Implementing Tasks in OpenMP/MPI C++.
The tasks are assignements for Parallel Programing Course at ITMO university.
The program aims to count words in a line. Any sequence of characters without separators is considered a word. Separators are spaces, tabs, newlines.
The main goal of the program is to find the maximum value of a vector. In order to achieve that, I use the OpenMP reduction clause. The program also studies the dependence of the runtime on the number of threads used (from 1 to 10) for a vector that contains 1,000,000 elements.
The program implements two dimension matrix multiplication with different loop orders. For each loop order, the computation is done for one thread then for multiple threads from 2 to 10. For each case the time of execution and the efficiency are measured. The arrays are shared among threads, meanwhile the indices are private to prevent different threads from editing them simultaneously.
MPI program that implements the dot product of two vectors distributed between processes. Two vectors with a size of at least 1,000,000 elements are initialized at process zero and filled with “1”, then they are sent in equal parts to all processes. Parts of vectors are scalar multiplied on each process, the result is sent to the root process and summed up. The total is displayed
MPI program in which two processes exchange messages, the time per exchange iteration is measured, and the dependence of the exchange time on the message length is studied. The latency and maximum achievable bandwidth of the communication network are calculated. The message length in bytes and the throughput in MB/s is displayed. The length of the message is changed in a loop starting from 1 element and increase to 1,000,000 elements, increasing by 10 times at each iteration.
MPI program in which the global vector addition operation is modeled by a doubling (cascade) scheme using point-to-point data transfers. Another execution is done using the MPI_Reduce procedure on as many processes as possible.
Each process stores an array of 1,000,000 elements equal to ‘1’
MPI program for ring topology exchange using the MPI_Sendrecv() function.
8.Delayed interactions. Scheme of an iterative method with exchange over a ring topology using pending requests - MPI
MPI allows for non-blocking operations to form whole packets of requests for communication operations MPI_Send_init and MPI_Recv_init, which are started by the MPI_Start or MPI_Startall functions. Checking for completion of execution is performed by conventional means using the functions of the WAIT and TEST families.
Finding out which process will perform the multiplication of two 500x500 square matrices faster
In the program each thread initializes integer array. The array values are generated in a way related to the threas’s rank number. In order to gather all the arrays in the root thread, the function MPI_Reduce was used. A function max_fun() was initialized to find the maximum value at each position of array. In order to check the correctness of execution, the MPI_MAX operation in the MPI_Reduce() function was used.
In order to implement the tasks, several functions were used: MPI_File_open() to open a file. MPI_File_write() to write the content in the file. MPI_File_close() to close the opened file. MPI_File_set_view() to select where to start typing inside the file. MPI_File_delete() to delete a file.