This project intended to compare the performance of a C/C++ console application using CPU & GPU programming. The console application consisted of doing a square matrix multiplication via several methods :
- [CPU Programming]
- [Sequential code]
- [Multithreading code (one thread per cell of the result matrix)]
- [Multithreading code (one thread per row of the result matrix)]
- [Multithreading code (each cell of the result matrix is assigned to one of the n threads)]
- [GPU Programming]
Each programming method was compared each other by doing a benchmark on the elapsed time for the computation only.
Information :
Using an ASUS X93S Series laptop whose the charasteristics are listed below :
| Charasteristics | Description |
|---|---|
| Processor | Intel Core i5 2430M 2.4 GHz ~ 2.9 GHz |
| Operating System | Windows 7 Home Premium |
| Chipset | Intel HM65 Express |
| Memory | DDR3 1333 MHz SDRAM, 4096 MB, (1 x 4096 MB) |
| Display | 18.4" 16:9 Full HD (1920x1080) LED Backlight |
| Graphic | NVIDIA® GeForce® GT 540M with 1GB DDR3 VRAM |
| Storage | 1 TB 7200 rpm |
| Optical Drive | DVD player |
| Card Reader | Card reader ( SD/ SDHC/ MS/ MS Pro/ MMC) |
| Webcam | 0.3 Mega Pixel Fixed web camera |
| Networking | Integrated 802.11 b/g/n, Bluetooth™ V2.1+EDR, 10/100/1000 Base T |
| Interface | 1 x Microphone-in jack, 1 x Headphone-out jack, 1 x VGA port / Mini D-sub 15 pins for external monitor, 1 x USB 3.0 port, 3 x USB 2.0 ports, 1 x RJ45 LAN Jack for LAN insert, 1 x HDMI |
| Audio | Built-in Speakers And Microphone, SonicFocus, Altec Lansing® Speakers |
| Battery | 6Cells : 5200 mAh 56 Whrs |
| Power Adapter | Output : 19 V DC, 6.3 A, 120 W Input : 100 -240 V AC, 50/60 Hz universal |
| Dimensions | 44.1 x 29.5 x 4.23 ~5.59 cm (WxDxH) |
| Weight | 4.11 kg (with 6 cell battery) |
| Note | Master HDD: 3.5” SATA, Second HDD: 2.5” SATA |
For this project, Visual Studio Community 2015 was used for CPU programming with the pthread library and Visual Studio 2013 was used for GPU programming with CUDA since CUDA is not supported in VC 2015.
CAUTION Visual Studio Community 2015 does not support CUDA yet. You need to use an older version of VC such as Visual Studio 2013.
First of all, you have to set that pthread library is used by typing 'pthreadVC2.lib' into the additional dependencies found at
'Property' > 'Links Editor' > 'Additional Dependencies' as shown on the picture below.
CAUTION It is possible that you need to type 'HAVE_STRUCT_TIMESPEC' into the preprocessor definition found at 'Property' > 'C/C++' > 'Preprocessor' > 'Preprocessor Definition' as shown on the picture below. Otherwise you will have this error message: Error C2011 'timespec' : redefinition of type 'struct'
All benchmarks presented here are resulted from a specific use of threading such as the number of threads doing the computation.
This first benchmark is created by executing the C++ sequential code with CPU. The code uses dynamic arrays and a 'Matrix' structure.
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.000000 s | 0.000000 s | 0.000000 s |
| 25 x 25 | 625 | 0.000144 s | 0.000106 s | 0.000198 s |
| 50 x 50 | 2,500 | 0.001092 s | 0.001143 s | 0.001261 s |
| 100 x 100 | 10,000 | 0.007243 s | 0.009130 s | 0.010531 s |
| 200 x 200 | 40,000 | 0.069280 s | 0.050119 s | 0.096451 s |
| 500 x 500 | 250,000 | 0.908748 s | 0.976781 s | 0.917461 s |
| 1,000 x 1,000 | 1,000,000 | 14.82270 s | 14.90280 s | 15.32160 s |
| 1,500 x 1,500 | 2,250,000 | 51.84170 s | 53.78350 s | 61.38590 s |
| 2,000 x 2,000 | 4,000,000 | 144.4250 s | 144.3840 s | 138.8950 s |
The second benchmark is created by executing the parallel code with CPU in C++ using pthread. The code uses dynamic arrays and a dynamic number of threads. Each cell in the result matrix is performed by one thread.
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.001818 s | 0.000263 s | 0.001539 s |
| 25 x 25 | 625 | 0.651238 s | 0.641366 s | 0.383544 s |
| 50 x 50 | 2,500 | 1.956470 s | 1.784130 s | 1.783190 s |
| 100 x 100 | 10,000 | 8.355190 s | 8.410570 s | 7.513020 s |
| 200 x 200 | 40,000 | 32.00290 s | 33.54810 s | 34.70680 s |
| 500 x 500 | 250,000 | 210.6790 s | 212.8600 s | 216.6650 s |
| 1,000 x 1,000 | 1,000,000 | 958.3180 s | 918.7720 s | 931.3840 s |
| 1,500 x 1,500 | 2,250,000 | 0.000000 s | 0.000000 s | 0.000000 s |
| 2,000 x 2,000 | 4,000,000 | 0.000000 s | 0.000000 s | 0.000000 s |
Before launching the console application for a square matrix of 1,000 x 1,000, the memory is low (but Visual Studio takes a big part of memory).
At the end of the console application for a square matrix of 1,000 x 1,000, the memory is full because all pointers take a lot of space and there are 1,000,000 of threads in memory.
When the application finished, all the memory is released.
NOTE Not using the 'Cell' structure as in the benchmarks 3 and 4.
The third benchmark is created by executing the parallel code with CPU in C++ using pthread. The code uses dynamic arrays and a dynamic number of threads. Each row in the result matrix is performed by one thread.
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.000655 s | 0.000203 s | 0.000496 s |
| 25 x 25 | 625 | 0.010548 s | 0.011697 s | 0.016335 s |
| 50 x 50 | 2,500 | 0.022197 s | 0.023205 s | 0.029736 s |
| 100 x 100 | 10,000 | 0.053766 s | 0.074381 s | 0.043908 s |
| 200 x 200 | 40,000 | 0.118335 s | 0.112993 s | 0.104711 s |
| 500 x 500 | 250,000 | 0.907414 s | 1.284270 s | 1.133060 s |
| 1,000 x 1,000 | 1,000,000 | 8.618210 s | 8.505290 s | 8.367430 s |
| 1,500 x 1,500 | 2,250,000 | 25.71800 s | 27.26910 s | 25.96100 s |
| 2,000 x 2,000 | 4,000,000 | 51.79730 s | 32.06580 s | 50.01410 s |
In the figure below you can see the performance of the computer when the mutiplication of two 2000 x 2000 matrix is performed.
In the figure below you can see that the memory empties at the end of the program and then all memory is free when the executable is closed.
NOTE Using the 'Cell' structure.
The fourth benchmark is created by executing the parallel code with CPU in C++ using pthread. The code uses dynamic arrays and a dynamic number of threads. Cells in the result matrix are performed by one of the threads according to a shifting algorithm.
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.003897 s | 0.001216 s | 0.001205 s |
| 25 x 25 | 625 | 0.001716 s | 0.001528 s | 0.001156 s |
| 50 x 50 | 2,500 | 0.006530 s | 0.002301 s | 0.002024 s |
| 100 x 100 | 10,000 | 0.007007 s | 0.052376 s | 0.011554 s |
| 200 x 200 | 40,000 | 0.056970 s | 0.046925 s | 0.050223 s |
| 500 x 500 | 250,000 | 0.920728 s | 1.089180 s | 1.085580 s |
| 1,000 x 1,000 | 1,000,000 | 11.11410 s | 10.93310 s | 10.65640 s |
| 1,500 x 1,500 | 2,250,000 | 39.03960 s | 37.91340 s | 38.91380 s |
| 2,000 x 2,000 | 4,000,000 | 101.9630 s | 102.7530 s | 100.9980 s |
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.001545 s | 0.001774 s | 0.030921 s |
| 25 x 25 | 625 | 0.001572 s | 0.007728 s | 0.001804 s |
| 50 x 50 | 2,500 | 0.003861 s | 0.003202 s | 0.005808 s |
| 100 x 100 | 10,000 | 0.006439 s | 0.011860 s | 0.011267 s |
| 200 x 200 | 40,000 | 0.039320 s | 0.059699 s | 0.054823 s |
| 500 x 500 | 250,000 | 0.713641 s | 0.761757 s | 0.725308 s |
| 1,000 x 1,000 | 1,000,000 | 8.023100 s | 7.817020 s | 7.643520 s |
| 1,500 x 1,500 | 2,250,000 | 29.89010 s | 30.31270 s | 30.57400 s |
| 2,000 x 2,000 | 4,000,000 | 92.73030 s | 83.12000 s | 86.39570 s |
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.003548 s | 0.002098 s | 0.031445 s |
| 25 x 25 | 625 | 0.002542 s | 0.002289 s | 0.001970 s |
| 50 x 50 | 2,500 | 0.004001 s | 0.003347 s | 0.002551 s |
| 100 x 100 | 10,000 | 0.016078 s | 0.025590 s | 0.041590 s |
| 200 x 200 | 40,000 | 0.054145 s | 0.049041 s | 0.071671 s |
| 500 x 500 | 250,000 | 0.593204 s | 0.782953 s | 0.628277 s |
| 1,000 x 1,000 | 1,000,000 | 7.991940 s | 7.606290 s | 7.705700 s |
| 1,500 x 1,500 | 2,250,000 | 26.71180 s | 24.19720 s | 24.50140 s |
| 2,000 x 2,000 | 4,000,000 | 77.18820 s | 73.15850 s | 74.93440 s |
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.003808 s | 0.003520 s | 0.037883 s |
| 25 x 25 | 625 | 0.004066 s | 0.026112 s | 0.004563 s |
| 50 x 50 | 2,500 | 0.007793 s | 0.006199 s | 0.004467 s |
| 100 x 100 | 10,000 | 0.013294 s | 0.018714 s | 0.009567 s |
| 200 x 200 | 40,000 | 0.059391 s | 0.042749 s | 0.041908 s |
| 500 x 500 | 250,000 | 0.631150 s | 0.729812 s | 0.594458 s |
| 1,000 x 1,000 | 1,000,000 | 7.877320 s | 8.428020 s | 8.243770 s |
| 1,500 x 1,500 | 2,250,000 | 32.18740 s | 30.61070 s | 29.90000 s |
| 2,000 x 2,000 | 4,000,000 | 89.21210 s | 89.62580 s | 91.95300 s |
NOTE Using the 'Cell' structure.
The fifth benchmark is created by executing the parallel code with GPU in CUDA C++ using pthread. The code uses dynamic arrays and a dynamic number of threads. Each cell in the result matrix is performed by one thread.
| Matrix Dimension | Number of cells | Elapsed Time 1 | Elapsed Time 2 | Elapsed Time 3 |
|---|---|---|---|---|
| 2 x 2 | 4 | 0.000126 s | 0.000085 s | 0.000083 s |
| 25 x 25 | 625 | 0.000096 s | 0.000091 s | 0.000142 s |
| 50 x 50 | 2,500 | 0.000159 s | 0.000177 s | 0.000185 s |
| 100 x 100 | 10,000 | 0.000546 s | 0.000498 s | 0.000567 s |
| 200 x 200 | 40,000 | 0.004514 s | 0.004495 s | 0.004509 s |
| 500 x 500 | 250,000 | 0.066477 s | 0.066450 s | 0.066408 s |
| 1,000 x 1,000 | 1,000,000 | 0.537625 s | 0.536510 s | 0.538056 s |
| 1,500 x 1,500 | 2,250,000 | 1.855070 s | 1.845630 s | 1.853430 s |
| 2,000 x 2,000 | 4,000,000 | 2.736540 s | 3.487340 s | 3.270900 s |
No conclusion yet
No one has contribute to the project.
- Error with the memory deleting:
- detected the 01 May 2016
- solved the 01 May 2016
- Error on benchmark 4 with matrix n x n where n > 2
- detected the 30 April 2016
- solved the 19 May 2016 by using pthread_join() function.