Scribd
Upload
21 views

Assignment 2

The document describes Assignment 2 which involves implementing an efficient parallel algorithm to solve the all-pairwise computation problem on a distributed-memory machine using MPI for communication between nodes and OpenMP for multi-threading within nodes. The algorithm must balance workload, minimize communication, and use loop unrolling. The implementation requires distributing a 2D input matrix across processes, computing dot products for pairs where i ≤ j, storing results in the upper triangle of a matrix, and conducting a self-check for correctness. A report on the algorithm, implementation, testing and issues is required.

Uploaded by

Alkesh Raj
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
21 views

Assignment 2

The document describes Assignment 2 which involves implementing an efficient parallel algorithm to solve the all-pairwise computation problem on a distributed-memory machine using MPI for communication between nodes and OpenMP for multi-threading within nodes. The algorithm must balance workload, minimize communication, and use loop unrolling. The implementation requires distributing a 2D input matrix across processes, computing dot products for pairs where i ≤ j, storing results in the upper triangle of a matrix, and conducting a self-check for correctness. A report on the algorithm, implementation, testing and issues is required.

Uploaded by

Alkesh Raj
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

In Assignment 1, you have designed and implemented an efficient parallel algorithm to solve

the all pairwise computation problem on shared-memory computing platforms and


implemented the algorithm using OpenMP. In Assignment 2, you are asked to design and
implement an efficient parallel algorithm to solve the same problem on a distributed-memory
machine. We assume that in this distributed-memory machine there are a number of processing
nodes and each processing node consists of a number of cores with shared memory. Thus, in
your algorithm design and implementation, you need to consider data/tasks distribution and
necessary communication among the processing nodes; you also need to consider how you will
assign tasks to different cores in each processing node. You will implement your algorithm
using MPI for data communication between processing nodes and OpenMP for thread
management and synchronization in each processing node. The all pairwise computation
problem is presented again below.

All Pairwise Computation is defined as performing computation (e.g., correlations) between


every pair of the elements in a given dataset. In this assignment, we consider a set of N
sequences, each being of length M, which are stored in a two-dimensional (2D) matrix of size
N by M (i.e., N rows and M columns). We then calculate a dot product for every possible pair
of sequences (row vectors) in the matrix.

A dot product of two sequences, or vectors is defined as


𝑀𝑀−1

𝑥𝑥 · 𝑦𝑦 = � 𝑥𝑥𝑖𝑖 𝑦𝑦𝑖𝑖
𝑖𝑖=0

If we allow a vector to pair with itself, then there are N(N+1)/2 such pairs in total for N vectors.
Assuming N is equal to 5 for example, we then have the following 15 vector pairs:

(0, 0)(0, 1)(0, 2)(0, 3)(0, 4)

(1, 1)(1, 2)(1, 3)(1, 4)

(2, 2)(2, 3)(2, 4)

(3, 3)(3, 4)

(4, 4)

In the above dot product formula, (i, j) denotes the pair of sequences i and j. After the dot
product computation for pair (i, j), we will have a single value as the output. Let 𝑎𝑎𝑖𝑖 and 𝑎𝑎𝑗𝑗 be
two sequences, namely the 𝑖𝑖 𝑡𝑡ℎ and 𝑗𝑗 𝑡𝑡ℎ row vectors in the 2D matrix, and 𝑐𝑐𝑖𝑖𝑖𝑖 = 𝑎𝑎𝑖𝑖 ∙ 𝑎𝑎𝑗𝑗 .

In your parallel algorithm design, you must consider how to (1) balance the workload, (2)
minimize communication overhead, and (3) use the loop unrolling technique with the unrolling
factor equal to 4 to improve the performance.

In your parallel algorithm implementation:


• Your program needs to ask for N (number of sequences) M (sequence length) and T
(number of threads in each processing node) as user defined parameters;
• In your program process 0 first creates an input two-dimensional matrix of size N by M
and then partitions and distributes the matrix evenly to other processes. As a result, the
matrix is evenly divided into p (the number of created processes) blocks of rows and
each process holds one block. It is not allowed for every process to store the whole
input matrix during the computation.
• Since the dot products of (i, j) and (j, i) will give the same result, you must only compute
the dot product of (i, j) for 𝑖𝑖 ≤ 𝑗𝑗 and store the results in the upper triangle of a 2D matrix
as shown in the above example;
• After the parallel computation, your main program must conduct a self-check for
correctness, i.e., perform a sequential computation using the same data set and then
compare the result of this computation with the parallel computation result.

Your implementation must be in C and use MPI and OpenMP as described earlier in this
assignment description.

You must write a report. The report must be concise, clear (3-6 A4 pages) and contain the
following sections:

1. Problem definition and requirements


2. Parallel algorithm design and implementation
3. Testing
4. Discussion
5. Known issues in program
6. Manual (e.g. how to run the program, input and output)

Your assignment will be marked on the efficiency of your algorithm, program logic and
readability, accuracy of results, and quality of your report.

You MUST attempt this assignment individually.

Submission Requirements

1. Your submission must be made by 11:59pm on Friday, 26 May, 2023 (Sydney time).
2. Create a tar or zip file that contains your report, makefile and source files (e.g., .c and .h
files). DO NOT INCLUDE ANY OBJECT OR BINARY FILES.
3. Submit only one .tar or .zip file.
Failure to follow these submission requirements may lead to loss of marks.

You might also like