Parallel Programming MP3 Solution

Overview



The purpose of this assignment is straightforward { for you to gain experience using pthreads and assess the performance of a parallel program. To achieve this, you will parallelize an implementation of the Mandelbrot Sequence and benchmark it. Students taking this course for 4 credit-hours will additionally use atomics to parallelize Histogram Sort.




Getting Started



2.1 Skeleton Programs




Pull the skeleton for MP3 from the release repository: https://github-dev.cs.illinois. edu/cs420-fa19/_release







2.2 References




We encourage you to start with the following links if you would like any additional infor-mation about the topics discussed in this MP:




https://en.wikipedia.org/wiki/Counting_sort




https://en.wikipedia.org/wiki/Embarrassingly_parallel https://www.geeksforgeeks.org/fractals-in-cc/




http://books.gigatux.nl/mirror/kerneldevelopment/0672327201/ch09lev1sec1. html




https://en.cppreference.com/w/c/atomic




Part A, Parallelizing the Mandelbrot Set



3.1 The Mandelbrot Set




The Mandelbrot Set is a set of complex numbers whose computation represents an embar-rassingly parallel problem; this means that no synchronization or communication among threads is required. We have provided a skeleton program, mandelbrot, that contains a serial implementation of the Mandelbrot Set which produces a PNG image, an example of which is shown in Figure 1. The parameters for the program are as follows:







1








































































Figure 1: An example image of the Mandelbrot Set.







-o <file { Specify the output le (by default: out.png).




-w <width { Specify the width of the generated image in pixels (by default 480). -h <height { Specify the height of the generated image in pixels (by default 480). -n <N { Run the program with N threads (your task is to implement this).




3.2 Your Task




In the section marked in the skeleton, create kData.numThreads threads and wait for their completion; hint, you are expected to use pthread create and pthread join. These threads should run generateMandelbrot with startX and endX representing a unique, non-overlapping sub-range of 0 to kData.width. You are free to divide up this workload anyway you choose; however, it is worth noting that the number of threads may not evenly divide with the image width (i.e. kData.width).







Once you have implemented your parallel version and veri ed its correctness (by verifying the generated image matches the serial version), you are to run the small benchmarking program we have included. To do this, load Python as follows:




module load python/3 && unset PYTHONPATH




Then run the benchmark with:




python3 benchmark.py <ImgSize




This will run the program with a number of threads, and produce results along the lines of:




Generating a XxX mandelbrot with thread counts: [1, 2, 4, ..., 20] The serial version ran for X s.







2









The parallel version, with 1 thread(s), ran for X s, a speedup of Xx.




The parallel version, with 2 thread(s), ran for X s, a speedup of Xx.




The parallel version, with 4 thread(s), ran for X s, a speedup of Xx.




...




The parallel version, with 20 thread(s), ran for X s, a speedup of Xx.




NOTE: For accurate results, you should not run the benchmark program on the login nodes, create a job script and run it using the same methodology from MP1 and MP2.







Try running the benchmark program with the following data sizes: 256, 512, 1024, 2048, 4096. Brie y comment about how changing the data size a ects the per-formance of the parallel version with four threads, and include the output of the benchmarking program for data size 2048 in your report.










For data size 2048... Plot Speedup vs. Number of Threads using the results of the benchmark program.










For data size 2048... How did the performance of the parallel version with a single thread compare to the serial version? Brie y explain your results.










For data size 2048... How did the performance of the parallel version with multiple threads compare to serial version? Brie y explain your results.










Using the campus cluster documentation to nd the number of cores for the compute nodes, what happened when the number of threads exceeded the number of available cores? What would you expect to happen?













Part B, Histogram Sort (4 Credits Only)



4.1 Your Task




To gain additional experience with pthreads and learn about atomic operations, you will parallelize histogram sort. Atomic operations are operations in which the value cannot be modi ed between the instant its value is read (to be returned) and the moment it is modi ed. This alleviates the need for a critical section, since data races cannot arise when using atomic operations. Like for Part A, a serial version has been provided for use as the basis of your parallel version. To parallelize it, you will have to:




Spawn numThreads threads and divvy up the workload. Change the type of the counts array to atomic int.




Replace the decrement operation on counts in populate sorted with a fetch-and-subtract operation.










3









Replace the increment operation on counts in compute counts with a fetch-and-add operation.




Use the atomic operations from the C11 stdatomics library.







We will be evaluating this part of the assignment based on the correctness of your parallel implementation.




4.2 Histogram Sort Program




The program sorts a randomly generated string of uppercase letters ranging from ’A’ to ’Z’.




The parameters for the program are as follows:




-s <seed { Specify the seed used to generate random values. -l <length { Specify the length of the string to be sorted.




-v { Specify that the program should print the sorted and unsorted strings (verbose output).




-n <N { Run the program with N threads (your task is to implement this).




Submission Guidelines



5.1 Expectations For This Assignment




The graded portions of this assignment are your parallelized Mandelbrot program and an-swers to the short-answer questions. Four credit hour students are additionally expected to complete Part B.




5.1.1 Points Breakdown




5 10pts: | Each Short Answer Question




50pts: | Parallelized Mandelbrot Program




50pts: | Parallelized Histogram Sort (4 cr. only)




5.2 Pushing Your Submission




Follow the git commands to commit and push your changes to the remote repo. Ensure that your les are visible on the GitHub web interface, your work will not be graded otherwise. Only the most recent commit before the deadline will be graded.


































4