Skip to content

[Feature request]: Establish method for measuring quality of regression/prediction of scaling #173

Description

@goranbs

Is your feature request related to a problem? Please describe

We do not have a good understanding of how good the resource prediction of the predicted scaling.

Currently we have implemented prediction of scaling using linear regression and power law regression.

We use origo as a point of reference when computing the linear or power law relation of the scaling of pipeline duration, cpu consumption and memory usage. This way, only one dryRun is needed to get a prediction of scaling. Of course, more dryRuns can be added to increase the number of points for the scaling.

Describe the solution you'd like

To establish a method that computes the quality of the regression prediction.

Describe alternatives you've considered

1.1 SST, SSR and SSER

These are common statistics measures for the error of regression.
This could be the basis of an analysis of the regression we are making as a function of how many dryRuns we are adding.

SST (Sum of Squares Total), SSR (Sum of squared residuals), SSER (Sum of squares regression error), SST = SSR + SSER

-- Regression analysis
---- dur, cpu, mem
---- y, f(x)
---- SSR = (y - f)^2
---- SST = (y - mean(y))^2
---- SSER = (f - mean(y))^2
---- SST = SSR + SSER
---- MSE = sqrt( SSR / (n - p) )

Introduction to statistics
statistics tutorial

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions