601.229 (S24): Assignment 4: Parallel Merge Sort

Due: Monday, April 8th by 11 pm

Assignment type: Pair, you may work with one partner

Getting started

Download csf_assign04.zip and unzip it.

You will modify the code in the parsort.c source file. You can compile the program using the provided Makefile. To run the program, the invocation is

./parsort filename threshold

where filename is the file containing the data to sort, and threshold is the number of elements below (inclusive) which the program should use a sequential sort.

Grading criteria

Your assignment grade will be determined as follows:

Your task

Your task is to write a program that will sort 64-bit signed integers (stored in a file in little-endian binary format), using a variation of Merge sort, modifying the data in the file so that the original data values are in sorted order from least to greatest.

In addition,

This might sound complicated! Fortunately, this program can be implemented quite easily in about 200 lines of C code.

Fork/join computation

The fork/join model of parallel computation is a technique for parallelizing divide and conquer algorithms.

The outline of a fork/join computation is the following:

if (problem is small enough)
  solve the problem sequentially
else {
  in parallel {
    solve the left half of the problem
    solve the right half of the problem
  }
  combine the solutions to the left/right halves of the problem
}

In the case of merge sort, a fork/join approach will look something like this:

if (number of elements is at or below the threshold)
  sort the elements sequentially
else {
  in parallel {
    recursively sort the left half of the sequence
    recursively sort the right half of the sequence
  }
  merge the sorted sequences into a temp array
  copy the contents of the temp array back to the original array
}

In your program, “sort the elements sequentially” should be delegated to the qsort function.

Recursively sorting in parallel can be implemented by using fork two times to create two child processes, and having each one recursively sort half of the array. (This will work because the data to be sorted will be accessed as a memory-mapped file that can be shared by all of the processes.) Note that the merge_sort function provided in the starter code is already a correct implementation of the sequential merge sort algorithm. You will just need to modify it to use child processes to do the recursive sorting in parallel.

Memory-mapped file I/O

The mmap system call allows a process to map file data into its address space. If the process passes the PROT_READ|PROT_WRITE options for the prot argument and MAP_SHARED option to the options argument, then any modifications the process makes to the memory within the file mapping will be written back to the actual file. Since descendants created with fork() share their initial memory space with their parent, the file only needs to be mmap’ed into memory once so long as each child works on a different region of mapped memory.

Let’s say that you want to map the contents of a file into memory so you can sort it. First, you will need to use the open syscall to open the file in read-write mode and get a file descriptor:

int fd = open(filename, O_RDWR);
if (fd < 0) {
  // file couldn't be opened: handle error and exit
}

Next, mmap will need to know how many bytes of data the file has. This can be accomplished using the fstat system call:

struct stat statbuf;
int rc = fstat(fd, &statbuf);
if (rc != 0) {
    // handle fstat error and exit
}
size_t file_size_in_bytes = statbuf.st_size;

Once the program knows the size of the file, creating a shared read-write mapping will allow the program, and all its descendants, to modify the file in-place in memory:

int64_t *data = mmap(NULL, file_size_in_bytes, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0)
// you should immediately close the file descriptor here since mmap maintains a separate
// reference to the file and all open fds will gets duplicated to the children, which will
// cause fd in-use-at-exit leaks.
// TODO: call close()
if (data == MAP_FAILED) {
    // handle mmap error and exit
}
// *data now behaves like a standard array of int64_t. Be careful though! Going off the end
// of the array will silently extend the file, which can rapidly lead to disk space
// depletion!

Passing in NULL for the requested mapping address gives mmap complete freedom to choose any address in memory to map. Since we don’t care where the file ends up in memory, so long as we can access it, this is what we want. Similarly, we want to map the entire file, so we set the offset to zero.

Note: Don’t forget to call munmap and close before returning from the topmost process in your program before returning to prevent leaking resources. Note that closing a file descriptor does not unmap the memory; both calls must be used.

Creating child processes

The fork() call can be used to spawn child processes from the current process. The child will start executing at the point of the fork call, and will share its initial memory space with the parent. Recall that the fork call will always return the pid zero to the newly started subprocess, and the actual pid to the parent process:

pid_t pid = fork()
if (pid == -1) {
    // fork failed to start a new process
    // handle the error and exit
} else if (pid == 0) {
    // this is now in the child process
}
// if pid is not 0, we are in the parent process
// WARNING, if the child process path can get here, things will quickly break very badly

You must make sure that the child branch exits after it has completed its work. Failure to make the child exit will allow it to continue executing through the parent’s code path, which will lead to memory corruption and other difficult-to-debug behaviors. We highly recommend that you hand over control to a function and exit the child process immediately afterwards:

if (pid == 0) {
    merge_sort(/* arguments to merge sort function */);
    // if merge_sort returns, assume it was successful
    exit(0); // child was successful!
    // everything past here is now unreachable in the child
}

To pause program execution until a child process has completed, we recommend using the waitpid call:

int wstatus;
// blocks until the process indentified by pid_to_wait_for completes
pid_t actual_pid = waitpid(pid_to_wait_for, &wstatus, 0);
if (actual_pid == -1) {
    // handle waitpid failure
}

The wstatus argument provides an opaque handle that can be used with special macros to query information about the how the subprocess exited. The WIFEXITED(wstatus) macro will evaluate to a true value if the subprocess exited normally, and the WEXITSTATUS(wstatus) macro can be used to retrieve the return code that the subprocess exited with:

if (!WIFEXITED(wstatus)) {
    // subprocess crashed, was interrupted, or did not exit normally
    // handle as error
}
if (WEXITSTATUS(wstatus) != 0) {
    // subprocess returned a non-zero exit code
    // if following standard UNIX conventions, this is also an error
}

Thus, the subprocess can notify its parent if its operation succeeded by returning a suitable return code. Remember to propagate error conditions up to the topmost process so it can report that the sort job failed using a non-zero error code.

Note: You must wait on every new process you start. This means that every fork call should have a corresponding waitpid call. Failure to due this in a long-running process creates a “pid leak”, and can lead to pid exhaustion and the inability to start any new processes on the system due to the accumulation of the “zombie processes” (yes this is the technical term). While the kernel and the init process will clean up your zombies after the topmost process exits, it is a good practice to ensure that you promptly deal with zombie processes in your program. We will be manually checking your code to ensure that you don’t leave zombies around while your program executes.

Be careful!

One challenging aspect of managing child processes is making sure that the parent process waits for its child process(es), evn if errors occur.

For example, here is an example of a parent process that does not correctly wait for child processes to complete:

pid_t child1 = fork();
if (child1 == 0) {
  // do child stuff and exit
} else if (child1 < 0) {
  fprintf(stderr, "Error: failed to create first child\n");
  exit(1);
}

pid_t child2 = fork();
if (child2 == 0) {
  // do child stuff and exit
} else if (child2 < 0) {
  fprintf(stderr, "Error: failed to create second child\n");
  exit(1);
}

waitpid(child1);
waitpid(child2);

Think about how it’s possible that the parent process could create a child process but exit before waiting for it.

Handling errors

If the parsort program encounters an error, it should print a message of the form

Error: explanation

to the standard error stream (i.e., stderr), and exit with a non-zero exit code.

Examples of errors that should be handled are:

Generating test data, running the program

You can create some random test data using the gen_rand_data executable we have included in the starter code:

make gen_rand_data
./gen_rand_data [size] [output filename]

For instance, to generate 1000 integers, you can use:

./gen_rand_data 8000 test.in

which will generate 8000 bytes of data (1000 int64s) and place it in a file called test.in. Be sure that your specified size is a multiple of 8 so that your parsort and is_sorted programs will function correctly!

You can also use the M suffix on the size argument to specify the output file size in megabytes. For example, the following command would create a file /tmp/test_1M.in with 131,072 8-byte integer values:

./gen_rand_data 1M /tmp/test_1M.in

We suggest using the /tmp directory (the system temporary directory) to create your test files to prevent accumulating many small test files alongside your assignment, and to avoid interactions with the network file server that might affect the performance of the program. However, if you do decide to use /tmp please note the following points:

If you create a few files that are no more than, say, 16 megabytes in size, and if you clean them up properly before logging out, you should be fine.

To check that your sort program works correctly, you can use the is_sorted program we have included in the starter code:

# generate the file with 1000 integers
./gen_rand_data 8000 data.in
# sort the file
./parsort data.in 500
# verify that the file is sorted correctly
make is_sorted
./is_sorted data.in

If the file is correctly sorted, is_sorted will print “Data values are sorted!”, otherwise, it will print an informative error message.

Remember that you are writing a parallel program that can consume resources at an exponential rate. If you are working on a shared system (e.g. the ugrad machines), you must ensure that you test in a responsible manner. If your program appears to be frozen, or taking an inordinate amount of time, you must immediately terminate it. You should test your program on small inputs first, before moving on to larger ones to contain the blast radius of any potential programming mistakes that you might have made. Do not suspend your program using ctrl-z; you must use ctrl-c to ensure that the entire process tree receives an interrupt signal and is terminated. Estimate the number of processes that your program will attempt to spawn with the given parameters before running the command. You should never try spawning more than a hundred processes at your highest limits on a shared system, and far fewer if they are expected to be long-running processes.

You can collect timing info for a given command be prefixing it with the time command:

time ./parsort test.in 1000

This will report your timing information in the following format:

real    0m0.010s
user    0m0.002s
sys     0m0.001s

You should use the real time reported to get the total wall-clock time your program takes, since the other times will serialize the time taken across all descendants. Since this will be very sensitive to system load, you should run each experiment multiple times, and eliminate any clear outliers before including a result in your report. You should also do you best to run your experiment when the host system is at low load (you can find the current load by using the top command). You will need to tweak the amount of data you test against until you are able to distinguish results between different threshold values on the same data size.

Hints and tips

Ensure that only the topmost process (i.e. the first process executed) ever attempts to open the file, map memory, and carry out cleanup. Attempting munmap() the file multiple times will lead to crashes and other unpredictable behaviour.

We highly recommend that you follow the guidance in the starter code and implement your program in C. Using C++ will make this assignment significantly harder.

Experiments and analysis

To make sure that your parsort program is exhibiting the expected degree of parallelism, we would like you to perform an experiment where you create a random data file of 16 megabytes in size, and then time sorting this file multiple times, while adjusting the threshold to achieve increasing amounts of parallel execution.

You should run the following commands:

make clean
make
mkdir -p /tmp/$(whoami)
./gen_rand_data 16M /tmp/$(whoami)/data_16M.in
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 2097152 
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 1048576
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 524288
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 262144
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 131072
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 65536
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 32768
cp /tmp/$(whoami)/data_16M.in /tmp/$(whoami)/test_16M.in
time ./parsort /tmp/$(whoami)/test_16M.in 16384
rm -rf /tmp/$(whoami)

The parsort commands start with a completely sequential sort, and then tests with increasing degrees of parallelism. For example, at the smallest threshold of 16384 elements, there will be 128 processes doing sequential sorting at the base cases of the recursion.

We have provided a shell script called run_experiments.sh which runs these commands, so to collect your experimental data you could just run the command

./run_experiments.sh

We suggest using one of the numbered ugrad machines (ugrad1.cs.jhu.edu to ugrad24.cs.jhu.edu) to do your experiment. When you log in, you can run the top command to see what processes are running. (Note that you can type q to exit from top.) If any processes are consuming significant CPU time, you should consider logging into a different system, until you find one where no processes are consuming significant CPU time.

When you run the commands, copy the output of the time command. The real time will indicate the amount of time that elapsed between when the program started and exited, which is a pretty good measure of how long the sorting took. You should see that decreasing the threshold decreased the total time, although depending on the number of CPU cores available, eventually a point of dimimishing returns will be reached.

In your README.txt, write a brief report which

  1. Indicates the amount of time that your parsort program took so sort the test data for each threshold value, and
  2. States a reasonable explanation for why you saw the times you did

For #2, think about how the computation unfolds, and in particular, what parts of the computation are being executed in different processes, and thus which parts of the computation could be scheduled by the OS kernel in parallel on different CPU cores. We don’t expect a completely rigorous and in-depth explanation, but we would like you to give an intuitive explanation for the results that you observed.

Note on the autograder

Passing the autograder will be a necessary but insufficient condition for full credit. This means that you may still lose functionality points, even if you pass all of the autograder tests. Due to the nature of testing parallel programs, there will be a significant number of points up for manual review, so please structure your code accordingly. Some of the things we may manually verify (bot not limited to) are:

Submitting

Edit the README.txt file to include the report and summarize each team member’s contributions.

You can create a zipfile of your solution using the command make solution.zip.

Submit your zipfile to Gradescope as Assignment 4.