow Field Forecasting - 3 Step 3: Build to the future Skeleton  K  1  K  2  K  3   d 0 d 1 d 2 d K            s 0 s 1 s 2 s K s K  1 s K  2 s K  3     0  1  2  K  K  1  K  2  K  3     flowfield interpolation Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint T Techs 10 10 . . . .

  11. Demonstration 1 - Flow field contains much applicable information - Prediction error builds slowly Su Summer 201 ESC ESCC/Inter ernet et2 J 2 Joint Tec echs 11 11 . . . .

  12. Demonstration 2 - Flow field contains little applicable information - Prediction error builds quickly Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint Tec echs 12 . . . .

  13. Current Status • Python code undergoing final testing For a copy  mfrey@bucknell.edu • Integration with e-Weather Center can begin in September 2011 • “Introducing Flow Field Forecasting” by Frey and Caudle For a copy  mfrey@bucknell.edu Summer er 2 201 01 ESCC/Inter ernet et2 J 2 Joint Tec echs 13 . . . .

Recommend


More recommend