Analyst solvers can be run on multiple processors within a single computer, multiple processors on a network, or on a dedicated cluster running either Windows or LINUX operating systems.
Analyst solvers use the Message Passing Interface (MPI) for interaction between parallel processes. Therefore you must download and install MPICH 1.2.5 developed at Argonne National Laboratory before more than one process can be used by a solver. MPICH 1.2.5 can be downloaded from our site via this link (use your support username/password to access). This needs to be installed on all machines that will be involved in the parallel analyses (INCLUDING the machine that Analyst is on). When installing the package, you can use the default installation path and properties (it should NOT be put in the Analyst installation directory).
Additionally, for LINUX cluster runs only, you need to have Windows SSH client executables plink.exe and pscp.exe in the Analyst the bin subdirectory of the Analyst install (only on the Analyst machine), and you must have accounts/permissions to launch jobs interactively on the cluster. The SSH client executables can be obtained directly from this link. Once they are present, you will need to open a DOS Command Prompt and change to the bin directory. Run “plink machname” (where machname is the hostname of the head node of the cluster, or the IP address of the head node). You will need to answer YES to the question about storing the server’s host key in the registry. Failure to perform this step will result in failures to launch the solver. You will also need to request a solver excutable for the LINUX platform from support@staarinc.com.
Configure the solver to run in parallel using the Parallelization tab on the solver setup dialog.
You may select between a Windows - Local, Windows - Network, and Unix Cluster analysis in this dialog. For multiprocessing on the machine that Analyst is licensed on, select Windows - Local, and enter the number of processors you want to include. To use other computers running Windows on the local network, select “Windows – Network”. For an analysis on a network of Windows machines, select the Network radio button. Note that this option may be disabled if your temporary path is not set properly (see the Path Settings dialog accessible from the Analyst main menu). You then need to specify a list of machines to run the analysis on. These machines must have either Analyst or the STAAR Remote Analysis Service installed. The list of machine should always include your local machine. The format for this list is:
Mach1:NumProcs, Mach2:NumProcs, Mach3, Mac4:NumProcs...
For example:
STAR5:2, STAR4:1,MOBILE1, STAR6:1
You can optionally follow each machine name by a colon and the number of processors. One-processor is the default unless otherwise specified. You can also set user-defined load balancing. After the number of processors for a specific machine, you can specify a load balance fraction. For example:
STAR5:2:0.7, STAR4:1:0.4,MOBILE1, STAR6:1
If a load balance fraction is specified for some machines, all unspecified values will be 1.0.
You must also provide login information (username and password) for the run. This information does not need to match the current user login, but it must be valid for all of the machines in the selection set. Note that this login information is always stored in an encrypted format for security.
To run on a remote LINUX machine or cluster that is SSH-enabled, check the Unix Cluster radio button on the solver setup Parallelization tab. Now pick the cluster from the list, and enter the number of processors to use. To create a new cluster entry, click the "Edit" button and fill in the required fields.