Qlogic Fast Fabric, Руководство пользователя

Страница: 34 / 174

3 – Getting Started
Running HPL
3-20 D000006-000 Rev A
Q
Create the file with a list of the chassis names (the TCP/IP Ethernet management
port names assigned above) or IP addresses (Use of names is recommended).
One entry per line. Such as:
Chassis1
Chassis2
For further details about the file format refer to the section “Selection of Chassis”
on page 5-4.
3. (All) Perform a health check using: all_analysis -e. If any errors are
encountered resolve the errors and rerun all_analysis -e until a clean run
occurs.
4. (All) Create a cluster configuration baseline using: all_analysis -b
5. (All) If desired, schedule regular runs of all_analysis via cron or other
mechanisms. Consult the Linux OS documentation for more information on
cron. Also consult the section “Health Check and Baselining Tools” on
page 5-69 for more information about all_analysis and its automated use.
3.11
Running HPL
As part of the installation process, a set of common MPI benchmarks have been
installed. One of the more popular measures of overall performance is HPL. This
is the application used to rate systems on the Top 500 list. The steps allow some
initial runs of HPL to be made and provide some initial baseline numbers. The
defaults provided should perform within 10%-20% of optimal HPL results for the
cluster. Tuning for that additional 10%-20% is beyond the scope of this document.
1. (Host) To run HPL, first select a configuration file appropriate to your cluster.
It is best to start with a small configuration to verify HPL has been properly
compiled:
a. cd /opt/iba/src/mpi_apps
b. /config_hpl 2t
will configure a two process test run of HPL.
2. (Host) Now create the file /opt/iba/src/mpi_apps/mpi_hosts listing the host
names of all the hosts. Depending of your selection of
VIADEV_PATH_METHOD in /opt/iba/src/mpi_apps/mpi.param.hpl the user
can specify Ethernet or IPoIB host names. The default config will allow either.
3. (Host) Now run HPL:
./run_hpl 2
Since this is a very small problem size the performance of the run will be much
lower than the potential of the machine. So do not worry about performance,
just whether or not the run was successful.