7. Extrae On-line User Guide

7.1. Introduction

Extrae On-line is a new module developed for the Extrae tracing toolkit, available from version 3.0, that incorporates intelligent monitoring, analysis and selection of the traced data. This tracing setup is tailored towards long executions that are producing large traces. Applying automatic analysis techniques based on clustering, signal processing and active monitoring, Extrae gains the ability to inspect and filter the data while it is being collected to minimize the amount of data emitted into the trace, while maximizing the amount of relevant information presented.

Extrae On-line has been developed on top of Synapse [1], a framework that facilitates the deployment of applications that follow the master/slave architecture based on the MRNet [2] software overlay network. Thanks to its modular design, new types of automatic analysis can be added very easily as new plug-ins into the on-line tracing system, just by defining new Synapse protocols.

This document briefly describes the main features of the Extrae On-line module, and shows how it has to be configured and the different options available.

7.2. Automatic analysis

Extrae On-line currently supports three types of automatic analysis:

  1. fine-grain structure detection based on clustering techniques,
  2. periodicity detection based on signal processing techniques,
  3. and multi-experiment analysis based on active monitoring techniques.

Extrae On-line has to be configured to apply one of these types of analysis, and then the analysis will be performed periodically as new data is being traced.

7.3. Structure detection

This mechanism aims at identifying the fine-grain structure of the computing regions of the program. Applying density-based clustering, this method is able to expose the main performance trends in the computations, and this information is useful to focus the analysis on the zones of real interest. To perform the cluster analysis, Extrae On-line relies on the ClusteringSuite tool [3].

At each phase of analysis, several outputs are produced:

  • A scatter-plot representation that illustrates the behavior of the main computing regions of the program, that enables a quick evaluation of potential imbalances.
  • A summary of several performance metrics per cluster.
  • On supported machines, a CPI stack model that attributes stall cycles to specific hardware components.
  • And a trace that is augmented with the clusters information, that allows to identify patterns of performance and variabilites.

Subsequent clustering results can be used to study the evolution of the application over time. In order to study how the clusters are evolving, the xtrack tool [4] can be used.

7.4. Periodicity detection

This mechanism allows to detect iterative patterns over a wide region of time, and precisely delimit where the iterations start. Once a period has been found, those iterations presenting less perturbations are selected to produce a representative trace, and the rest of the data is basically discarded. The result of applying this mechanism is a compact trace where only the representative iterations are traced in full detail, and for the rest of the execution we can optionally keep summarized information in the form of phase profiles or a “low resolution” trace.

Please note that applying this technique to a very short execution, or if no periodicity can be detected in the application, may result in an empty trace depending on the configuration options selected (see section Configuration).

7.5. Multi-experiment analysis

This mechanism employs active measurement techniques in order to simulate different execution scenarios under the same execution. Extrae On-line is able to add controlled interference into the program to simulate different computation loads, network bandwidth, memory congestion and even tuning some configurations of the parallel runtime (currently supports MPI Dynamic Load Balance (DLB) runtime). Then, the application behavior can be studied under different circumstances, and tracking can be used to analyze the impact of these configurations on the program performance. This technique aims at reducing the number of executions necessary to evaluate different parameters and characteristics of your program.

7.6. Configuration

In order to activate the On-line tracing mode, the user has to enable the corresponding configuration section in the Extrae XML configuration file. This section is found under trace-control > remote-control > online. The default configuration is already ready to use:

<online enabled="yes"

The available options for the <online> section are the following:

  • enabled Set to yes to activate the On-line tracing mode.
  • analysis Choose from clustering, spectral and gremlins.
  • frequency Set the time in seconds after which a new phase of analysis will be triggered, or auto to let Extrae decide this automatically.
  • topology Set the desired tree process tree topology, or auto to let Extrae decide this automatically.

Depending on the analysis selected, the following specific options become available.

7.6.1. Clustering analysis options

<clustering config="cl.I.IPC.xml"/>
  • config Specify the path to the ClusteringSuite XML configuration file.

7.6.2. Spectral analysis options

<spectral max_periods="0"
  <spectral_advanced enabled="no"
    <periodic_zone detail_level="profile" />
    <non_periodic_zone detail_level="bursts" min_duration="3s" />

The basic configuration options for the spectral analysis are the following:

  • max_periods Set to the maximum number of periods to trace, or all to explore the whole run.
  • num_iters Set to the number of iterations to trace per period.
  • min_seen Minimum repetitions of a period before tracing it (0 to trace the first time that you encounter it).
  • min_likeness Minimum percentage of similarity to compare two equivalent periods.

Also, some advanced settings are tunable in the <spectral_advanced> section:

  • enabled Set to yes to activate the spectral analysis advanced options.
  • burst_threshold Filter threshold to keep all CPU bursts that add up to the given total time percentage.
  • detail_level Specify the granularity of the data stored for the non-representative iterations of the periodic region, and in the non-periodic regions. Choose from none (everything is discarded), profile (phase profile at the start of each iteration/region) or bursts (trace in bursts mode).
  • min_duration Minimum duration in seconds of the non-periodic regions for emitting any information regarding that region into the trace.

7.6.3. Gremlins analysis options

<gremlins start="0"
          loop="no" />
  • start Number of gremlins at the beginning of the execution.
  • increment Number of extra gremlins at each analysis phase. Can also be a negative value to indicate that you want to remove gremlins.
  • roundtrip Set to yes if you want to start adding gremlins after you decrease to 0, or vice-versa, start removing gremlins after you reach the maximum.
  • loop Set to yes if you want to go back to the initial number of gremlins and repeat the sequence of adding/removing gremlins after you have finished a complete sequence.


[3]You can download it from https://tools.bsc.es/downloads.
[4]You can download it from https://tools.bsc.es/downloads.