Privatization

What is Privatization? Privatization, as used in this context, refers to the process of taking global variables and making them local to a specific task or thread within a shared-memory context. This allows for multiple tasks or threads to run in parallel, each with their own instance of the variable, without affecting other tasks or threads.

Why is Privatization Necessary? In traditional MPI programming models, global variables are shared among all processes, which can lead to data replication and memory issues when running on many-core architectures. Privatization helps alleviate these issues by making each task or thread have its own copy of the variable, reducing the need for inter-process communication and improving performance.

How is Privatization Achieved? Privatization in MPC is achieved through compiler-based transformations that extend the Thread-Local Storage (TLS) model. The __task keyword is introduced to create a new level between thread-local and process-level variables, allowing tasks to have their own context and share data with other tasks.

Compiling a Privatized Program

Privatization in MPC can be achieved by compiling a program with mpc_cc, a compiler wrapper similar to mpicc. This allows developers to utilize privatization features while still benefiting from the flexibility and scalability of MPI.

To compile a privatized program using mpc_cc, you can use this line:

mpc_cc -fmpc-privatize myprogram.c -o ./a.out

the -fmpc-privatize is not required since privatization is activated by default in MPC unless mpc-compiler-additions is disabled on build. To disable privatization you can use -fno-mpc-privatize. Let’s try with a C code that would be problematic on thread-based runtime without privatization:

#include <stdio.h>
#include <mpi.h>

int rank;

int main(int argc, char **argv){
    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    // here rank is not mixed up when running thread-based
    return 0;
}

The previous compilation command line ensures the following code behaves properly when launched with the following mpcrun options :

mpcrun -n=2 -p=1 ./a.out