In Catla-HS, there is an additional component called PredictorRunner to facilitate performance change's fitting and predition. With the use of multiple fitting analysis, we can establish the prediction model for evaluating MapReduce job performance.
The component currently supports:
- linear fitting
- multivariate linear fitting
- logarithmic fitting
- exponential fitting
- polynomial fitting
More models can be adapted within the system following the common interface of Predictor.
The interface of Predictor:
package cn.edu.bjtu.cdh.catla.prediction;
public interface Predictor {
public void fit();
public void predict();
public void plot();
public void plotError();
}Then we create a factory for creating instance of Predictor to support multiple fitting models, like:
package cn.edu.bjtu.cdh.catla.prediction;
import java.util.Map;
public class PredictorFactory {
public static Predictor createInstance(Map<String,String>options,String type,String[] xfields,String[] yfields, String[][] x_data,String[][] y_data) {
if(type.equals("poly")) {
return new PolyPredictor(options,xfields,yfields,x_data,y_data);
}else if (type.equals("line")) {
return new LinePredictor(options,xfields,yfields,x_data,y_data);
}else if (type.equals("lineXs")) {
return new LineXsPredictor(options,xfields,yfields,x_data,y_data);
}else if (type.equals("exp")) {
return new ExpPredictor(options,xfields,yfields,x_data,y_data);
}else if (type.equals("log")) {
return new LogPredictor(options,xfields,yfields,x_data,y_data);
}
return null;
}
}So when running the PredictorRunner, the runner can implement the functions fit(), predict(), plot() and plotError().
Implementing a specific fitting model, for example, for polynomial fitting, we can define:
package cn.edu.bjtu.cdh.catla.prediction;
import java.util.Map;
import org.knowm.xchart.SwingWrapper;
import org.knowm.xchart.internal.chartpart.Chart;
import cn.edu.bjtu.cdh.catla.visualization.CatlaChart;
import cn.edu.bjtu.cdh.catla.visualization.ChartFactory;
public class PolyPredictor extends CommonPredictor implements Predictor {
// raw data
private Map<String, String> options;
private String[][] x_data, y_data;
// estimated parameters for the prediction model
private double[] model_result;
// real transformed data for fitting model
private double[][] x;
private double[] y;
// estimated/predicted y based on input x
private double[] predicted_ys;
// input and output x,y fields
private String[] x_fields, y_fields;
public PolyPredictor(Map<String, String> options, String[] xfields, String[] yfields, String[][] xdata,
String[][] ydata) {
this.options = options;
this.x_data = xdata;
this.y_data = ydata;
this.x_fields = xfields;
this.y_fields = yfields;
}
public void fit() {
this.x = toParseDoubleMatrix(x_data);
this.y = toParseDoubleList(y_data[0]);
int n = Integer.parseInt(options.getOrDefault("-n", "3"));
this.model_result = MathUtil.dxsFitting(x[0], y, n);
}
public void predict() {
double[] predicted_ys = new double[x_data[0].length];
for (int i = 0; i < x_data[0].length; i++) {
double predicted_y = MathUtil.getDxsValueByX(x[0][i], new double[] { this.model_result[0],
this.model_result[1], this.model_result[2], this.model_result[3] });
predicted_ys[i] = predicted_y;
}
this.predicted_ys = predicted_ys;
}
public void plot() {
// plot data transformation
String[][] plot_x = new String[x_data[0].length][x_data.length];
for (int i = 0; i < plot_x.length; i++) {
for (int j = 0; j < plot_x[i].length; j++) {
plot_x[i][j] = x_data[j][i];
}
}
String[][] plot_y = new String[x_data[0].length][2];
for (int i = 0; i < plot_y.length; i++) {
plot_y[i][0] = y_data[0][i];
plot_y[i][1] = predicted_ys[i] + "";
}
// print plot dataset
printTable("x", plot_x);
printTable("y", plot_y);
String plot_type = options.getOrDefault("-plot_type", "line");
String plot_title = options.getOrDefault("-plot_title", "");
// show graph
CatlaChart chart = ChartFactory.createChart(plot_type, plot_title, x_fields[0], y_fields[0], plot_x, plot_y,
new String[] { "real", "predicted" });
new SwingWrapper<Chart>(chart.getChart()).displayChart();
}
public void plotError() {
// plot data transformation
String[][] plot_x = new String[y_data[0].length][1];
for (int i = 0; i < y_data[0].length; i++) {
plot_x[i][0] = y_data[0][i];
}
String[][] plot_y = new String[x_data[0].length][1];
for (int i = 0; i < plot_y.length; i++) {
plot_y[i][0] = predicted_ys[i] + "";
}
// print plot dataset
printTable("x", plot_x);
printTable("y", plot_y);
// show graph
CatlaChart chart = ChartFactory.createChart("scatter", "difference between real and predicted values", "real", "predicted", plot_x, plot_y,
new String[] { "difference" });
new SwingWrapper<Chart>(chart.getChart()).displayChart();
}
}Here are some usage examples to perform prediction analysis in Catla-HS.
public static void testPoly() {
String[] args=new String[] {
"-dir","E:\\CatlaHS\\bobyqa-two\\history",
"-predictor","poly",
"-x","Order",
"-y","totalTimeCost",
"-plot_type","line",
"-plot_title","Tuning summary",
"-n","3"
};
PredictorRunner.main(args);
}The -dir parameter should be assigned with the path of history folder or project folder.
The -predictorparameter represents model used to fit, current value set includes poly, line, lineXs, exp, log, and other models defined by users in the future.
The -x parameter represents the fields used for x-axis.
The -y parameter represents the fields used for y-axis, you can use multiple fields split by ,.
The -plot_title parameter repesents caption of graph.
The -plot_type parameter represents style of graph.
The -n parameter represents other parameters use for specific models.
Here are the results:
Fitting model:
Comparison between real values and predicted values:
Other models can be similiarly estimated using PredictorRunner in Catla-HS.
You can find full code implementation here.