Scribd
Upload
162 views12 pages

Parallel Gaussian Elimination Methods

This document summarizes parallel Gaussian elimination algorithms for solving systems of linear equations. It describes the major steps of Gaussian elimination and Gauss-Jordan elimination. It then discusses different partitioning methods for parallelizing the algorithms, including static partitioning where the matrix is divided into fixed blocks and distributed to processors, and dynamic partitioning where the partitioning is adjusted at each step to balance the workload. Hybrid methods that combine static and dynamic approaches are also mentioned. The document concludes by assigning homework to implement a parallel Gaussian elimination algorithm using dynamic partitioning or a hybrid method.

Uploaded by

firostar
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
162 views12 pages

Parallel Gaussian Elimination Methods

This document summarizes parallel Gaussian elimination algorithms for solving systems of linear equations. It describes the major steps of Gaussian elimination and Gauss-Jordan elimination. It then discusses different partitioning methods for parallelizing the algorithms, including static partitioning where the matrix is divided into fixed blocks and distributed to processors, and dynamic partitioning where the partitioning is adjusted at each step to balance the workload. Hybrid methods that combine static and dynamic approaches are also mentioned. The document concludes by assigning homework to implement a parallel Gaussian elimination algorithm using dynamic partitioning or a hybrid method.

Uploaded by

firostar
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd

Parallel Gaussian Elimination

Major Steps in Gauss-Jordan


1. Transform Coefficient Matrix to Upper
Triangular form
2. Back-Substitution
Transforming to Upper Triangular
forin+1 to Mdo
x[ i] b[ i] /a[ i, i]
for j i+1 toM do
b[ j] b[ j] x[ i] × a [ j, i]
endfor
endfor
Back-Substitution
forin-1 to 0 do
for j I to N do
sum = x[j] * a[i,j]
end for
x[i] = b[i] - sum
endfor
Data-Dependency Graph

[Link]
Partition (PCAM)
1 0 2 2 0 6
Rank 0 Rank idle
0 1 2 0 -3 0

0 0 22 -1 0 4
Rank 1
0 0 2 13 0 -5

0 0 0 1 16 0

Rank 2
0 0 5 3 5 -32

Static Partition Method


Steps – Static Partition Method
1. n number of rows in each stripe
2. Scatter matrix to ranks
3. for i 0 to rows-1
1. Adjust pivot row
2. Broadcast pivot row
3. Sweep matrix
4. Gather matrix
5. Solve via back-substitution
Partition (PCAM)
1 0 2 2 0 6
Completed
0 1 2 0 -3 0 rows stored
Rank 0 0 0 22 -1 0 4

Rank 1 0 0 2 13 0 -5

0 0 0 1 16 0
Rank 2
0 0 5 3 5 -32

Dynamic Partition Method


Steps – Dynamic Partition Method
1. for i 0 to rows-1
1. n number of rows per stripe ((rows-i)/ranks)
2. Scatter matrix to ranks
3. Adjust pivot row
4. Broadcast pivot row
5. Sweep matrix
6. Gather matrix
2. Solve via back-substitution
Static vs Dynamic
Static Dynamic
• Broadcast once • Broadcast for each row
• Gather once • Gather for each row
• Ranks idle after stripe • Load re-balanced at each
completed step
• Simpler code ?? • Code more complex??
Hybrid Methods
• Bandwidth Reduction (Gibbs-Poole-
Stockmeyer algorithm)
– [Link]
rithms/BOOK/BOOK3/[Link]
– [Link]
• Simpler solutions?
– Overlapping communications with computations
– Mix Static and Dynamic Methods
Homework
Use Gaussian Elimination to solve AX=B

• A ‘[Link]’ (first NxN values)


– 250x250, 1000x1000, and 3000x3000
• B ‘[Link]’ (multiple Right Hand Sides)
• X matrix of solution vectors
• Use Dynamic Partitioning or a Hybrid Method
• Run with 24, 96, 192 cores. (total of nine tests)
• Turn in code, sum of solution matrix, plot of
flops per test. (grouped bar chart)

You might also like