Parallel Monte Carlo Simulation of HIV Molecular Evolution in Response to Immune Surveillance - Parallel Computing for Bioinformatics and Computational Biology - page 39

rate is approximately 10 − 5 per nucleotide per replication cycle [19], and the recombi-
nation rate is two orders of magnitude higher, approximately 10 − 3 per nucleotide per
replication cycle for virus infecting T-cells [20]. The length of the HIV-1 replication
cycle is estimated at between 1 and 3 days [21].
2.4 PARALLELIZATION WITH MPI
The simulation model described earlier will benefit from parallelization by simply
having replicate simulations run simultaneously. Replicates have identical starting
conditions, but because of the stochastic nature of mutation, recombination, selection,
and other events, the states of replicate simulations diverge. This problem is referred
to as embarrassingly, or naturally, parallel because of the ease with which a parallel
programming model is implemented. More formally, this is an example of a single
program multiple data (SPMD) parallel programming model [1]. MPI is particularly
well suited to this programming model and was used to run replicate simulations in
parallel. MPI provides a set of calls (an interface) to a library of subroutines (Fortran)
or functions (C) that control communication between processors. Descriptions of the
MPI-1 and MPI-2 standards can be found at http: // www-unix.mcs.anl.gov / mpi / .A
good primer on MPI is the topic Using MPI [22].
2.4.1 Initialization and Termination of MPI
The basic parallelization strategy applied here is to run each replicate simulation on a
separate processor and afterward gather results from each process, summarize these,
and write a single output file to disk. Figure 2.6 gives the Fortran 90 pseudocode in the
main program for this approach. These instructions are executed on each processor.
Near the beginning of the code is a C preprocessor directive that begins with #
include . This directive identifies the MPI header file for Fortran, mpif.h , which
contains definitions of variables and constants used by MPI. For example, the default
communicator, MPI_COMM_WORLD , which defines the communication context and
set of all processes used, is defined in this header file. This directive must be used in
every program and subprogram that makes MPI calls. Invoking preprocessing usually
requires setting a compiler switch or identifying files for automatic preprocessing
with the suffix “.F” or “.F90,” as opposed to “.f” or “.f90” for files that do not require
preprocessing.
Then, after some variable declarations, MPI is initialized with a call to the MPI
subroutine MPI_INIT . This must be the first MPI call. The variable istat is used
to store the error status of the MPI call. Following this statement, two calls are made,
one to MPI_COMM_SIZE to query the number of processors being used ( npes ) and
the other to MPI_COMM_RANK to identify the processor number on which execution
is taking place ( mype ). The next statement identifies the replicate number being run
on the processor. Processors are counted from zero, so the replicate number is mype
+1 . This information is usedwhen summarizing the results from replicate simulations.