Partner Newsletter Q4 2010 Intel Cluster Ready Articles 2

TIPS & TRICKS

Intel® Cluster Checker 1.6 – Using the New Capabilities

By Andres More

Intel® Cluster Checker version 1.6 provides a wide range of new features and performance enhancements. This article provides suggestions for using a few of the new capabilities.

Specify Network Fabrics Centrally
The engine supports a centralized approach for network fabric selection inside the global_configuration container in the configuration file. This makes it easier to specify the network fabrics to be used by the different test modules.

The global configuration option with multiple entries is used to define the network fabrics to be used by test modules that exercise the Intel® MPI Library. The <network> container is used to hold the different fabrics that should be entered using the <fabric> container. Multiple fabrics can be configured and each one must include the corresponding MPI device (<device>). Optional attributes are available for the user to specify a custom name to the fabric (name=) and its state (enabled=). The default state is enabled.

[standard]<global_configuration>
<network>
<fabric name="IB" enabled="on">
<device>rdssm</device>
</fabric>
<fabric name="SM-TCP" enabled="on">
<device>shm:tcp</device>
</fabric>
</network>
</global_configuration>
[/standard]
Gather Information from Previous Runs
There’s now a --report option for gathering information from previous runs. To take advantage of this capability, instruct cluster-check to look for the latest output logs in the log directory and generate a report with descriptive information about each execution. The value you enter represents how many of the most recent logs should be examined to generate the output.

Files are ordered by date, with the latest files first. A maximum of five logs are examined. In addition to the latest five logs, all certification logs are checked for a successful execution. You can change the amount of output by using the --verbose value, where higher values produce more descriptive output.

[standard]Intel® Cluster Checker - Last Output Report
---------------------------------------------
This cluster was successfully certified at: 2010-07-13 15:36:02.
Log file: /home/user/config-certification-20100703.153650.xml

Log: /home/user/config-certification-20100703.153650.xml
Execution Type: Certification
Overall Status: Succeeded
Date: 2010-07-13 15:36:02
Version: 1.6
Command Line: /opt/intel/clck/1.6/cluster-check --certification 1.1
[/standard]
Parse the Output Log Files
Intel Cluster Checker 1.6 includes the source code for an XML parser. This code is provided as a reference for you to process output log files.

For an example of how to implement parsing of the log kernel messages, see the modules/dmesg.pm file inside the distributed package.

Set Extra Intel® MPI Library Options
The configuration file now allows you to provide extra Intel® MPI Library options by using an XML attribute in the configuration file. Specifically, you use an 'options' XML attribute for the <device> tag. The options will be reordered as required by MPI, placing global ones first.

For instance, in the examples below, the first line increases the verbosity of the MPI library run-time messages. The second specifies the TCP network to use. The third line enables the multi-rail fabric combination feature using MPI options, and the last example shows how to select the tag matching interface (TMI*) transport.

[standard]<device options="-genv I_MPI_DEBUG 5">ssm</device>
<device options="-genv I_MPI_TCP_NETMASK ib0">ssm</device>
<device options="-genv I_MPI_OFA_NUM_ADAPTERS 2">shm:ofa</device>
<device options="-genv I_MPI_TMI_LIBRARY /usr/lib/libtmi.so -genv I_MPI_TMI_CONFIG
/etc/tmi.conf -genv I_MPI_TMI_PROVIDER mx"> tmi </device>
[/standard]
Learn More
Please check the Release Summary for details on the many new version 1.6 features and improvements.

Refer to the Intel® MPI Library Reference Manual for more details on MPI device selection and the available configuration options.

______

Andres More is a software engineer at Intel - Argentina. Andres is a former college teacher and software consultant who was educated at the Universidad Nacional de La Plata and the Universidad Nacional de Córdoba.
For more complete information about compiler optimizations, see our Optimization Notice.