PMLIB - Analysis: Square Matrix Multiplication

Partitioned Memory Parallel Programming Library (PMLIB)

Based on BSP-RAMP: Partitioned Memory Parallel Programming Framework

Prof. Subodh Kumar

Prof. Sorav Bansal

Tarun Beri

Indian Institute of Technology, Delhi

Array Summation

Experimental Configuration

Four 64-bit Ubuntu Linux 8.04.2 Eight Core Machines [Intel Xeon CPU E5450 3.00 GHz with 16 GB physical memory];

Two machines have one Tesla C1060 GPU Card each

Watch Dog/Kernel Execution Timeout Disabled; mpiexec with 4 processes on different machines;

CUDA Version 3.1; OpenMPI Version 1.78; OpenMP Version 3.0; gcc Version 4.2.4

Equal task partitioning among all processing elements

Array Length

Computation Time (in seconds)

Serial Task

1 host CPU Task

1 host GPU Task

1 host Task
CPU+GPU

Cluster CPU Task

Cluster GPU Task

Cluster Task
CPU+GPU

5000000

11.450

1.621

0.56

1.81

0.74

0.69

1.01

6000000

13.53

2.01

0.56

2.08

0.88

0.72

1.10

7000000

15.77

2.19

0.58

2.45

1.02

0.71

1.24

8000000

18.11

2.53

0.67

2.69

1.29

0.74

1.50

9000000

20.36

2.90

0.59

2.90

1.26

0.77

1.53

10000000

22.66

3.22

0.58

3.20

1.38

0.80

1.69

20000000

45.16

6.27

0.91

6.02

2.75

1.04

2.89

30000000

67.83

9.36

1.11

8.93

4.13

1.52

4.18

40000000

90.35

12.67

Out of Threads

11.70

5.52

1.85

5.40

50000000

112.51

15.62

Out of Threads

14.95

6.79

2.18

6.72

60000000

135.76

19.19

Out of Threads

17.47

8.58

2.52

8.10

Comparison of Serial and PMLIB Tasks

Comparison of PMLIB Tasks

Comparison of Serial Task, Fastest PMLIB Task and Slowest PMLIB Task

Array Length

Computation Time (in seconds)

Serial Task

1 host CPU Task

1 host GPU Task

1 host Task
CPU+GPU

Cluster CPU Task

Cluster GPU Task

Cluster Task
CPU+GPU

Fastest PMLIB Task

Slowest PMLIB Task

5000000

11.450

1.621

0.56

1.81

0.74

0.69

1.01

0.557

1.81

6000000

13.53

2.01

0.56

2.08

0.88

0.72

1.10

0.56

2.08

7000000

15.77

2.19

0.58

2.45

1.02

0.71

1.24

0.58

2.45

8000000

18.11

2.53

0.67

2.69

1.29

0.74

1.50

0.67

2.69

9000000

20.36

2.90

0.59

2.90

1.26

0.77

1.53

0.59

2.90

10000000

22.66

3.22

0.58

3.20

1.38

0.80

1.69

0.58

3.22

20000000

45.16

6.27

0.91

6.02

2.75

1.04

2.89

0.91

6.27

30000000

67.83

9.36

1.11

8.93

4.13

1.52

4.18

1.11

9.36

40000000

90.35

12.67

Out of Threads

11.70

5.52

1.85

5.40

1.85

12.67

50000000

112.51

15.62

Out of Threads

14.95

6.79

2.18

6.72

2.18

15.62

60000000

135.76

19.19

Out of Threads

17.47

8.58

2.52

8.10

2.52

19.19

Fastest LPC Task

Slowest LPC Task

Array Length

Computation Time (in seconds)

% Speedup of PMLIB Task over Serial Task

Serial Task

Fastest PMLIB Task

Slowest PMLIB Task

Fastest PMLIB Task

Slowest PMLIB Task

5000000

11.450

0.557

1.81

2053.98

633.85

6000000

13.53

0.56

2.08

2414.05

649.57

7000000

15.77

0.58

2.45

2733.08

643.98

8000000

18.11

0.67

2.69

2689.40

673.30

Max PMLIB Speedup over Serial Task

9000000

20.36

0.59

2.90

3463.94

701.62

61.13x

10000000

22.66

0.58

3.22

3938.72

702.95

20000000

45.16

0.91

6.27

4948.60

720.76

30000000

67.83

1.11

9.36

6113.03

724.42

40000000

90.35

1.85

12.67

4876.80

712.88

50000000

112.51

2.18

15.62

5158.46

720.50

60000000

135.76

2.52

19.19

5394.41

707.55

PMLIB Tasks slower than Serial Task

PMLIB Tasks faster than Serial Task