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
Square Matrix Multiplication
Experimental Configuration 14
Four 64-bit Ubuntu Linux 8.04.2 Eight Core Machines [Intel Xeon CPU E5450 3.00 GHz with 16 GB physical memory]; 5
Two machines have one Tesla C1060 GPU Card each 9
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
Square Matrix Dimension  Matrix Multiplication 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
100 0.001 0.001 0.53 0.54 0.02 0.54 0.54
200 0.01 0.01 0.53 0.54 0.02 0.54 0.55
300 0.04 0.02 0.53 0.54 0.04 0.55 0.57
400 0.09 0.03 0.54 0.54 0.07 0.55 0.59
500 0.18 0.10 0.54 0.54 0.12 0.57 0.62
1000 5.22 0.77 0.66 0.89 0.60 0.72 0.92
1500 21.34 2.98 0.99 2.90 1.91 1.04 1.90
2000 45.91 6.15 1.63 6.73 3.50 1.56 3.67
2500 106.01 14.99 2.70 13.57 6.06 2.37 6.42
3000 172.24 25.61 4.29 24.35 11.20 3.50 9.95
3500 307.82 45.33 6.48 40.64 16.93 5.01 17.52
4000 433.02 65.62 9.46 57.08 22.31 6.94 22.31
4500 695.93 96.05 13.22 90.62 34.85 9.36 31.97
5000 919.17 126.21 18.05 126.20 42.70 12.32 42.78
Comparison of Serial and PMLIB Tasks
Comparison of PMLIB Tasks
Comparison of Serial Task, Fastest PMLIB Task and Slowest PMLIB Task
Square Matrix Dimension  Matrix Multiplication 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
100 0.001 0.001 0.53 0.54 0.02 0.54 0.54 0.001 0.54
200 0.01 0.01 0.53 0.54 0.02 0.54 0.55 0.01 0.55
300 0.04 0.02 0.53 0.54 0.04 0.55 0.57 0.02 0.57
400 0.09 0.03 0.54 0.54 0.07 0.55 0.59 0.03 0.59
500 0.18 0.10 0.54 0.54 0.12 0.57 0.62 0.10 0.62
1000 5.22 0.77 0.66 0.89 0.60 0.72 0.92 0.60 0.92
1500 21.34 2.98 0.99 2.90 1.91 1.04 1.90 0.99 2.98
2000 45.91 6.15 1.63 6.73 3.50 1.56 3.67 1.56 6.73
2500 106.01 14.99 2.70 13.57 6.06 2.37 6.42 2.37 14.99
3000 172.24 25.61 4.29 24.35 11.20 3.50 9.95 3.50 25.61
3500 307.82 45.33 6.48 40.64 16.93 5.01 17.52 5.01 45.33
4000 433.02 65.62 9.46 57.08 22.31 6.94 22.31 6.94 65.62
4500 695.93 96.05 13.22 90.62 34.85 9.36 31.97 9.36 96.05
5000 919.17 126.21 18.05 126.20 42.70 12.32 42.78 12.32 126.21
Fastest LPC Task Slowest LPC Task
Square Matrix Dimension  Matrix Multiplication Time (in seconds) % Speedup of PMLIB Task over Serial Task
                         Serial Task Fastest PMLIB Task Slowest PMLIB Task Fastest PMLIB Task Slowest PMLIB Task
100 0.001 0.001 0.54 94.36 0.24
200 0.01 0.01 0.55 172.32 1.90
300 0.04 0.02 0.57 184.28 6.23
400 0.09 0.03 0.59 247.63 14.67 Max PMLIB Speedup over Serial Task
500 0.18 0.10 0.62 188.35 29.25 74.64x
1000 5.22 0.60 0.92 868.15 569.52
1500 21.34 0.99 2.98 2147.04 715.60
2000 45.91 1.56 6.73 2942.27 682.07
2500 106.01 2.37 14.99 4469.61 707.25
3000 172.24 3.50 25.61 4927.79 672.46
3500 307.82 5.01 45.33 6142.32 678.99
4000 433.02 6.94 65.62 6242.61 659.92
4500 695.93 9.36 96.05 7432.48 724.55
5000 919.17 12.32 126.21 7463.53 728.27
 PMLIB Tasks slower than Serial Task  PMLIB Tasks faster than Serial Task