Description
The purpose of this assignment is to gain hands-on experience in practical use of message passing
interface (MPI) in real-world applications, thereby consolidating your understanding of
asynchronous message passing and communicators. In addition, you will get familiar with the
message-passing scheme used in common spatial-decomposition applications, using the parallel
molecular dynamics (MD) program, pmd.c, as an example.
(Part I—Asynchronous Messages)
Modify pmd.c such that, for each message exchange, it first calls MPI_Irecv, then MPI_Send, and
finally MPI_Wait. The asynchronous messages make the deadlock-avoidance scheme unnecessary,
and thus there is no need to use different orders of send and receive calls for even- and odd-parity
processes. In addition to just MPI_Send, insert other computations that do not depend on the
received messages between MPI_Irecv and MPI_Wait.
• Submit the modified source code, with your modifications clearly marked.
• Run both the original pmd.c and the modified program on 16 cores (requesting 4 nodes with 4
cores per node in your Slurm script), and compare the execution time for InitUcell = {3,3,3},
StepLimit = 1000, and StepAvg = 1001 in pmd.in (keep all the other parameter values as
they are as downloaded from the course home page) and vproc = {2,2,4} (i.e., nproc = 16) in
pmd.h. Which program runs faster? Repeat the comparison three times and report the average
runtime of both programs. Submit the timing data.
(Part II—Communicators)
Following the lecture note on “In situ analysis of molecular dynamics simulation data using
communicators,” modify pmd.c such that as many number of processes as that for MD simulations
is spawned to calculate the probability density function (PDF) for the atomic velocity.
• Submit the modified source code (name it pmd_split.c), with your modifications clearly
marked.
• Run the modified program on 16 cores (requesting 2 nodes with 8 cores per node in your Slurm
script), with which 8 cores perform MD simulation and the other 8 cores calculate PDF. In
pmd.h, choose vproc[3] = {2,2,2} and nproc = 8. Also, specify InitUcell = {5,5,5},
StepLimit = 30, and StepAvg = 10 in pmd.in. Submit the plot of calculated PDFs at time
steps 10, 20, and 30.
