11.5 I/O Performance Considerations

Before building and running your parallel application, I/O performance issues on the HP XC cluster must be considered.

The I/O control system provides two basic types of standard file system views to the application:

•Shared•Private

11.5.1Shared File View

Although a file opened by multiple processes of an application is shared, each core maintains a private file pointer and file position. This means that if a certain order of input or output from multiple cores is desired, the application must synchronize its I/O requests or position its file pointer such that it acts on the desired file location.

Output requests to standard output and standard error are line-buffered, which can be sufficient output ordering in many cases. A similar effect for other files can be achieved by using append mode when opening the file with the fopen system call:

fp = fopen ("myfile", "a+");

11.5.2 Private File View

Although the shared file approach improves ease of use for most applications, some applications, especially those written for shared-nothing clusters, can require the use of file systems private to each node. To accommodate these applications, the system must be configured with local disk.

For example, assume /tmp and/tmp1 have been configured on each compute node.

Now each process can open up a file named /tmp/myscratch or /tmp1/myotherscratch and each would see a unique file pointer. If these file systems do not exist local to the node, an error results.

It is a good idea to use this option for temporary storage only, and make sure that the application deletes the file at the end.

C example: fd = open ("/tmp/myscratch", flags)Fortran example: open (unit=9, file="/tmp1/myotherscratch" )

11.6 Communication Between Nodes

On the HP XC system, processes in an MPI application run on compute nodes and use the system interconnect for communication between the nodes. By default, intranode communication is done using shared memory between MPI processes. For information about selecting and overriding the default system interconnect, see the HP-MPI documentation.

11.7 Using MPICH on the HP XC System

MPICH is a freely available portable implementation of MPI. For additional information on MPICH, see the following URL:

http://www-unix.mcs.anl.gov/mpi/mpich1/ .

Verify with your system administrator that MPICH has been installed on your system. The HP XC System Software Administration Guide provides procedures for setting up MPICH.

MPICH jobs must not run on nodes allocated to other tasks. HP strongly recommends that all MPICH jobs request node allocation through either SLURM or LSF and that MPICH jobs restrict themselves to using only those resources in the allocation.

Launch MPICH jobs using a wrapper script, such as the one shown in Figure 11-1. The following subsections describe how to launch MPICH jobs from a wrapper script with SLURM or LSF,

11.5 I/O Performance Considerations 113