0% found this document useful (0 votes)

53 views

Parallel Processing

This document describes parallel algorithms for matrix multiplication and analyzes their performance. It first describes a traditional algorithm that divides the matrix into rows and performs parallel multiplication. Cannon's algorithm is also described, which shifts submatrices to align for multiplication. Performance is analyzed by varying matrix size and number of processors. Results show computation time increases with matrix size while communication time increases with more processors. Cannon's algorithm has higher computation time than the traditional approach.

Uploaded by

prv142

Available Formats

Download as DOCX, PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

53 views

Parallel Processing

Uploaded by

prv142

Available Formats

Download as DOCX, PDF, TXT or read online on Scribd

You are on page 1/ 35

Assignment 3

Course EC 566 Parallel Processing

Fall 2013

Matrix Multiplication

Parallel algorithms for multiplying two n x n dense, square matrices.

Traditional approach is used with simple modified parallelization
Cannons algorithm is implemented for matrix multiplication.
Product of A ^ k and Determinant are calculated in both approaches .

Traditional Algorithm

Matrix is divided into 1-D row parallelization is implemented.

The matrix is divided into p rows, each of size (n/p) x n.
This modified approach is chosen rather than the normal block matrix
operations .
the normal traditional algorithm, the n x n matrix is divided into p sub-block
matrices, each of (n/p) x (n/p).
For this approach number of processors used is always a perfect square.
Any other number would not divide the matrix properly into sub-block
matrices.
The 1D row parallelization can help us in the division issue.

Modified approach is implemented as below,

Scatter the agglomerated row to all p processes.

Perform all-to-all broadcast to acquire other process elements for
multiplication. MPI_Allgather operation is used for the all-to-all broadcast.
Perform sequential multiplication at each process to calculate the product.
Gather all the products at the root for the multiplied result.

Cost Analysis

With 1-D row agglomeration, all-to-all broadcast requires (p-1) steps of

communication between neighbors as in a 1-D array or a ring.
Each step involves a message of size m, which is (n 2p), takes (ts + m*tw).
Summing over p - 1 steps
Tcomm = (ts + n2p * tw) (p 1)
Each process computes the multiplied sub-result, which involves (n 2p * n)
multiplications.
Tcomp = n3 p
Total parallel run-time is Tp ~ n3 p + (ts + n2p * tw) (p 1)

Results & Graphs .

According to the problem,

n - the size of the matrix is varied from 64 to 512.

k - power of A is varied from 2 to 6
p - number of processors are varied from 2 to 8.

Table for 2 processors, size of the matrix varying from 64 to 512, Power of matrix
varying from 2 to 6 .

Total time VS Matrix Multiplication power for matrix size varying from 64 to 512
25
20
15

512
256
128

64
5
0
2

Communication time VS Matrix Multiplication power for matrix size varying from 64
to 512

0.2
0.18
0.16
0.14
0.12

512

0.1

256

0.08

128
64

0.06
0.04
0.02
0
2

Computation time VS Matrix Multiplication power for matrix size varying from 64 to
512 .
25
20
15

512
256
128

64
5
0
2

Table for 4 processors, size of the matrix varying from 64 to 512, Power of matrix
varying from 2 to 6

Total time VS Matrix Multiplication power for matrix size varying from 64 to 512
18
16
14
12
512

256

128
64

6
4
2
0
2

Communication time VS Matrix Multiplication power for matrix size varying from 64
to 512
6.000000
5.000000
4.000000
512
256

3.000000

128
64

2.000000
1.000000
0.000000
2

Computation time VS Matrix Multiplication power for matrix size varying from 64 to
512.
12.000000
10.000000
8.000000
512
256

6.000000

128
64

4.000000
2.000000
0.000000
2

Table for 8 processors, size of the matrix varying from 64 to 512, Power of matrix
varying from 2 to 6

Total time VS Matrix Multiplication power for matrix size varying from 64 to 512
18
16
14
12
512

256

128
64

6
4
2
0
2

Communication time VS Matrix Multiplication power for matrix size varying from 64
to 512
9.000000
8.000000
7.000000
6.000000
512

5.000000

256

4.000000

128
64

3.000000
2.000000
1.000000
0.000000
2

Computation time VS Matrix Multiplication power for matrix size varying from 64 to
512.
18
16
14
12
512

256

128
64

6
4
2
0
2

Total time VS Matrix Multiplication power for matrix size 64 and no of processors
varying from 4 to 8.
0.03
0.03
0.02
Processors2
Processors4

0.02

processors8
0.01
0.01
0
1

Computation time VS Matrix Multiplication power for matrix size 64 and no of

processors varying from 4 to 8.
0.01
0.01
0.01
0.01

Processors2
Processors4

processors8

0
0
0
0
1

Communication time VS Matrix Multiplication power for matrix size 64 and no of

processors varying from 4 to 8.
0.02
0.02
0.02
0.01
0.01

Processors2

0.01

Processors4

0.01

processors8

0.01
0
0
0
1

Analysis

Computation time is reduced when the processors are increased.

Communication time increased when the processors are increased
Communication time is negligible when compared to the computation time.
Total time of the algorithm decreases as the processors increased.
As the size of the matrix increases the communication, computation time
increases drastically.

Merits anda Demerits .

This algorithm can use a max of n processors.

It requires (n2/p) memory .
Each process adds (n2) memory.
Isoefficiency function of this algorithm is (p3/2).

Cannons Algorithm

Cannon's algorithm uses a mesh of s ^2 processes that are connected as a

torus.
Process (i,j) at location (i,j) initially begins with submatrices A {i,j} and B{i,j}.
Asthe algorithm progresses, the submatrices are passed left and upwards.
Initially P{i,j} begins with A{i,j} and B{i,j}.
Elements are moved from their initial positions to align them so that the
correct submatrices are multiplied with one another.
Note that submatrices on the diagonal don't actually require alignment.
Alignment is done by shifting the ith row of A by i positions left and the jth
column of B by j positions up.
Each process, P{i,j} multiplies its current submatrices and adds to a cumulative
sum.
The submatrices are shifted left and upwards.
The above two steps are repeated through the remaining submatrices.

Cost Analysis

The initial shift operation requires a time of 2 (t s + tw * n2/p).

Each p single step shift takes (ts + tw * n2/p) time.
The total communication time is given as,
Tcomm = 2 (ts + tw * n2/p) * p
Each process performs p multiplications of ( n/p * n/p) submatrices.
Tcomp = n3/p
Total parallel run time of this algorithm is,
Tp = 2 p (ts + tw * n2/p) + n3/p

Results & Graphs .

According to the problem,

n - the size of the matrix is varied from 64 to 512.

k - power of A is varied from 2 to 6

Table for 4 processors, size of the matrix varying from 64 to 512, Power of matrix
varying from 2 to 6 .

Total time VS Matrix Multiplication power for matrix size varying from 64 to 512
100
90
80
70
60

512

256

128
64

30
20
10
0
2

Communication time VS Matrix Multiplication power for matrix size varying from 64
to 512
30
25
20
512
256

128
64

10
5
0
2

Computation time VS Matrix Multiplication power for matrix size varying from 64 to
512.
70
60
50
512

256
30

128
64

20
10
0
2

Analysis

Computation time is increased when the Matrix size is increased.

Communication time increased when the Matrix size is increased

Communication time is negligible when compared to the computation time.
Total time of the algorithm increases as the Matrix size is increased.
As the size of the matrix increases the communication, computation time
increases drastically.

Merits anda Demerits .

This algorithm can use a max of n processors.

It requires (n2/p) memory .
Each process adds (n2) memory.
Isoefficiency function of this algorithm is (p3/2).

Code for Traditional Algorithm

#include <mpi.h>
#include <stdio.h>
#define n 64 // size of matrix
#define pow 2 // power of matrix A
int main (int argc, char *argv[])
{
int prcss, id;
int prv, nxt;
int rw, clm;
int nrw1, nclm1, brw1, bclm1;
int i, j, k, x, y;
float det1 = 2, det2 = 1;
double cst, cend, cdiff, commt = 0, compt = 0;
rw = n;
clm = n;
MPI_Status stat;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &id);
MPI_Comm_size(MPI_COMM_WORLD, &prcss);
// Ring structure
prv = id - 1;
nxt = id + 1;
if (id == 0)
prv = prcss - 1;
if (id == (prcss - 1))

nxt = 0;
nrw1 = prcss;
nclm1 = 1;
brw1 = n/prcss;
bclm1 = n;
// building matrix A
float a[rw*clm];
if (id == 0)
{
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
{
if (i == j)
a[i + i*clm] = 1;
else
a[j + i*clm] = 0;
}
}
a[0] = 2;
}
// defining data-type
MPI_Datatype blocktype1;
MPI_Datatype blocktype2;
MPI_Type_vector(brw1, bclm1, clm, MPI_INT, &blocktype2);
MPI_Type_create_resized(blocktype2, 0, sizeof(int), &blocktype1);
MPI_Type_commit(&blocktype1);
int disp1[nrw1*nclm1];
int count1[nrw1*nclm1];
for (i = 0; i < nrw1; i++)
{
for (j = 0; j < nclm1; j++)
{
disp1[j + nclm1*i] = j*bclm1 + i*clm*brw1;
count1[j + nclm1*i] = 1;
}
}
// horizontal scatter operation
float b[brw1*bclm1];

cst = MPI_Wtime();
MPI_Scatterv(a, count1, disp1, blocktype1, b, brw1*bclm1, MPI_FLOAT, 0,
MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
cend = MPI_Wtime();
cdiff = cend - cst;
commt = commt + cdiff;
// horizontal all-gather operation
float c[rw*clm];
cst = MPI_Wtime();
MPI_Allgather(&b, brw1*bclm1, MPI_FLOAT, &c, brw1*bclm1, MPI_FLOAT,
MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
cend = MPI_Wtime();
cdiff = cend - cst;
commt = commt + cdiff;
// multiplication
float p[brw1*bclm1];
cst = MPI_Wtime();
for (y = 1; y < pow; y++)
{
for (i = 0; i < brw1; i++)
{
for (j = 0; j < bclm1; j++)
{
p[j + i*bclm1] = 0;
for (k = 0; k < bclm1; k++)
p[j + i*bclm1] = p[j + i*bclm1] + ( b[k + i*bclm1] *
c[j + k*bclm1] );
}
}

for (i = 0; i < brw1; i++)

{
for (j = 0; j < bclm1; j++)
b[j + i*bclm1] = p[j + i*bclm1];
}
}
cend = MPI_Wtime();
cdiff = cend - cst;

compt = compt + cdiff;

// gather Product
float mul[rw*clm];
cst = MPI_Wtime();
MPI_Gather(&p, brw1*bclm1, MPI_FLOAT, &mul, brw1*bclm1, MPI_FLOAT, 0,
MPI_COMM_WORLD);
cend = MPI_Wtime();
cdiff = cend - cst;
commt = commt + cdiff;
//Result Matrix
if (id == 0)
{
// Matrix A
printf("Initial Matrix A is\n");
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
printf(" %5.0f ", a[j + i*clm]);
printf("\n");
}
printf("\nMultiplied Matrix A^%d is\n", pow);
if (pow > 1)
{
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
printf(" %5.0f ", mul[j + i*clm]);
printf("\n");
}
}
if (pow == 1)
{
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
printf(" %5.0f ", a[j + i*clm]);
printf("\n");
}
}

// printing Determinant
det2 = mul[0];
printf("\nDet of A is %5f\nDet of A^%d is %5f\n", det1, pow, det2);
// printing Time
printf("Total communication time is %f\n", commt);
printf("Total computation time is %f\n", compt);
printf("Total parallel run time is %f\n", commt+compt);
}
MPI_Barrier (MPI_COMM_WORLD);
MPI_Finalize();
}
Output of Traditional Algorithm.
Initial Matrix A is
2
0
0

0
0
0

1
0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
1
0

0
0
0

0
1
0

0
0
0

0
0

0
0
0

0
0

1
0
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0
1

0
0
0

0
0
1

0
0
0

0
0
1

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
1

0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
1

0
0

0
0
0

0
0

0
0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

Multiplied Matrix A^2 is

4
0
0

0
0
0

1
0
0

0
0
0

1
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
1
0

0
0
0

0
0

1
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0
1

0
0
0

0
0

0
0
0

0
1

0
0

0
0
0

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0
1

0
0
0

0
0

0
0
1

0
0

Determinant of A is 2.000000
Determinant of A^2 is 4.000000
Total communication time is 0.001938
Total computation time is 0.001689
Total Run time is 0.003627
Code for Cannons Algorithm
#include <mpi.h>
#include <stdio.h>
#define n 64 // size of matrix
#define pow 2 // power of matrix A
int main (int argc, char *argv[])
{
int prcss, id;
int prv, nxt;
int rw, clm;
int nrw1, nclm1, brw1, bclm1;
int i, j, k, x, y;
int sqrt;
float det1, det2;
double csrt, cend, cdiff, commt = 0, compt = 0;
rw = n;
clm = n;
MPI_Status stat;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &id);
MPI_Comm_size(MPI_COMM_WORLD, &prcss);
// ring structure
prv = id - 1;
nxt = id + 1;
if (id == 0)
prv = prcss - 1;
if (id == (prcss - 1))

0
0

nxt = 0;
if (prcss == 4)
sqrt = 2;
else
sqrt = 4;
nrw1 = sqrt;
nclm1 = sqrt;
brw1 = n/sqrt;
bclm1 = n/sqrt;
// building matrix A & B
float a[rw*clm];
float b[rw*clm];
float temp1[rw*clm];
float temp2[rw*clm];
float temp3[rw*clm];
float mul[rw*clm];
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
{
if (i == j)
a[i + i*clm] = 1;
else
a[j + i*clm] = 0;
temp1[j + i*clm] = a[j + i*clm];
temp3[j + i*clm] = a[j + i*clm];
}
}
a[0] = 2;
temp1[0] = 2;
temp3[0] = 2;
det1 = 2;
// defining data-type
MPI_Datatype blocktype1;
MPI_Datatype blocktype2;
MPI_Type_vector(brw1, bclm1, clm, MPI_LONG, &blocktype2);
MPI_Type_create_resized(blocktype2, 0, sizeof(int), &blocktype1);
MPI_Type_commit(&blocktype1);

int disp1[nrw1*nclm1];
int count1[nrw1*nclm1];
for (i = 0; i < nrw1; i++)
{
for (j = 0; j < nclm1; j++)
{
disp1[j + nclm1*i] = j*bclm1 + i*clm*brw1;
count1[j + nclm1*i] = 1;
}
}
float b1[brw1*bclm1];
float c1[brw1*bclm1];
float p[brw1*bclm1];

for (y = 1; y < pow; y++)

{
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
{
a[j + i*clm] = temp1[i*clm + ((i+j)%clm)];
b[j + i*rw] = temp3[j + ((i+j)%rw)*rw];
}
}
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
{
temp1[j + i*clm] = a[j + i*clm];
temp2[j + i*clm] = b[j + i*clm];
}
}
// scatter operation
csrt = MPI_Wtime();
MPI_Scatterv(&a, count1, disp1, blocktype1, &b1, brw1*bclm1,
MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
cend = MPI_Wtime();
cdiff = cend - csrt;

commt = commt + cdiff;

csrt = MPI_Wtime();
MPI_Scatterv(&b, count1, disp1, blocktype1, &c1, brw1*bclm1,
MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
cend = MPI_Wtime();
cdiff = cend - csrt;
commt = commt + cdiff;
csrt = MPI_Wtime();
for (i = 0; i < brw1; i++)
{
for (j = 0; j < bclm1; j++)
{
p[j + i*bclm1] = 0;
p[j + i*bclm1] = p[j + i*bclm1] + ( b1[j + i*bclm1] * c1[j
+ i*bclm1] );
}
}
cend = MPI_Wtime();
cdiff = cend - csrt;
compt = compt + cdiff;

for (x = 1; x < rw; x++)

{
csrt = MPI_Wtime();
for (i = 0; i < n; i++)
{
for (j = 0; j < n; j++)
{
a[j + i*clm] = temp1[(j+1)%clm + i*clm];
b[j + i*rw] = temp2[j + ((1+i)%rw)*rw];
}
}
cend = MPI_Wtime();
cdiff = cend - csrt;
compt = compt + cdiff;
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
{
temp1[j + i*clm] = a[j + i*clm];
temp2[j + i*clm] = b[j + i*clm];

}
}
// scatter operation
csrt = MPI_Wtime();
MPI_Scatterv(&a, count1, disp1, blocktype1, &b1, brw1*bclm1,
MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
MPI_Scatterv(&b, count1, disp1, blocktype1, &c1, brw1*bclm1,
MPI_FLOAT, 0, MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
cend = MPI_Wtime();
cdiff = cend - csrt;
commt = commt + cdiff;
csrt = MPI_Wtime();
for (i = 0; i < brw1; i++)
{
for (j = 0; j < bclm1; j++)
p[j + i*bclm1] = p[j + i*bclm1] + ( b1[j + i*bclm1] *
c1[j + i*bclm1] );
}
cend = MPI_Wtime();
cdiff = cend - csrt;
compt = compt + cdiff;
}
// gather operation
csrt = MPI_Wtime();
MPI_Gatherv(&p, brw1*bclm1, MPI_FLOAT, &mul, count1, disp1,
blocktype1, 0, MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
MPI_Scatter(&mul, rw*clm, MPI_FLOAT, &temp1, rw*clm, MPI_FLOAT, 0,
MPI_COMM_WORLD);
MPI_Barrier (MPI_COMM_WORLD);
cend = MPI_Wtime();
cdiff = cend - csrt;
commt = commt + cdiff;
}
if (id == 0)
{

printf("Initial Matrix A is\n");

for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
printf(" %5.0f ", temp3[j + i*clm]);
printf("\n");
}
printf("\nThe multiplied matrix A^%d is\n", pow);
if (pow > 1)
{
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
printf(" %5.0f ", mul[j + i*clm]);
printf("\n");
}
}
if (pow == 1)
{
for (i = 0; i < rw; i++)
{
for (j = 0; j < clm; j++)
printf(" %5.0f ", temp3[j + i*clm]);
printf("\n");
}
}
det2 = mul[0];
// printing Determinant
printf("\nDet of A is %5f\nDet of A^%d is %5f\n", det1, pow, det2);
// printing Time
printf("Total communication time is %f\n", commt);
printf("Total computation time is %f\n", compt);
printf("Total parallel run time is %f\n", commt+compt);
}
MPI_Barrier (MPI_COMM_WORLD);
MPI_Finalize();
Output of Cannons Algorithm

Initial Matrix A is
2
0
0

0
0
0

1
0
0

0
0
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
1
0

0
0
0

0
0

0
1

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

1
0

0
0

0
0
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0
1

0
0
0

0
0
1

0
0
0

0
0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

The Multiplied Matrix A^2 Is

4
0
0

0
0
0

1
0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0
0

0
0

0
0
0

0
1
0

0
0
0

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0
0
0

0
0

0
1

0
0

0
1

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0

0
0
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0
0
0

0
0

1
0

0
0

0
0
0
0
0

0
0

0
0
0
0
0

1
0

0
0

0
0
0
0
0

0
0

0
0
0
1
0

0
0

1
0
0
0
0

0
0

0
0
0

0
0

0
0
0
0
0

0
0

0
0
0

0
0

1
0

0
0
0
0
0

0
0

0
0
0

0
0

0
0
0
0
0

0
0

0
0
0

0
0

1
0

0
0

0
0
0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0

1
0

0
0

0
0
0

0
0

0
0
0

0
0
1

0
0
0

Determinant of A is 2.000000
Determinant of A^2 is 4.000000
Total communication time is 0.082163
Total computation time is 0.013703
Total Runtime is 0.095866

0
0

0
1

0
0

0
0
0

0
0

0
1
0

0
0

0
1

0
0

0
0
0

0
0

0
1

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
1

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
1

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

0
0
0

0
0

English s3 Unit 1 The Wild Robot
100% (2)
English s3 Unit 1 The Wild Robot
154 pages
Full Chapter Daily Coding Problem Get Exceptionally Good at Coding Interviews by Solving One Problem Every Day Second Edition Lawrence Wu PDF
100% (8)
Full Chapter Daily Coding Problem Get Exceptionally Good at Coding Interviews by Solving One Problem Every Day Second Edition Lawrence Wu PDF
54 pages
Learn To Build With PHP - A Crash Course
No ratings yet
Learn To Build With PHP - A Crash Course
29 pages
Assignment 1: Name Class Date Period Sbuid Netid Email
No ratings yet
Assignment 1: Name Class Date Period Sbuid Netid Email
4 pages
Jews in Turkey - Cambridge Dictionary-Libre
100% (1)
Jews in Turkey - Cambridge Dictionary-Libre
3 pages
Exercise 9
No ratings yet
Exercise 9
5 pages
Dense Matrix Algorithms: Ananth Grama, Anshul Gupta, George Karypis, and Vipin Kumar
No ratings yet
Dense Matrix Algorithms: Ananth Grama, Anshul Gupta, George Karypis, and Vipin Kumar
55 pages
A Practical Performance Comparison of Parallel Matrix Multiplication Algorithms On Networks of Workstations
No ratings yet
A Practical Performance Comparison of Parallel Matrix Multiplication Algorithms On Networks of Workstations
2 pages
Report - Viber String
No ratings yet
Report - Viber String
26 pages
High Performance Computing Matrix Mul.
No ratings yet
High Performance Computing Matrix Mul.
15 pages
Gauss
No ratings yet
Gauss
7 pages
Pseudo Code of Mpi Programs
No ratings yet
Pseudo Code of Mpi Programs
22 pages
Ee8218 Lab2
No ratings yet
Ee8218 Lab2
7 pages
Fa19 BCS 134
No ratings yet
Fa19 BCS 134
4 pages
Assignment 2
No ratings yet
Assignment 2
2 pages
Content PDF
No ratings yet
Content PDF
14 pages
Class18 - Linalg II Handout PDF
No ratings yet
Class18 - Linalg II Handout PDF
48 pages
Chained Matrix Multiplication
No ratings yet
Chained Matrix Multiplication
32 pages
Chapter 9 - Parallel Computation Problems
No ratings yet
Chapter 9 - Parallel Computation Problems
43 pages
Project 1: Parallel Implementation of Matrix Multiplication: C C B B B B A A A A
No ratings yet
Project 1: Parallel Implementation of Matrix Multiplication: C C B B B B A A A A
1 page
Floyd's Algorithm: Input N: Number of Vertices A (0..n-1) (0..n-1) - Adjacency Matrix
No ratings yet
Floyd's Algorithm: Input N: Number of Vertices A (0..n-1) (0..n-1) - Adjacency Matrix
7 pages
278 hw5
No ratings yet
278 hw5
20 pages
As 3
No ratings yet
As 3
2 pages
Advanced Computer Architecture 1
No ratings yet
Advanced Computer Architecture 1
14 pages
Parallel Algorithms Underlying MPI Implementations
No ratings yet
Parallel Algorithms Underlying MPI Implementations
55 pages
Assign-01
No ratings yet
Assign-01
14 pages
Lect7 Parallel System
No ratings yet
Lect7 Parallel System
20 pages
Ass Parallel
No ratings yet
Ass Parallel
11 pages
Lecture 4: Principles of Parallel Algorithm Design (Part 4)
No ratings yet
Lecture 4: Principles of Parallel Algorithm Design (Part 4)
27 pages
Cannon Strassen DNS Algorithm
No ratings yet
Cannon Strassen DNS Algorithm
10 pages
hpc_scaling
No ratings yet
hpc_scaling
56 pages
Canon's Algorithm
No ratings yet
Canon's Algorithm
11 pages
LEC6 parallelAlg-Broadcasting
No ratings yet
LEC6 parallelAlg-Broadcasting
15 pages
Parallel Algorithms Underlying MPI Implementations
No ratings yet
Parallel Algorithms Underlying MPI Implementations
55 pages
Introduction To Parallel Programming: Parallel Methods For Matrix Multiplication
No ratings yet
Introduction To Parallel Programming: Parallel Methods For Matrix Multiplication
50 pages
Lec8 MPIalgorithmDesign
No ratings yet
Lec8 MPIalgorithmDesign
12 pages
UNIT-8 Forms of Parallelism: 8.1 Simple Parallel Computation: Example 1: Numerical Integration Over Two Variables
No ratings yet
UNIT-8 Forms of Parallelism: 8.1 Simple Parallel Computation: Example 1: Numerical Integration Over Two Variables
12 pages
Assign01
No ratings yet
Assign01
19 pages
PL01 Guiao
No ratings yet
PL01 Guiao
3 pages
MPI Matrix Multiplication 1 PDF
No ratings yet
MPI Matrix Multiplication 1 PDF
23 pages
Semester of MCS-BU: Name: Enayatullah Atal ID: K1S20MCS0003 Date: 4/12/2020 Class: 1 Subj: Advanced Algorithm D&A
No ratings yet
Semester of MCS-BU: Name: Enayatullah Atal ID: K1S20MCS0003 Date: 4/12/2020 Class: 1 Subj: Advanced Algorithm D&A
5 pages
3 Sardar Anisul Haque Thesis UWO
No ratings yet
3 Sardar Anisul Haque Thesis UWO
152 pages
Lab 3
No ratings yet
Lab 3
23 pages
Computational Complexity: ECE 3340 - David Mayerich
No ratings yet
Computational Complexity: ECE 3340 - David Mayerich
12 pages
Project Report: Weight Sharing On CNN
No ratings yet
Project Report: Weight Sharing On CNN
6 pages
Par - 1 In-Term Exam - Course 2018/19-Q2
No ratings yet
Par - 1 In-Term Exam - Course 2018/19-Q2
9 pages
RG2-ParallelizationPrinciples-HPCAI-Jan2020
No ratings yet
RG2-ParallelizationPrinciples-HPCAI-Jan2020
40 pages
20bce2126 PDC Lab Da 3
No ratings yet
20bce2126 PDC Lab Da 3
11 pages
To Print - Dynprog2
No ratings yet
To Print - Dynprog2
46 pages
Department of Computer Scienc2
No ratings yet
Department of Computer Scienc2
5 pages
Efficient Parallel Implementation of The Fox Algorithm
No ratings yet
Efficient Parallel Implementation of The Fox Algorithm
8 pages
CO-2 (2)
No ratings yet
CO-2 (2)
22 pages
Matrix Chain Mult
No ratings yet
Matrix Chain Mult
11 pages
MEGA MAC a Merged Accumulation Based App
No ratings yet
MEGA MAC a Merged Accumulation Based App
4 pages
EXERCISE- 4[1] (1)
No ratings yet
EXERCISE- 4[1] (1)
8 pages
Quiz For Chapter 7 With Solutions
No ratings yet
Quiz For Chapter 7 With Solutions
8 pages
Improving Blocked Matrix-Matrix Multiplication Routine by Utilizing AVX-512 Instructions On Intel Knights Landing and Xeon Scalable Processors
No ratings yet
Improving Blocked Matrix-Matrix Multiplication Routine by Utilizing AVX-512 Instructions On Intel Knights Landing and Xeon Scalable Processors
11 pages
Chapter 14: Parallel Algorithms
No ratings yet
Chapter 14: Parallel Algorithms
23 pages
Sanggjin: Design Study (2 2) Muling Peng, Hong, Doboli, Wendy Computer Engineering Stony
No ratings yet
Sanggjin: Design Study (2 2) Muling Peng, Hong, Doboli, Wendy Computer Engineering Stony
2 pages
Dynamic Programming
No ratings yet
Dynamic Programming
43 pages
gauravkumar_221it027@it301_Lab2
No ratings yet
gauravkumar_221it027@it301_Lab2
28 pages
HW 6
No ratings yet
HW 6
9 pages
Matrices with MATLAB (Taken from "MATLAB for Beginners: A Gentle Approach")
From Everand
Matrices with MATLAB (Taken from "MATLAB for Beginners: A Gentle Approach")
Peter Kattan
3/5 (4)
Line Drawing Algorithm: Mastering Techniques for Precision Image Rendering
From Everand
Line Drawing Algorithm: Mastering Techniques for Precision Image Rendering
Fouad Sabry
No ratings yet
Duolingo English Test 2
No ratings yet
Duolingo English Test 2
3 pages
Robert J. Lemke Oliver
No ratings yet
Robert J. Lemke Oliver
6 pages
CCNA Presentation
No ratings yet
CCNA Presentation
127 pages
Encoders With PROFIBUS-DP User Manual
No ratings yet
Encoders With PROFIBUS-DP User Manual
43 pages
Chapter 12 - ABAP List Viewer Outputs-Part 1 - SAP ABAP - Hands-On Test Projects With Business Scenarios
No ratings yet
Chapter 12 - ABAP List Viewer Outputs-Part 1 - SAP ABAP - Hands-On Test Projects With Business Scenarios
31 pages
Unit 3 Test Yourself A. Phonetics
No ratings yet
Unit 3 Test Yourself A. Phonetics
4 pages
Cultural Differences Between India and Australia
No ratings yet
Cultural Differences Between India and Australia
12 pages
Cazane Mora
No ratings yet
Cazane Mora
58 pages
Spiritual Gifts: To Know, To Love, To Serve
No ratings yet
Spiritual Gifts: To Know, To Love, To Serve
10 pages
Lecture 5-Methods and Method OverLoading
No ratings yet
Lecture 5-Methods and Method OverLoading
28 pages
Module 1 - Blockchain
No ratings yet
Module 1 - Blockchain
5 pages
Bulgarian Historical Review 2021-1-2 Table of Content.
No ratings yet
Bulgarian Historical Review 2021-1-2 Table of Content.
2 pages
Linux System Administration: Linux Professional Institute (LPI) Certification Level 1 Exam 101
No ratings yet
Linux System Administration: Linux Professional Institute (LPI) Certification Level 1 Exam 101
76 pages
Food Report
No ratings yet
Food Report
20 pages
Whole Numbers and The Four Operations: Reteach
No ratings yet
Whole Numbers and The Four Operations: Reteach
10 pages
Hindi (Devanagari
No ratings yet
Hindi (Devanagari
18 pages
Unit 8
No ratings yet
Unit 8
56 pages
P.1 MTC Term 1 Notes
No ratings yet
P.1 MTC Term 1 Notes
39 pages
Principles of Pragmatics Excerpt - Geoffrey Leech (Banter)
0% (1)
Principles of Pragmatics Excerpt - Geoffrey Leech (Banter)
7 pages
11.4 Sum and Difference Identities
No ratings yet
11.4 Sum and Difference Identities
23 pages
Steps in Creating Simple Mail Merge
No ratings yet
Steps in Creating Simple Mail Merge
5 pages
Present An Issue in Varied Ways To Disseminate Information Using The
No ratings yet
Present An Issue in Varied Ways To Disseminate Information Using The
28 pages
Old Testament Life and Literature - Gerald A. Larue
No ratings yet
Old Testament Life and Literature - Gerald A. Larue
492 pages
Life 2e - Beginner - Unit 11 Test - Word
No ratings yet
Life 2e - Beginner - Unit 11 Test - Word
10 pages
Communication Skills
No ratings yet
Communication Skills
3 pages
David Crystal
0% (1)
David Crystal
1 page