PERI Performance Data Collection - PERI

PERI Performance Data Collection

From PERI

Jump to: navigation, search

Contents

About this page

This page provides info and links about PERI performance data collection activities.

Additional info related to the PERI architecture tiger team is on this page.

Database at RENCI

There is a PostgreSQL database set up on perfdb.renci.org to store the TAU and PerfTrack data.

Loading TAU data

For TAU profiles, the strategy is to (1) copy the data to perfdb.renci.org, (2) load the data into the database.

To do either of these things, you will need an account on the RENCI host. To request an account, please contact Dan Gunter.

1. Copy to perfdb.renci.org

Data should be copied under /scratch/peri. This should be group-writable by group peri.

To copy data there, do something like:

rsync -avz $localdir perfdb.renci.org:/scratch/peri/$remotedir

For $remotedir the data is laid out by code, then by host:

/scratch/peri/{gtc,s3d,flash}/{jaguar,intrepid}/$rundir/<files>

The convention for $rundir is: <experiment>_<num_processors>_<date>, e.g. for an S3D run the chemistry may be the experiment name so the directory name would be c2h4_4096_20080816. The whole command would then be:

rsync -avz myrun/data/ perfdb.renci.org:/scratch/peri/s3d/jaguar/c2h4_4096_20080816

2. Load the data into the TAU database

1. Add TAU and Java to your path. For tcsh:

% set path=(/home/sameer/jdk1.6.0_01/bin $path)
% set path=(/home/sameer/tau2/x86_64/bin $path)
% perfdmf_configure

and give the login name, password ('y' to store password), and download the drivers in your area.

2a. Launch the paraprof GUI

% paraprof
  • right click on the DB tab in the Manager window and add Application/Experiment/Trial using the GUI

2b. Or, run the command-line application.

  • First, create an "application" in the database with:
% perfdmf_createapp -n <appname>
  • Next, use the application id this returns, <appid>, to create a new experiment:
% perfdmf_createexp -a <appid> -n <experiment_name>
  • Finally, use the experiment id this returns, <expid>, to upload the trial.
  • If you have created a packed profile (using paraprof --pack <file.ppk>), you may upload this dataset using:
% perfdmf_loadtrial -e <expid> -n <trialName> <file.ppk>
  • If you prefer to specify full names, this is possible too. See perfdmf_loadtrial --help. For example:
% perfdmf_createapp -n "GTC"
application id = 3
% perfdmf_createexp -a 3  -n "Jaguar"
experiment id = 8
% perfdmf_loadtrial -e 8 -n "4096" jaguar_gtc_4k.ppk
or
% perfdmf_loadtrial -a "GTC" -x "Jaguar" -n "4096" jaguar_gtc_4k.ppk