Scribd
Upload
4 views

Assignment_4

The assignment involves multiplying N sparse matrices using MPI, OpenMP, and CUDA, with a focus on parallelization across multiple nodes and CPU cores. Students must implement a CUDA kernel for matrix multiplication and follow specific input/output formats for the matrices. A report detailing the parallelization strategy and performance analysis is also required alongside the source files and a Makefile for execution.

Uploaded by

Aneeket1 Yadav
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
4 views

Assignment_4

The assignment involves multiplying N sparse matrices using MPI, OpenMP, and CUDA, with a focus on parallelization across multiple nodes and CPU cores. Students must implement a CUDA kernel for matrix multiplication and follow specific input/output formats for the matrices. A report detailing the parallelization strategy and performance analysis is also required alongside the source files and a Makefile for execution.

Uploaded by

Aneeket1 Yadav
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

COL380 Assignment 4

(2024-25 Sem II)

You are given N sparse matrices Ai, 0 ≤ i < N (of the kind given in Assignment 1 – You may employ
the generate_matrix function from assignment 1 for testing as well). Your task is to multiply this
sequence of matrices: produce ∏Ai in the same format as the input matrices (all integers are 64-bit
long long and all operations are modulo maximum long long). This sequence multiplication is to be
parallelized across multiple nodes (< 8), each with multiple CPU cores (up to 16) and 1 GPU.
Combine MPI, OpenMP, and CUDA (same versions as you have used in the earlier assignments).

You will need to implement a CUDA kernel for multiplying two sparse matrices. You may re-use
the OpenMP parallelization of matrix multiplication. You will distribute these pair-wise
multiplications on all available MPI processes.

The command line of your program has one input string: the path of a readable folder, call it F.
Folder F will contain N+1 files, named matrix1 .. matrixN, and size, respectively. The file
F/size has a two integers (separated by white-spaces). The first denotes N and the second denotes
k, the block size used for all matrices in this folder. Each file matrixi has the following format:

<Height:int> <Width:int>
<Number of non-zero blocks: int>
For each non-zero block:
<Block starting row number:int> <Block starting column number:int>
For each non-zero block (in the same order as above)
k*k <value:int>

Correct matrix files ensure that <Width> value in matrixi equals <Height> value in
matrixi+1. It is possible (if rare) that a matrix’s width or height is not divisible by k – in that case,
the last block will have fewer than k elements.

The output will be a single matrix produces in the execution directory with the name matrix in
the same format as input matrices. (Note that F may be read-only – do not attempt to write there.)

For performance tuning, expect input matrix sizes to be around s*1024 (with a small single digits s).
N may be up to a few hundred. The value of k would be small (e.g., 4, 16, 32). Fewer than 10% of
the values in each matrix would be non-zero.

Note that finding the most efficient order of multiplications in the parallel context is a hard
problem. Consider greedy heuristics to reduce the number of operations. For example, if
multiplying a pair of consecutive matrices produces fewer non-zero blocks, it may be multiplied
first. Remember, the time to compute the heuristic ordering is included in the total multiplication
time.

Submission:
Submit any number of source files, but must also include a Makefile, which produces an
executable a4. The tester will call make, followed by a4 foldername (of course, with
appropriate mpirun).
In addition to the source file, also submit a report in file report.pdf (only for part 2) explaining
your parallelization strategy. Include your performance analysis in the report itself. (No csv files are
required.) The total execution time will include file IO times (and will be determined by the slowest
MPI process). You may also separately analyze the relative performances of your CUDA and
OpenMP parts.

You might also like