Intel® MPI Library Reference Manual for Linux* OS
To use the Intel MPI Library on Intel® Xeon Phi™ Coprocessor (codename: Knights Corner), ensure that:
Each host and each Intel® Xeon Phi™ coprocessor must have a unique IP address and symbolic name, which is the same as to classic cluster.
Password-less access between host and Intel® Xeon Phi™ Coprocessor by ssh is established.
If the connection fails, check the possible causes and solutions in the following table:
Possible Cause | Solution |
---|---|
The version of Intel® MIC Software Stack you used is out of date. | Install a newer version. |
The iptables service is running on the host. | Stop that service. |
The route is incomplete. | Add the missing routes. |
Refer to the system administrator and Intel® MIC Software Stack readme to configure the settings for the IP connectivity.
When using Intel MPI Library on an Intel Xeon Phi coprocessor, consider Intel Xeon Phi coprocessor card to be another cluster node with a different Intel® architecture. The way that MPI features work for the Intel Xeon Phi coprocessor is similar to the way they work for an Intel® Xeon processor.
For example, MPI libraries may be available on both Intel Xeon processor and Intel Xeon Phi coprocessor through an NFS share that has the same path for Intel Xeon processor host and Intel Xeon Phi coprocessor; MPI tasks may be started from Intel Xeon processor host or Intel Xeon Phi coprocessor, and so on.
To build an application for running on the Intel Xeon Phi coprocessor and the host node, go through the following steps:
Establish environment settings for the compiler and for the Intel MPI Library:
(host)$ . <compiler_installdir>/bin/compilervars.sh intel64em64t
(host)$ . <mpi_installdir>/intel64em64t/bin/mpivars.sh
Build your application for Intel MIC Architecture, for example:
(host)$ mpiicc -mmic test.c -o test_hello.mic
Build your application for Intel 64 Architecture, for example:
(host)$ mpiicc test.c -o test_hello
To run an application on the Intel Xeon Phi coprocessor and the host node, go through the following steps:
Ensure that NFS is properly set-up between the hosts and the Intel Xeon Phi coprocessor, which is the recommended way for using Intel MPI Library on Intel MIC Architecture.
For information on how to set up NFS on the Intel Xeon Phi coprocessor, see http://software.intel.com/en-us/articles/intel-mpi-library-for-linux-kb/all/ or http://software.intel.com/mic-developer .
Establish environment settings for the Intel MPI Library:
(host)$ . <mpi_installdir>/intel64em64t/bin/mpivars.sh
Launch the executable file from host, for example:
(host)$ export I_MPI_MIC=1
(host)$ mpirun -n 2 -host <host ID> ./test_hello : -n 2 -host <coprocessor ID> ./test_hello.mic
See Intel® MPI Library for Linux* OS Reference Manual for -configfile, -hostfile and -machinefile options which also can be used.
To run the application on Intel Xeon Phi coprocessor only, follow the steps described above except for the step of building the application for Intel 64 Architecture. Meanwhile, ensure that the hostfile only contains the Intel Xeon Phi coprocessor name.
For more details, see http://software.intel.com/en-us/articles/intel-mpi-library-for-linux-kb/all/ and http://software.intel.com/mic-developer.