This lab should be done with your lab partner.

Lab 9 Goals:

• shared memory parallel programming using pthreads
• synchronization
• 2D arrays and file I/O in C
• pointers, dynamic memory allocation, and pass by reference
• debugging threaded programs

First, both you and your partner should run update31 to create the handin directory structure for lab 9, then you should copy your lab 6 solution into your lab9 subdirectory:

  $ update31
  $ cd cs31/labs/09
  $ pwd
  /home/your_user_name/cs31/labs/09
  $ cp ../06/gol.c .
  $ cp ../06/*.txt .
  $ ls
  Makefile gol.c  and some .txt files

With the 09 starting point code, I've given you a Makefile and some .txt input files. You should create more .txt input files to test larger sized and longer running threaded version, but use smaller ones (and printing command line options) to test for correctness.

This lab is designed to give you some practice writing and debugging multithreaded Pthreads programs, using synchronization primitives, and experience designing and running scalability experiments.

You will implement a parallel version of Conway's Game of Life using your sequential solution as a starting point for this lab. See the lab 6 assignment to remind yourself about the sequential version of the program, and about the rules of the game. You will evaluate the scalability of your implementation as you increase the problem size and the number of threads.

Your lab 9 solution will use the same method as lab 6 for initializing a dynamically allocated board from an input file given as a command line argument, and will have the same print or not-print option after every iteration.

You should begin lab 9 by first fixing any errors in your lab 6 sequential solution that you copied over into your labs/09 subdirectory.

Parallel Version

The parallel version should be structured as follows:

1. The main thread should begin by initializing the game board in the same way as the sequential version. The gettimeofday timers in your code should not include this step but should include all other parts of the execution.
2. After initializing the game board, the main thread spawns off worker threads that will play multiple rounds of game of life, each thread computing just its portion of cells of the new board each round. Grid cells are allocated to threads using either a row-wise or column-wise partitioning specified by a command line argument.
3. If printing after each round is enabled, you should designate one thread to be the printing thread, and only that thread should do the printing after each round (otherwise the output can be interleaved in crazy ways).
4. The main thread should print out the time of the gol computation part of the program before exiting (this time should NOT include the time to initialize the board from the input file, but should measure the rest of the computation).

You will need to think about where and how to synchronize thread actions to remove race-conditions. For example, you need to ensure that no thread starts the next round of play until all threads are done with the current round, otherwise threads will not read consistent values for the given round. If printing is enabled, you need to ensure that the printing thread prints a consistent version of the game board.

A run with no printing enabled should only printout the final timing result of the gettimeofday timer and should not include any calls to usleep in the run.

Command line arguments

Your program will use the two command line arguments from lab 6 and additionally add three new command line arguments to:

1. specify the number of threads (the degree of parallelism)
2. specify how to parallelize the GOL program (0: row-wise grid cell allocation, 1: column-wise grid cell allocation).
3. specify if the per-thread board allocation should be printed out or not (0:don't print configuration information, 1: print allocation information)

$ ./gol
usage: ./gol infile.txt print[0:1] ntids partition[0:1] print_alloc[0:1]

Here are some example command lines:

# run with config values read from file1.txt, do not print the board after each round
# create 8 threads, use row-wise partitioning, print per-thread partitioning details:
./gol file1.txt  0  8  0  1     

# run with config file file2.txt, print the board after each round, create
# 15 threads, use column-wise partitioning, don't print per-thread partitioning:
./gol file2.txt  1  15  1  0   

The print per-thread board allocation command line option will help you debug the grid cell allocation scheme to ensure that you are correctly allocating rows or columns across the worker threads.

Your program should handle badly formed command lines and check for bad values entered (like a negative number of threads). It should print out an error message and exit for badly formed command lines and handle most bad input values similarly.

Partitioning

You will implement two different ways of parallelizing the computation: one partitions the board's rows across threads, the other partitions the board's columns across threads (i.e each thread computes the result for some number of rows or columns). As an example, suppose that the board is 8x8, and that there are 4 threads. Here is how the threads (with logical tids 0-3) would be assigned to computing elements of C using row and column partitioning:

row partitioning                        column partitioning
----------------                        ----------------
0 0 0 0 0 0 0 0                         0 0 1 1 2 2 3 3  
0 0 0 0 0 0 0 0                         0 0 1 1 2 2 3 3
1 1 1 1 1 1 1 1                         0 0 1 1 2 2 3 3
1 1 1 1 1 1 1 1                         0 0 1 1 2 2 3 3
2 2 2 2 2 2 2 2                         0 0 1 1 2 2 3 3
2 2 2 2 2 2 2 2                         0 0 1 1 2 2 3 3
3 3 3 3 3 3 3 3                         0 0 1 1 2 2 3 3
3 3 3 3 3 3 3 3                         0 0 1 1 2 2 3 3

When the number of threads does not evenly divide the dimension by which you are partitioning, divide the rows (or columns) up so that there is at most a difference of 1 assigned row (or column) between any two threads. For example, for an 8x7 board and 3 threads, you would partition rows and columns like this among the 3 threads (with tids 0-2):

row partitioning                        column partitioning
----------------                        ----------------
0 0 0 0 0 0 0                           0 0 0 1 1 2 2   
0 0 0 0 0 0 0                           0 0 0 1 1 2 2
0 0 0 0 0 0 0                           0 0 0 1 1 2 2 
1 1 1 1 1 1 1                           0 0 0 1 1 2 2 
1 1 1 1 1 1 1                           0 0 0 1 1 2 2 
1 1 1 1 1 1 1                           0 0 0 1 1 2 2 
2 2 2 2 2 2 2                           0 0 0 1 1 2 2 
2 2 2 2 2 2 2                           0 0 0 1 1 2 2 

In this partitioning scheme, a single row (or column) is assigned to exactly one thread; a single row (or column) is never split between two or more threads. Thus, in the above example threads 0 and 1 have one more row each of work to do than thread 2 in the row-wise partitioning, and thread 0 has one more column of work to do than threads 1 and 2 in the column-wise partitioning.

Printing Partitioning Information

When the 5th command line option is non-zero, your program should print thread partitioning information. Each thread should print:

1. Its thread id (its logical one not its pthread_self value)
2. Its start and end row index values
3. The total number of rows it is allocated
4. Its start and end column index values
5. The total number of columns it is allocated

The threads will run in different orders, so you may see each thread's output in any order, which is fine. You can avoid some interleaving of individual threads output by calling fflush after printf:

printf("tid %d my values ...\n", mytid, ...);
fflush(stdout);   // force the printf output to be printed to the terminal

Here are some example runs of my program with different numbers of threads partitioning a 100x100 grid. The total number of rows and columns per thread are shown in parentheses (your output does not need to be identical to mine but must include all required parts):

# 9 threads, column-wise partitioning
% gol big 0 9 1 1
tid  0: rows:  0:99 (100) cols:  0:11 (12)
tid  1: rows:  0:99 (100) cols: 12:22 (11)
tid  2: rows:  0:99 (100) cols: 23:33 (11)
tid  4: rows:  0:99 (100) cols: 45:55 (11)
tid  3: rows:  0:99 (100) cols: 34:44 (11)
tid  5: rows:  0:99 (100) cols: 56:66 (11)
tid  6: rows:  0:99 (100) cols: 67:77 (11)
tid  7: rows:  0:99 (100) cols: 78:88 (11)
tid  8: rows:  0:99 (100) cols: 89:99 (11)

# 6 threads, row-wise partitioning
% gol big 0 6 0 1
tid  0: rows:  0:16 (17) cols:  0:99 (100)
tid  1: rows: 17:33 (17) cols:  0:99 (100)
tid  3: rows: 51:67 (17) cols:  0:99 (100)
tid  2: rows: 34:50 (17) cols:  0:99 (100)
tid  4: rows: 68:83 (16) cols:  0:99 (100)
tid  5: rows: 84:99 (16) cols:  0:99 (100)

# 17 threads, row-wise partitioning
% gol big 0 17 0 1
tid  0: rows:  0: 5 ( 6) cols:  0:99 (100)
tid  1: rows:  6:11 ( 6) cols:  0:99 (100)
tid  3: rows: 18:23 ( 6) cols:  0:99 (100)
tid  2: rows: 12:17 ( 6) cols:  0:99 (100)
tid  4: rows: 24:29 ( 6) cols:  0:99 (100)
tid  5: rows: 30:35 ( 6) cols:  0:99 (100)
tid  6: rows: 36:41 ( 6) cols:  0:99 (100)
tid  7: rows: 42:47 ( 6) cols:  0:99 (100)
tid 10: rows: 60:65 ( 6) cols:  0:99 (100)
tid 11: rows: 66:71 ( 6) cols:  0:99 (100)
tid 12: rows: 72:77 ( 6) cols:  0:99 (100)
tid 13: rows: 78:83 ( 6) cols:  0:99 (100)
tid  8: rows: 48:53 ( 6) cols:  0:99 (100)
tid 15: rows: 90:94 ( 5) cols:  0:99 (100)
tid 14: rows: 84:89 ( 6) cols:  0:99 (100)
tid  9: rows: 54:59 ( 6) cols:  0:99 (100)
tid 16: rows: 95:99 ( 5) cols:  0:99 (100)

Note that for the 17 thread run there are enough threads to see "out of order" execution due to the exact scheduling of threads on the cpus.

Correctness

In addition to printing out per-thread board partitioning information to verify correct allocation, you should also ensure that your parallel version solves GOL correctly. To do this run your parallel version on some of the same input files as your sequential one with printing enabled. The results should be identical. For example, starting with the oscillator input file the start and end board should look like this (you would of course call system("clear") between each iteration):

$ ./gol oscillator.txt 1 4 0 0

start board:

- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - @ @ @ - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 

end board:

- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - @ - - - - - 
- - - - - @ - - - - - 
- - - - - @ - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 
- - - - - - - - - - - 

total time for 21 iterations of 11x11 is 5.493663 secs

If you see odd output, either your parallelization of GOL is incorrect, or you are missing synchronization, or both. You can remove the call to system("clear") to see the history of each round to help you debug incorrect GOL computation (of course run for a small number of iterations).

Additional Requirements

In addition to the parallelization and correctness requirements described above, you solution must:

• use a makefile
• be free of valgrind errors
• be well designed and well commented

For the pthread implementation:

• The man pages for pthread functions are very helpful. In addition I have links to pthreads documentations and tutorials

• use perror to print out error messages from failed system and pthread library calls. perror prints out detailed information about the error that occurred. You call it passing in a string with program-specific error message. For example:

  big_buff = malloc( ...);
  if(!big_buff) {
    perror("mallocing big buffer failed\n");
  }
  

  See the man page for perror for more information and for the include files you need to add to your program.

• You will need to use some synchronization in your solution. You may use any of the pthread synchronization primitives: semaphores, condition variables, and/or barriers. Remember that before using any of these, you need to initialize them, and that all threads need to have access to the shared synchronization variables (either pass a reference to them to the main thread function or declare them as static global variables). Here are some code snippets for using the barrier synchronization primitive:

  // declare a global pthread_barrier_t var
  static pthread_barrier_t barrier;
  
  // initialize it somewhere before using it:
  if(pthread_barrier_init(&barrier, NULL, num_threads)){
     perror("Error: with pthread barrier init\n");
     exit(1); 
  }
  
  // threads than can call pthread_barrier_wait to synchronize: 
  ret = pthread_barrier_wait(&barrier);
  if(ret != 0 && ret != PTHREAD_BARRIER_SERIAL_THREAD) {
    perror("Error: can't wait on pthread barrier\n");
    exit(1);
  }
  

• With the exception of pthread synchroniztion primatives, you should not use any other global variables in your program. Any values that threads need should be passed to them via the (void *arg) parameter. Look at the parallel max program we did on Tuesday for an example of how to do this.

• A guide for using gdb to debug pthread programs

Before running cs31handin: Make sure to run make clean to remove all executable files, and remove (or move) any extra or temporary files out of your submit directory. You can move files somewhere else using the mv command:

mv badfile ../.        # move it up one directory
mv gol1.c ../.         # move a confusing duplicate .c file out of your handin
                       # directory

Once you are satisfied with your solution, hand it in by typing handin31 at the unix prompt.

Only one of you or your partner should run handin31 to submit your joint solutions If you accidentally both run it, send me email right away letting me know which of the two solutions I should keep and which I should discard (you don't want the grader to just guess which joint solution to grade).

You may run handin31 as many times as you like, and only the most recent submission will be recorded. This is useful if you realize, after handing in some programs, that you'd like to make a few more changes to them.