共计 5112 个字符，预计需要花费 13 分钟才能阅读完成。
XJCO3221 Coursework 2 1
School of Computing, University of Leeds
XJCO3221
Parallel computation
Coursework 2: Distributed memory parallelism with MPI
Deadline: 10am, Tuesday 19th April 2022
If you have any queries about this coursework please visit the MPI Discussion Forum on
Minerva (found in the Learning Resources folder for this module). If your query is not resolved
by previous answers, post a new message.
This piece of work is worth 20% of the final module grade.
Learning objectives
• Use collective communication to distribute the problem and accumulate the answer.
• Implement a binary tree using point-to-point communication.
• Perform timing runs for parallel scaling and interpret your findings.
Task
Your task is to implement an MPI-C program that performs matrix–vector multiplication in
parallel. That is, given the N × N matrix A and the N−vector x defined on rank 0, you need to
calculate b = Ax, i.e.
in parallel, and ensure the full N−vector b is on rank 0 at the end of the calculation.
For convenience, rather than store the matrix A as a two–dimensional float array, it is instead
stored as a one–dimensional array of size N2
. This ensures the rows of the matrix are stored
adjacent in memory, which makes the use of collective communication easier. With this choice,
the matrix element at row row and column col is A[row*N+col], and the serial code that performs
the multiplication is
i n t row , c o l ;
f o r (row=0; row<N; row++) {b [ row] = 0. 0 f ;
f o r (c o l =0; col<N; c o l++)
b [row] += A[row∗N+c o l] ∗ x [c o l] ;
}•••••••• 1
XJCO3221 Coursework 2 2
You are provided with code that initialises the matrix A and solution vector b on rank 0, with the
problem size N specified as a command line argument. It also allocates memory for the vector
x on all processes, and the per–process matrix A perProc and per–process b perProc. You will
need to make sure each process receives a copy of the full vector x, and a partition of A by rows.
Each process will then calculate the multiplication for their portion of the matrix, leaving the
answer in b perProc. The b perProc for each process will then need to be combined into the final
solution b on rank 0.
Once you have this working using collective communication, you should then implement an alternative
distribution of x to all processes that uses point-to-point communication in a binary tree
pattern. This method should be automatically called if the number of processes numProcs is a
power of 2; if it is not a power of 2, your code should continue to use your original version. Your
submitted code should therefore include two versions of the distribution of x, each being called
depending on the number of processes specified when launching with mpiexec.
When you have the final version of your code, perform timing runs on cloud-hpc1.leeds.ac.uk
to determine the parallel execution time as the number of processes is varied. The combination
of numbers of processes and nodes is given in the file readme.txt, and you should insert your
results and discussion into this file, and include it when submitting. You should also interpret
your results in terms of your understanding of MPI and architecture used to run batch jobs on
cloud-hpc1.leeds.ac.uk. For the times, you should use the time output already provided in the
code.
Guidance
The provided files are:
cwk2.c : Starting point for your solution.
cwk2 extras.h : Routines to allocate, deallocate, and initialise the problem.
Do not modify this file or avoid calling its functions.
readme.txt : Use to provide results of your timing runs, and interpretation.
makefile : A simple makefile (usage optional).
It is recommended that you proceed incrementally, testing your code after each stage of your
calculations. For the binary tree, it is useful to insert print statements showing which rank each
process is sending to or receiving from, to ensure that all sends have corresponding receives.
Point–to–point communication was covered in Lecture 9 and collective communication in Lecture
- For the binary tree, you may like to consider an upside–down version of second variant
considered in Lecture 11, as it is easier to code. You may find the following useful:
1<<n : Evaluates as 2n
.
if(n && ((n&(n-1))==0) ) : Evaluates as true if n is a power of 2.
int lev=1; while(1<<lev<=p)lev++; : Finds the number of levels lev in a binary
tree with p nodes in the first layer.
XJCO3221 Coursework 2 3
You are required to use cwk1 extras.h unaltered, but are expected to add new routines to cwk1.c
(this is the only file that you should edit).
Marks
There are 20 marks in total: - marks : Distribution of matrix A and solution vector b between processes.
- marks : Distribution of x to all processes, using a binary tree when the number of
processes is a power of 2. Max. 2 marks if there is no binary tree version. - marks : Correct parallel implementation of matrix–vector multiplication.
- marks : Results from timing runs and interpretation.
Note that you can achieve a first class mark without implementing the binary tree, in which case
your code should use the same method of distribution of x for all values of numProcs.
All submissions must compile and execute on cloud-hpc1.leeds.ac.uk.
Your timings should be taken using the batch submission system on cloud-hpc1.leeds.ac.uk.
A description of how to submit MPI jobs using slurm is provided in Worksheet 2.
Submission
Your are required to submit cwk1.c and readme.txt.
Do not modify cwk2 extras.h, or copy any of the content to another file and then
modify. This file will be replaced with a different version as part of the assessment, so it is
imperative that your code still calls these routines.
Disclaimer
This is intended as an individual piece of work and, while discussion of the work
is encouraged, what you submit should be entirely your own work. Code similarity
tools will be used to check for collusion, and online source code sites will be checked.
Ensure you retain the receipt for your submission as you may be asked to produce it
at a later date in the event of a disputed submission time.
The standard late penalty of 5% per day applies for work submitted after the deadli