auyu0408 / MPI_hw4

多處理機平行程式設計 hw4

MPI Hw4 Pthread Smooth

File

• bmp.h: bmp picture headerfile
• Smooth.cpp: original bmp picture smooth file
• input.bmp: picture we want to smooth
• hw4.cpp: program that use busy waiting and mutex to complete
• hw4_bar.cpp: program that use pthread_barrier_wait() to complete
• hw4_sem.cpp: program that use barrier with semaphore to complete
• hw4_con.cpp: program that use condition variables to complete

Building

• Compile

g++ -g -Wall -o [Output Filename] hw4.cpp

• Execute

  • n: numbers of threads want to have

./[Output Filename] n

• Result:

  • Computer 1

  • Computer 2

• Compare with original Smooth process:

Implementation

• Read "input.bmp", get starting time.

• Initialize variable.

• Create thread, in each thread:

  1. Calculate the size and starting point I need to smooth.
  2. barrier, get correct bmp data.
  3. Smoothing.
  4. Waiting each thread complete.
  5. Repeat 2~4 1000 times.

• Waiting thread end.

• Save "output.bmp", get end time.

• Get total time.

• I tried all 3 method and pthread_barrier_wait() when doing my homework, all of them have similar result.

• The barrier method I implement in homework is "busy waiting and mutex".

Result Analysis

• The execution time will be influenced by the number of processor in the computer and the number of thread.
• The picture below is the different number of thread with the CPU usage percentage.

  • thread = 2

    • 

  • thread = 4

    • 

  • thread = 6

    • 

  • thread = 8

    • 
• We could found that if our thread number < total processor, we can't have the best usage of CPU.
  • If our thread number > total processor, we will split CPU resources equally.
• About the execution time, it will have different answer in different environment.
  • In some 4 CPU(s) computer it could nearly = (single thread)/4.
  • But if your computer is using by other thread, it may not as fast as above.
• I guess the ratio is because the computer only have 4 CPUs (2c*2p).

Difficulties

• calculating total time

  • In the beginning, I use clock() to get start and end time. However, clock() will measure your CPU time use by my process, so I only get total time.
  • Then, I changed my function into clock_gettime(). Then I measure execution time correctly.

• swap function

  • Each time when we smooth, we need to swap "BMPSaveData" and "BMPData" in the beginning to get the data we calculate.
  • At first, I forgot that I just need to swap at one of the threads, I swap it at all threads, then I got a strange answer.
  • After, discussing with my classmate, I found the mistake.
  • The wrong code could also produce a correct answer when total thread = 1. When the total thread > 1, it would produce a file without nothing inside it.

References

• 並行程式設計