From Tau Wiki
Jump to: navigation, search

Performance Regression with TAU

Performance Regression analysis is the study of how an application's performance varies with time. In many cases an application that is being actively developed might benefit from running a test each night to closely monitor changes in the application's parallel performance.

This page provides some information about how TAU can provide performance regressions. The basic strategy is to construct a set of scripts that can build and run your application with TAU on a particular machine and also to use TAU to gather the resulting performance data for analysis.

Downloading TAU/PDT

you can use these commands to download TAU and PDT (needed if you want to do source level instrumentation):

tar xzf pdt.tgz
tar xzf tau.tgz

Then you can then configure and build PDT/TAU as you would normally.

Next build and run your application with TAU (see TAUCompiler).

Creating a workspace on the TAU Portal

The TAU web portal provides a convenient way to store the performance regression data. Your performance data will be uploaded to our database servers in Oregon and you can access your data via ParaProf or PerfExplorer from anywhere with web access.

Create a user account and workspace on the portal here.

Upload data to the TAU Portal

Use paraprof to quickly pack your performance data

paraprof --pack application.ppk

You can use the script to upload your data to the TAU Portal: up -u [username] -p [password] -w [workspace name] -e [experiment name] application.ppk