wiki:AdaptParallelization

Adapting a model's parallelization for PDAF

Overview

Like many numerical models, PDAF uses the MPI standard for the parallelization. In the description below, we assume that the model is parallelized using MPI.

PDAF supports a 2-level parallelization: First, the numerical model can be parallelized and can be executed using several processors. Second, several model tasks can be computed in parallel, i.e. a parallel ensemble integration can be performed. This 2-level parallelization has to be initialized before it can be used. The templates-directory templates/ contains the file init_parallel_pdaf.F90 that can be used as a template for the initialization. The required variables are defined in mod_parallel.F90, which is stored int he same directory and can also be used as a template. If the numerical model itself is parallelized, this parallelization has to be adapted and modified for the 2-level parallelization of the data assimilation system generated by adding PDAF to the model. The necessary steps are described below.

Three communicators

MPI uses so-called 'communicators' to define sets of parallel processes. In order to provide the 2-level parallelism, three communicators need to be initialized that define the processes that are involved in different tasks of the data assimilation system. The required communicators are initialized in the routine init_parallel_pdaf and called

  • COMM_model - defines the processes that are involved in the model integrations
  • COMM_filter - defines the processes that perform the filter analysis step
  • COMM_couple - defines the processes that are involved when data are transferred between the model and the filter

The parallel region of an MPI parallel program is initialized by calling MPI_init. By calling MPI_init, the communicator MPI_COMM_WORLD is initialized. This communicator is pre-defined by MPI to contain all processes of the MPI-parallel program. Often it is sufficient to conduct all parallel communication using only MPI_COMM_WORLD. Thus, numerical models often use only this communicator to control all communication. However, as MPI_COMM_WORLD contains all processes of the program, this approach will not allow for parallel model tasks. In order to allow parallel model tasks, it is required to replace MPI_COMM_WORLD by an alternative communicator that is split for the model tasks. We will denote this communicator COMM_model. If a model code already uses a communicator distinct from MPI_COMM_WORLD, it should be possible to use that communicator.

/pics/communicators_PDAFonline.png
Figure 1: Example of a typical configuration of the communicators using a parallelized model. In this example we have 12 processes over all, which are distributed over 3 model tasks (COMM_model) so that 3 model states can be integrated at the same time. COMM_couple combines each set of 3 communicators of the different model tasks. The filter is executed using COMM_filter which uses the same processes of the first model tasks, i.e. COMM_model 1 (Fig. credits: A. Corbin)

Using COMM_model

Frequently the parallelization is initialized in the model by the lines:

      CALL MPI_Init(ierr)
      CALL MPI_Comm_Rank(MPI_COMM_WORLD, rank, ierr)
      CALL MPI_Comm_Size(MPI_COMM_WORLD, size, ierr)

(The call to MPI_init is mandatory, while the second an third line are optional) If the model itself is not parallelized, the MPI-initialization will not be present. Please see the section 'Non-parallel models' below for this case.

Subsequently, one can define COMM_model by adding

      COMM_model = MPI_COMM_WORLD

In addition, the variable COMM_model has to be declared in a way such that all routines using the communicator can access it. The parallelization variables of the model are frequently hold in a module. In this case, it is easiest to add COMM_model as an integer variable here. (The example declares COMM_model and other parallelization-related variables in mod_parallel.F90)

Having defined the communicator COMM_model, the communicator MPI_COMM_WORLD has to be replaced by COMM_model in all routines that perform MPI communication, except in calls to MPI_init, MPI_finalize, and MPI_abort. The changes described by now must not influence the execution of the model itself. Thus, after these changes, one should ensure that the model compiles and runs correctly.

Initializing the communicators

Having replaced MPI_COMM_WORLD by COMM_model enables to split the model integration into parallel model tasks. For this, the communicator COMM_model has to be redefined. This is performed by the routine init_parallel_init, which is supplied with the PDAF package. The routine should be added to the model usually directly after the initialization of the parallelization described above. The routine init_parallel_pdaf also defines the communicators COMM_filter and COMM_couple that were described above. The provided routine init_paralllel_init is a template implementation. Thus, it has to be adjusted for the model under consideration. In particular one needs to ensure that the routine can access the variables COMM_model as well as rank and size (See the initialization example above. These variables might have different names in a model). If the model defines these variables in a module, a USE statement can be added to init_parallel_pdaf as is already done for mod_parallel.

The routine init_parallel_pdaf splits the communicator MPI_COMM_WORLD and (re-)defines COMM_model. If multiple parallel model tasks are used, by setting n_modeltasks to a value above 1, COMM_model will actually be a set of communicators with one for each model task. In addition, the variables npes_world and mype_world are defined. If the model uses different names for these quantities, like rank and size, the model-specific variables should be re-initialized at the end of init_parallel_pdaf. The routine defines several more variables that are declared and held in the module mod_parallel. It can be useful to use this module with the model code as some of these variables are required when the initialization routine of PDAF (PDAF_init) is called.

Arguments of init_parallel_pdaf

The routine init_parallel_pdaf has two arguments, which are the following:

SUBROUTINE init_parallel_pdaf(dim_ens, screen)
  • dim_ens: An integer defining the ensemble size. This allows to check the consistency of the ensemble size with the number of processes of the program. If the ensemble size is specified after the call to init_parallel_pdaf (as in the example) it is recommended to set this argument to 0. In this case no consistency check is performed.
  • screen: An integer defining whether information output is written to the screen (i.e. standard output). The following choices are available:
    • 0: quite mode - no information is displayed.
    • 1: Display standard information about the configuration of the processes (recommended)
    • 2: Display detailed information for debugging

Compiling the extended program

This completes the adaptation of the parallelization. The compilation of the model has to be adjusted for the added files holding the routine init_parallel_pdaf and the module mod_parallel. One can test the extension by running the compiled model. It should run as without these changes, because mod_parallel defines by default that a single model task is executed (n_modeltasks=1). If screen is set to 1 in the call to init_parallel_pdaf, the standard output should include lines like

 Initialize communicators for assimilation with PDAF

                  PE configuration:
   world   filter     model        couple     filterPE
   rank     rank   task   rank   task   rank    T/F
  ----------------------------------------------------------
     0       0      1      0      1      0       T
     1       1      1      1      2      0       T
     2       2      1      2      3      0       T
     3       3      1      3      4      0       T

These lines show the configuration of the communicators. This example was executed using 4 processes and n_modeltasks=1. (In this case, the variables npes_filter and npes_model will have a value of 4.)

To test parallel model tasks one has to set the variable n_modeltasks to a value larger than one. Now, the model will execute parallel model tasks. For n_modeltasks=4 and running on a total of 4 processes the output from init_parallel_pdaf will look like the following:

 Initialize communicators for assimilation with PDAF

                  PE configuration:
   world   filter     model        couple     filterPE
   rank     rank   task   rank   task   rank    T/F
  ----------------------------------------------------------
     0       0      1      0      1      0       T
     1              2      0      1      1       F
     2              3      0      1      2       F
     3              4      0      1      3       F

In this example only a single process will compute the filter analysis (filterPE=.true.). There are now 4 model tasks, each using a single process. Thus, both npes_filter and npes_model will be one.

Using multiple model tasks can result in the following effects:

  • The standard screen output of the model can by shown multiple times. This is due to the fact that often the process with rank=0 performs screen output. By splitting the communicator COMM_model, there will be as many processes with rank 0 as there are model tasks.
  • Each model task might write file output. This can lead to the case that several processes try to generate the same file or try to write into the same file. In the extreme case this can result in a program crash. For this reason, it might be useful to restrict the file output to a single model task. This can be implemented using the variable task_id, which is initialized by init_parallel_pdaf and holds the index of the model task ranging from 1 to n_modeltasks. (For the ensemble assimilation, it can be useful to switch off the regular file output of the model completely. As each model tasks holds only a single member of the ensemble, this output might not be useful. In this case, the file output for the state estimate and perhaps all ensemble members should be done in the pre/poststep routine of the assimilation system.)

Non-parallel models

If the numerical model is not parallelized (i.e. serial), there are two possibilities: The data assimilation system can be used without parallelization (serial), or parallel model tasks can be used in which each model task uses a single process. Both variants are described below.

Serial assimilation system

The data assimilation program can be compiled for serial processing without linking a real MPI library. As in the PDAF code calls to MPI functions are implemented, the file nullmpi.F90 available in the directroy templates should be compiled and liked. An example for this gives the case make.arch/linux_gfortran.h. nullmpi.F90 provides the functionality of the MPI functions for the case that only a single process is used and hence no real communication is performed. A shortened header file mpif.h is provided in the directory src/dummympi. This file is in general not compatible with real MPI libraries and should only be used without parallelization.

Even without parallelization, the call to init_parallel_pdaf described above is still required. The routine will simple initialize the parallelization variables for a single-process case.

Adding parallelization to a serial model

In order to use parallel model tasks with a model that is not parallelized, the procedure is generally as described for the fully parallel case. However, one has to add the general initialization of MPI to the model code (or to init_parallel_pdaf). This is the lines

      CALL MPI_Init(ierr)
      CALL MPI_Comm_Rank(MPI_COMM_WORLD, mype_world, ierr)
      CALL MPI_Comm_Size(MPI_COMM_WORLD, npes_world, ierr)
      COMM_model = MPI_COMM_WORLD

together with the USE statement for mod_parallel should be added. Subsequently, the call to init_parallel_pdaf has to be inserted at the beginning of the model code. At the end of the program one should insert

    CALL  MPI_Barrier(MPI_COMM_WORLD,ierr)
    CALL  MPI_Finalize(ierr)

The module mod_parallel.F90 from the template directory provides subroutines for the initialization and finalization of MPI. Thus, if this module is used, the is no need to explicitly add the call to the MPI functions, but one can simply add

    CALL init_parallel()

at the beginning of the program. This has to be followed by

    CALL init_parallel_pdaf(dim_ens, screen)

to initialize the variables for the parallelization of PDAF. At the end of the program one should then insert

    CALL finalize_parallel()

in the source code.

If the program is executed with these extensions using multiple model tasks, the issues discussed in 'Compiling the extended program' can occur. This one has to take care about which processes will perform output to the screen or to files.

Last modified 4 years ago Last modified on Dec 11, 2019, 6:02:25 PM