1. 1 Reading Report 9 Yin Chen 29 Mar 2004 Reference: Multivariate Resource Performance Forecasting in the Network Weather Service, Martin Swany and Rich Wolski http://sc-2002.org/paperpdfs/pap.pap292.pdf

2. 2 Problems Overview Frequently, monitor data is used as a prediction of future performance. The Common Method  “Forecast “ based on the last value : Assume the performance of a given transfer will be the same as it was the last time it was performed.  i.e., For network bandwidth estimation, many users simply conduct lengthy data transfer between end-hosts, observe the throughput, and use that observation as the prediction. Two Problems  Not clear that the last observation is a good estimation.  Require the resource be “loaded” enough : On high-throughput networks with lengthy round-trip times, enough data must be transferred to match the bandwidth-delay product for the end-to-end route. Costly -- waste resource and lost time while the probe take place.

3. 3 Network Weather Service’s Solutions Using statistical techniques Combine other performance monitoring tools  Combine infrequent, irregularly spaced expensive measurements with regularly spaced, but far less intrusive probes  Combine short NWS bandwidth probes and previous HTTP history transfers  Combine data from two or more measurement streams  Combine heavy-weight operations with lightweight, inexpensive probes of a resource

4. 4 Forecasting Using Correlation The univariate forecaster represent the current and future performance of a single dataset. The multivariate forecaster operate on some subset of a collection measurement series, characterized by units and frequency, with a combination of correlation and forecasting. The correlator serves to map X values onto Y values where the X values are plentiful and “cheap” and the Y values are “rare” and expensive.  Given two correlated variables, knowledge of one at a given point in time provides information as to the value of the other.  Fix the value of one of the variables can make reasonable assumptions about the probability of various value of the other variable, based on their history of correlation.

5. 5 Possible Correlator Methodology Possible Mapping Methods :  Linear regression  Traditional correlation operate on datasets of equal sizes  Assume that the variables in question are related linearly Problems -- measurement data gathered form a variety of sources:  Data sets may be of different sizes  Data items in each set may be gathered with different frequencies or with different regularities  The units of measure may be different between data sets.

6. 6 Correlator Methodology of the Work Rank Correlation Techniques  A rank correlation measure sorts datasets and does linear correlation based on their position in the sorted list (their “rank”).  It is non-parametirc, appropriated because the distribution of the data cannot be certain of. Cumulative Distribution Function (CDF)  Defined as the probability for some set of sample points that a value is less than or equal to some real valued x.  If the probability density function (PDF) is known, CDF is defined as:  The reason of using CDF Practicality of delivering this information to Grid A CDF dataset may be compressed with varying degrees of loss. i.e., a handful of points can describe the CDF with linear interpolation between specified points.

7. 7 Correlator Methodology of the Work (Cont.) Correlator Methodology  Approximate the CDF by :  Where count x is the total elements in the sample set. Which is equivalent to  Rank measurements in datasets of different size by computing the CDF for each of them.  Use the position of the value in X (x target ) to produce a value of Y (y forecast )  This gives the value in Y (in CDF Y ) associated with the position x target  If the CDF position of x target (CDF x (x target )) is between measured Y values, linear interpolation is used to determine the value of y forecast  This mapping is similar to quantile-quantile comparision.  The x target can be chosen in a variety of ways.  Choose x target by using the NWS forecasters to produce the value X target = Forecast(X)

8. 8 Correlator Methodology of the Work (Cont.) Figure 1 shows the NWS(X) values. Figure 2 shows the HTTP(X) values. To compute the forecast for Y use the position in the CDF of x target which yields some real number between 0 and 1 This predictive method forms both the X and Y CDFs. At each prediction cycle, a univariate forecast of X is taken and that X value’s position is determined. That value is used to determine the value in Y at the corresponding position, which is the prediction.

9. 9 Correlator Methodology of the Work (Cont.) Error  Compute the Mean Normalized Error Percentage (MNEP) as the last prediction’s error over the current average value.  This is required due to the forecast-oriented application of the system  The mean used (mean t ) is the current mean and is reevaluated at each timestep.

10. 10 Experimental Methodology Generate default-sized NWS bandwidth measurements (64K bytes), using the TCP/IP end-to-end sensor between a parit of machines every 10 sec. Initiate a 16MB HTTP transfer every min. and record the observed transfer bandwidth. Figure 3 shows the NWS data Figure 4 shows the HTTP data The exact correlative relationship is not clear visually, but the shapes appear similar.

11. 11 Experimental Methodology (Cont.) Goal  Use this data to investigate the accuracy with which the short 64K NWS transfer could be used to predict the long 16MB HTTP transfers given successively longer periods. Why HTTP?  Widely used in the Internet  A key components of SOAP and Web services  Used more and more in high-performance computing  This data represents: The relationship between short and long network experiments. Instrumentation data being fed into to the forecasting service. Compare with 3 prediction methods  Last value method  The univariate NWS forecast applied solely to the 16MB probe history  The new correlative methods Why 16MB?  To limit intrusiveness  Less variability in the observed bandwidth as the amount of data transferred gets larger.  Any transient periods of high congestion are amortized over the longer transfer time.  Not so long as to make things easy, but long enough to represent real applications and real data.

12. 12 Results Figure 5 shows the MAE of the univariate compared to the multivariate forecasters.  Multivariate forecaster performs better.  As Y become less frequent, the univariate forecasters lose accuracy.  Using the new forecasting technique, the NWS can predict 16MB bandwidth using 64K measurements with a mean error of 0.47 megabits/sec. Figure 6 shows MNEP of the MAE.  The MNEP shows the percentage of the current average value that an error represents.  At 500 min. the 64K transfers can predict 16MB transfer bandwidth to within 7% on the average. Figure 5. Comparison of MAE between univariate and multivariate forecasts for different frequencies of HTTP measurements. Figure 6. Comparison of MNEP of MAE between univariate and multivariate forecasts for different frequencies of HTTP measurements.

13. 13 Results (Cont.) Figure 7 shows the square root of the MSE of both types of forecasts.  This value bears resemblance to the standard deviation and tends to indicate the variance of the error.  The new method predict relatively high levels of accuracy. Figure 8 show the MNEP of the MAE for the CDF forecaster and the “Last Value” predictor.  Last Value predictor has over 100% error. END… Figure 7. Comparison of the square root of the MSE between univariate and multivariate forecasts for different frequencies of HTTP measurements. Figure 8. Comparison of MNEP of MAE between “Last Value” and multivariate forecasts.