mpirun - what you get is what you want?


The following examples of the mpirun command show the different ways to request CPUs and how coprocessor mode (-mode CO) and virtual node mode (-mode VN) differ.

Request CPUs
mpirun		 -mode CO -mode VN
get tasks nodes used get tasks nodes used
-partition R000_Nx 32 32 64 32
-np 32 32 32 32 16
-np 64 64 128 64 32

Remarks to the examples:

  • In Communication coprocessor (CO) mode
    one CPU does all the computations, the other communication, 512MB memory
    • requesting 32 CPUs with -partition R000_Nx or -np 32 the requested number of CPUs is available and the whole partition is used

  • In Virtual node (VN) mode
    each CPU does computation and communication, each has 256MB memory
    • requesting a predefined partition R000_Nx the 64 CPUs are available and the whole nodebook is used
    • specifying -np 32 you get the requested 32 CPUs but half of the nodebook is left unused

  • If memory is not the problem and -mode VN is the best choice 
    mpirun -partition R000_Nx
    or 
    mpirun -np 64
    must be specified to get the maximum possible number of CPUs and to use the whole nodebook

  • Because the command 
    mpirun -np 64 -mode CO
    specifies a number of tasks larger than the supported size of a partition (32, 128 or n*512) the next higher partition is allocated.
last change 22.08.2005 | Ulrike Schmidt | Print