Skip to content
TACC Stats is an automated resource-usage monitoring and analysis package.
C Python Shell HTML M4 Makefile JavaScript
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
monitor updating dist Apr 11, 2019
tacc_stats new bokeh function signatures adhered to in Aug 13, 2019
.gitignore cleaned up refactored Aug 21, 2018
LICENSE Update Oct 2, 2018 fixed a few bugs in config Aug 23, 2018
tacc_stats.ini new bokeh function signatures adhered to in Aug 13, 2019

tacc_stats Documentation {#mainpage}


Developers and Maintainers

R. Todd Evans ( Bill Barth (

Original Developer

John Hammond


The tacc_stats package provides the tools to monitor resource usage of HPC systems at multiple levels of resolution.

The package is split into an autotools-based monitor subpackage and a Python setuptools-based tacc_stats subpackage. monitor performs the online data collection and transmission in a production environment while tacc_stats performs the data curation and analysis in an offline environment.

Installing monitor will build and install an System V service, /etc/init.d/taccstats. This service launches a daemon with an overhead of 3-9% on a single core when configured to sample at a frequency of 1Hz. It is typically configured to sample at 10 minute intervals, with samples taken at the start and end of every job as well. tacc_stats sends the data directly to a RabbitMQ server over the administrative ethernet network. RabbitMQ must be installed and running on the server in order for the data to be received.

Installing the tacc_stats module will setup a Django-based web application along with tools for extracting the data from the RabbitMQ server and feeding them into a PostgreSQL database.

Code Access

To get access to the tacc_stats source code clone this repository:

git clone


monitor subpackage

First ensure the RabbitMQ library and header file are installed on the build and compute nodes


./configure --enable-rabbitmq; make; make install will then successfully build the tacc_stats executable for many systems. If Xeon Phi coprocessors are present on your system they can be monitored with the --enable-mic flag. Additionally the configuration options, --disable-infiniband, --disable-lustre, --disable-hardware will disable infiniband, Lustre Filesystem, and Hardware Counter monitoring. Not enabling RabbitMQ will result in a legacy build of tacc_stats that relies on the shared filesystem to transmit data. This mode is not recommended. If libraries or header files are not found than add their paths to the include and library paths with the CPPFLAGS and/or LDFLAGS vars as usual.

There will be a configuration file, /etc/tacc_stats.conf, after installation. This file contains the fields





SERVER should be set to the RabbitMQ server, QUEUE to the system name, PORT to the RabbitMQ port (5672 should be ok), and FREQ to the desired sampling frequency in seconds.

An RPM can be built for subpackage deployment using the tacc_statsd.spec file. The most straightforward approach to build this is to setup your rpmbuild directory then

rpmbuild -ba tacc_statsd.spec --define 'rmqserver rabbitmqservername' --define 'system systemname'

where the rmqserver will be the RabbitMQ SERVER hostname and system will be the QUEUE in tacc_stats.conf.

After installation the executable /opt/tacc_statsd/tacc_stats, service /etc/init.d/taccstats, and config file /etc/tacc_stats.conf should exist. If the rpm was used for installation tacc_stats will be chkconfig'd to start at boot time and be running. tacc_stats can be started, stopped, and restarted using taccstats start, taccstats stop, and taccstats restart.

In order to notify tacc_stats of a job beginning echo the job id into /var/run/TACC_jobid. It order to notify it of a job ending echo - into /var/run/TACC_jobid. This can be accomplished in the job scheduler prolog and epilog for example.

Job Scheduler Configuration

In order for tacc_stats to correcly label records with JOBIDs it is required that the job scheduler prolog and epilog contain the lines

echo $JOBID > jobid_file


echo - > jobid_file

To perform the pickling of this data it is also necessary to generate an accounting file that contains the following information in the following format


for example,


If using SLURM the script that installed with tacc_stats may be used.

tacc_stats subpackage

To install TACC Stats on the machine where data will be processed, analyzed, and the webserver hosted follow these steps:

  1. Download the package and setup the Python virtual environment.
$ virtualenv machinename --system-site-packages
$ cd machinename; source bin/activate
$ git clone

tacc_stats is a pure Python package. Dependencies should be automatically downloaded and installed when installed via pip. The package must first be configured however.
2. The initialization file, tacc_stats.ini, controls all the configuration options and has the following content and descriptions

## Basic configuration options - modify these
# machine       = unique name of machine/queue
# server        = database and rmq server hostname
# data_dir      = where data is stored
machine         = ls5
data_dir        = /hpc/tacc_stats_site/%(machine)s
server          =

## RabbitMQ Configuration
# RMQ_SERVER    = RMQ server
# RMQ_QUEUE     = RMQ server
rmq_server      = %(server)s
rmq_queue       = %(machine)s

## Configuration for Web Portal Support
acct_path       = %(data_dir)s/accounting/tacc_jobs_completed
host_list_dir   = %(data_dir)s/hostfile_logs
pickles_dir     = %(data_dir)s/pickles
archive_dir     = %(data_dir)s/archive
host_name_ext   = %(machine)
batch_system    = SLURM

Set these paths as needed. The raw stats data will be stored in the archive_dir and processed stats data in the pickles_dir. machine should match the system name used in the RabbitMQ server QUEUE field. This is the only field that needs to match anything in the monitor subpackage. 3. Install tacc_stats

$ pip install -e tacc_stats/
  1. Start the RabbitMQ server reader in the background, e.g.
$ nohup > /tmp/listend.log

Raw stats files will now be generated in the archive_dir. 5. A PostgreSQL database must be setup on the host. To do this, after installation of PostgreSQL and the tacc_stats Python module

$ sudo su - postgresql
$ psql
# CREATE DATABASE machinename_db;
# CREATE USER taccstats WITH PASSWORD 'taccstats';
# ALTER ROLE taccstats SET client_encoding TO 'utf8';
# ALTER ROLE taccstats SET default_transaction_isolation TO 'read committed';
# ALTER ROLE taccstats SET timezone TO 'UTC';
# ALTER DATABASE machinename_db OWNER TO taccstats;
# GRANT ALL PRIVILEGES ON DATABASE machinename_db TO taccstats;
# \q


$ python migrate

This will generate a table named machinename_db in your database.

  1. Setup cron jobs to process raw data and ingest into database. Add the following to your cron file
*/15 * * * * source /home/rtevans/testing/bin/activate;; > /tmp/ls5_update.log 2>&1
  1. Next configure the Apache server (make sure it is installed and the mod_wsgi Apache module is installed) A sample configuration file, /etc/httpd/conf.d/ls5.conf, looks like
LoadModule wsgi_module /stats/stampede2/lib/python3.7/site-packages/mod_wsgi/server/
WSGISocketPrefix run/wsgi

<VirtualHost *:80>


WSGIDaemonProcess s2-stats python-home=/stats/stampede2 python-path=/stats/stampede2/tacc_stats:/stats/stampede2/lib/python3.7/site-packages user=rtevans
WSGIProcessGroup s2-stats
WSGIScriptAlias / /tacc_stats/site/tacc_stats_site/ process-group=s2-stats
WSGIApplicationGroup %{GLOBAL}

<Directory /stats/stampede2/tacc_stats/tacc_stats/site/tacc_stats_site>
Require all granted
  1. Start up Apache

Running can be run manually by:

$ ./ [start_date] [end_date] [-dir directory] [-jobids id0 id1 ... idn]

where the 4 optional arguments have the following meaning

  • start_date : the start of the date range, e.g. "2013-09-25" (default is today)
  • end_date : the end of the date range, e.g. "2013-09-26" (default is start_date)
  • -dir : the directory to store pickled dictionaries (default is set in tacc_stats.ini)
  • -jobids : individual jobids to pickle (default is all jobs)

No arguments results in all jobs from the previous day getting pickled and stored in the pickles_dir defined in tacc_stats.ini. On Stampede argumentless is run every 24 hours as a cron job set-up by the user.

Pickled data format: generated

Pickled stats data will be placed in the directory specified by pickles_dir. The pickled data is contained in a nested python dictionary with the following key layers:

job       : 1st key Job ID
 host     : 2nd key Host node used by Job ID
  type    : 3rd key TYPE specified in tacc_stats
   device : 4th key device belonging to type

For example, to access Job ID 101's stats data on host c560-901 for TYPE intel_snb for device cpu number 0 from within a python script:

pickle_file = open('101','r')
jobid = pickle.load(pickle_file)

The value accessed by this key is a 2D array, with rows corresponding to record times and columns to specific counters for the device. To view the names for each counter add


or for a short version



(C) 2011 University of Texas at Austin


This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License as published by the Free Software Foundation; either version 2.1 of the License, or (at your option) any later version.

This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details.

You should have received a copy of the GNU Lesser General Public License along with this library; if not, write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

You can’t perform that action at this time.