1. Copyright 2004 David J. Lilja1 Errors in Experimental Measurements Sources of errors Accuracy, precision, resolution A mathematical model of errors Confidence intervals For means For proportions How many measurements are needed for desired error?

2. Copyright 2004 David J. Lilja2 Why do we need statistics? 1. Noise, noise, noise, noise, noise! OK – not really this type of noise

3. Copyright 2004 David J. Lilja3 Why do we need statistics? 2. Aggregate data into meaningful information. 445 446 397 226 388 3445 188 1002 47762 432 54 12 98 345 2245 8839 77492 472 565 999 1 34 882 545 4022 827 572 597 364

4. Copyright 2004 David J. Lilja4 What is a statistic? “A quantity that is computed from a sample [of data].” Merriam-Webster → A single number used to summarize a larger collection of values.

5. Copyright 2004 David J. Lilja5 What are statistics? “A branch of mathematics dealing with the collection, analysis, interpretation, and presentation of masses of numerical data.” Merriam-Webster → We are most interested in analysis and interpretation here. “Lies, damn lies, and statistics!”

6. Copyright 2004 David J. Lilja6 Goals Provide intuitive conceptual background for some standard statistical tools. Draw meaningful conclusions in presence of noisy measurements. Allow you to correctly and intelligently apply techniques in new situations. → Don’t simply plug and crank from a formula.

7. Copyright 2004 David J. Lilja7 Goals Present techniques for aggregating large quantities of data. Obtain a big-picture view of your results. Obtain new insights from complex measurement and simulation results. → E.g. How does a new feature impact the overall system?

8. Copyright 2004 David J. Lilja8 Sources of Experimental Errors Accuracy, precision, resolution

9. Copyright 2004 David J. Lilja9 Experimental errors Errors → noise in measured values Systematic errors Result of an experimental “mistake” Typically produce constant or slowly varying bias Controlled through skill of experimenter Examples Temperature change causes clock drift Forget to clear cache before timing run

10. Copyright 2004 David J. Lilja10 Experimental errors Random errors Unpredictable, non-deterministic Unbiased → equal probability of increasing or decreasing measured value Result of Limitations of measuring tool Observer reading output of tool Random processes within system Typically cannot be controlled Use statistical tools to characterize and quantify

11. Copyright 2004 David J. Lilja11 Example: Quantization → Random error

12. Copyright 2004 David J. Lilja12 Quantization error Timer resolution → quantization error Repeated measurements X ± Δ Completely unpredictable

13. Copyright 2004 David J. Lilja13 A Model of Errors ErrorMeasured value Probability -Ex – E½ +Ex + E½

14. Copyright 2004 David J. Lilja14 A Model of Errors Error 1Error 2Measured value Probability -E x – 2E¼ -E +Ex¼ -Ex¼ +E x + 2E¼

15. Copyright 2004 David J. Lilja15 A Model of Errors

16. Copyright 2004 David J. Lilja16 Probability of Obtaining a Specific Measured Value

17. Copyright 2004 David J. Lilja17 A Model of Errors Pr(X=x i ) = Pr(measure x i ) = number of paths from real value to x i Pr(X=x i ) ~ binomial distribution As number of error sources becomes large n → ∞, Binomial → Gaussian (Normal) Thus, the bell curve

18. Copyright 2004 David J. Lilja18 Frequency of Measuring Specific Values Mean of measured values True value Resolution Precision Accuracy

19. Copyright 2004 David J. Lilja19 Accuracy, Precision, Resolution Systematic errors → accuracy How close mean of measured values is to true value Random errors → precision Repeatability of measurements Characteristics of tools → resolution Smallest increment between measured values

20. Copyright 2004 David J. Lilja20 Quantifying Accuracy, Precision, Resolution Accuracy Hard to determine true accuracy Relative to a predefined standard E.g. definition of a “second” Resolution Dependent on tools Precision Quantify amount of imprecision using statistical tools

21. Copyright 2004 David J. Lilja21 Confidence Interval for the Mean c1c2 1-α α/2

22. Copyright 2004 David J. Lilja22 Normalize x

23. Copyright 2004 David J. Lilja23 Confidence Interval for the Mean Normalized z follows a Student’s t distribution (n-1) degrees of freedom Area left of c 2 = 1 – α/2 Tabulated values for t c1c2 1-α α/2

24. Copyright 2004 David J. Lilja24 Confidence Interval for the Mean As n → ∞, normalized distribution becomes Gaussian (normal) c1c2 1-α α/2

25. Copyright 2004 David J. Lilja25 Confidence Interval for the Mean

26. Copyright 2004 David J. Lilja26 An Example ExperimentMeasured value 18.0 s 27.0 s 35.0 s 49.0 s 59.5 s 611.3 s 75.2 s 88.5 s

27. Copyright 2004 David J. Lilja27 An Example (cont.)

28. Copyright 2004 David J. Lilja28 An Example (cont.) 90% CI → 90% chance actual value in interval 90% CI → α = 0.10 1 - α /2 = 0.95 n = 8 → 7 degrees of freedom c1c2 1-α α/2

29. Copyright 2004 David J. Lilja29 90% Confidence Interval a n0.900.950.975 ………… 51.4762.0152.571 61.4401.9432.447 71.4151.8952.365 ………… ∞1.2821.6451.960

30. Copyright 2004 David J. Lilja30 95% Confidence Interval a n0.900.950.975 ………… 51.4762.0152.571 61.4401.9432.447 71.4151.8952.365 ………… ∞1.2821.6451.960

31. Copyright 2004 David J. Lilja31 What does it mean? 90% CI = [6.5, 9.4] 90% chance real value is between 6.5, 9.4 95% CI = [6.1, 9.7] 95% chance real value is between 6.1, 9.7 Why is interval wider when we are more confident?

32. Copyright 2004 David J. Lilja32 Higher Confidence → Wider Interval? 6.5 9.4 90% 6.19.7 95%

33. Copyright 2004 David J. Lilja33 Key Assumption Measurement errors are Normally distributed. Is this true for most measurements on real computer systems? c1c2 1-α α/2

34. Copyright 2004 David J. Lilja34 Key Assumption Saved by the Central Limit Theorem Sum of a “large number” of values from any distribution will be Normally (Gaussian) distributed. What is a “large number?” Typically assumed to be >≈ 6 or 7.

35. Copyright 2004 David J. Lilja35 How many measurements? Width of interval inversely proportional to √n Want to minimize number of measurements Find confidence interval for mean, such that: Pr(actual mean in interval) = (1 – α)

36. Copyright 2004 David J. Lilja36 How many measurements?

37. Copyright 2004 David J. Lilja37 How many measurements? But n depends on knowing mean and standard deviation! Estimate s with small number of measurements Use this s to find n needed for desired interval width

38. Copyright 2004 David J. Lilja38 How many measurements? Mean = 7.94 s Standard deviation = 2.14 s Want 90% confidence mean is within 7% of actual mean.

39. Copyright 2004 David J. Lilja39 How many measurements? Mean = 7.94 s Standard deviation = 2.14 s Want 90% confidence mean is within 7% of actual mean. α = 0.90 (1-α/2) = 0.95 Error = ± 3.5% e = 0.035

40. Copyright 2004 David J. Lilja40 How many measurements? 213 measurements → 90% chance true mean is within ± 3.5% interval

41. Copyright 2004 David J. Lilja41 Proportions p = Pr(success) in n trials of binomial experiment Estimate proportion: p = m/n m = number of successes n = total number of trials

42. Copyright 2004 David J. Lilja42 Proportions

43. Copyright 2004 David J. Lilja43 Proportions How much time does processor spend in OS? Interrupt every 10 ms Increment counters n = number of interrupts m = number of interrupts when PC within OS

44. Copyright 2004 David J. Lilja44 Proportions How much time does processor spend in OS? Interrupt every 10 ms Increment counters n = number of interrupts m = number of interrupts when PC within OS Run for 1 minute n = 6000 m = 658

45. Copyright 2004 David J. Lilja45 Proportions 95% confidence interval for proportion So 95% certain processor spends 10.2-11.8% of its time in OS

46. Copyright 2004 David J. Lilja46 Number of measurements for proportions

47. Copyright 2004 David J. Lilja47 Number of measurements for proportions How long to run OS experiment? Want 95% confidence ± 0.5%

48. Copyright 2004 David J. Lilja48 Number of measurements for proportions How long to run OS experiment? Want 95% confidence ± 0.5% e = 0.005 p = 0.1097

49. Copyright 2004 David J. Lilja49 Number of measurements for proportions 10 ms interrupts → 3.46 hours

50. Copyright 2004 David J. Lilja50 Important Points Use statistics to Deal with noisy measurements Aggregate large amounts of data Errors in measurements are due to: Accuracy, precision, resolution of tools Other sources of noise → Systematic, random errors

51. Copyright 2004 David J. Lilja51 Important Points: Model errors with bell curve True value Precision Mean of measured values Resolution Accuracy

52. Copyright 2004 David J. Lilja52 Important Points Use confidence intervals to quantify precision Confidence intervals for Mean of n samples Proportions Confidence level Pr(actual mean within computed interval) Compute number of measurements needed for desired interval width