lezahlie/Stencil-Project

Performance Analysis of Iterative Stencil Loops using Pthreads and OpenMPI

Authors: Leslie Horace, James Simmons
Project: Stencil-Project
Purpose: Data Collection and Analysis for Parallel applications of 9-pt Stencil Iterations
Note: It is reccommended to open this readme with a markdown viewer

Prequsite Installations (Ubuntu-Debian)

Note: Documentation links included for other environment installations

Update OS Packages First!

• sudo apt update && sudo apt upgrade

OpenMPI

• sudo apt install openmpi-bin openmpi-dev openmpi-common openmpi-doc libopenmpi-dev

Python3 [min-version 3.5]

• sudo apt install python3

Pip3 [optional]

• sudo apt install python3-pip

Matplotlib

• pip3 install matplotlib

Seaborn

• pip3 install seaborn

Numpy

• pip3 install numpy

Opencv

• pip3 install opencv-python

Project Files Overview

GCC Programs

---

Note: <arg> is an input, [arg] is optional, arg1 | arg2 means arg1 or arg2.

1. Makefile

• Usage: make [clean] [all | <program>]
• Compiles/cleans all project programs

2. make-2d.c

• Usage: ./make-2d <rows> <cols> <outfile>
• Generates a matrix and initializes values to represent a boilerplate

3. print-2d.c

• Usage: ./print-2d <infile>
• Prints a matrix to console (debug_level=2)

4. stencil-2d.c

• Usage: ./stencil-2d <num_iterations> <infile> <outfile> <all_stacked_file(optional)>
• Serial version of 9-pt stencil algorithm

5. pth-stencil-2d.c

• Usage: ./pth-stencil-2d <num_iterations> <infile> <outfile> <debug_level[0-2]> <num_threads> <all_stacked_file(optional)>
• Pthread version of 9-pt stencil algorithm

6. mpi-stencil-2d.c

• Usage:
• OpenMPI version of 9-pt stencil algorithm

---

GCC Utility & Header Files

---

1. utilities.c

• User defined functions shared utilized by entire project

2. utilities.h

• Header file containing structs, macros, and protoypes in "utilities.c"

3. mpi_utils.c

• User defined functions for "mpi-stencil-2d.c"

4. mpi_utils.h

• Header file containing structs, macros, and protoypes for "mpi-stencil-2d.c"
• "utilities.h" is linked here, giving access to all prototype functions

5. timer.h

• Gets the current time in microseconds

---

Python Scripts

---

Note: All pythons scripts save output to log files in ./logs/ subfolder

1. generate-matrix.py

• Usage: python3 generate-matrix.py <#rows/cols>
  • <#rows/cols> can be comma delimited lists with no spaces, e.g., "1,2,4"
• Runs "make-2d"
  • Only creates square matrices

2. gather-data.py

• Usage: python3 gather-data.py <pthread/mpi> OPTIONAL: [size] [process/threads] [iterations]
  • If at least one optional arg is entered, all must be entered in the defined order
  • Optional args [size] and [process/threads] can be entered as a comma delimited string list with no spaces, e.g., "1,2,4"
• "generate-matrix.py" must be ran prior to "gather-data.py" with the same matrix sizes
• Runs "pthread-stencil-2d" or "mpi-stencil-2d"

3. analyze-data.py

• Usage: python3 analyze-data.py <pthread/mpi>
• Parses time data files generated from "gather_data.py"
• Calculates speedup, efficiency, and serial fraction
• Generates graphs for time and all calculated data

4. create-video.py

• Usage: python3 create-video.py <stackedfile>
  • <stackedfile> name must include in order, <#rows> <#cols> <#iterations> demimited by any non-numeric chars
  • e.g., xyz-100.abc100.def25.raw-h3ll0 will parse 100 rows x 100 cols x 25 iterations, any #'s after <#iterations> is excluded
• (Alt) Usage: python3 <rows> <columns> <iterations>
  • Utilizes the Makefile and C files to create a raw file, with the rows columns and iterations specified
  • Takes the stacked file generated and creates an mp4 video from it
• Creates a video from stacked raw file

5. debug-data.py

• Usage: python3 debug-data.py
• Runs "stencil-2d", "pthread-stencil-2d", and "mpi-stencil-2d"
• Diffs final matrix files for "stencil-2d" and "pthread-stencil-2d"
• Diffs "pthread-stencil-2d" and "mpi-stencil-2d" to cross check

---

Slurm Job Scheduling Scripts

---

1. submit-jobs.bash

• Usage: bash ./submit-jobs.bash
• Dynamically schedules all jobs
• Defines sbatch flags and input params
  • Schedules "sub-generate.bash"
  • Schedules "sub-gather.bash"
• Calls

2. sub-generate.bash

• Usage: sbatch ./sub-generate.bash <matrix-size | matrix-size-list>
  • Both <matrix-size> and <matrix-size-list> count as 1 string parameter
• Job script to create input matrices
• Compiles the programs
• Runs "generate-matrix.py"

3. sub-gather.bash

• Usage: sbatch --mem=<#GB> --cpus-per-task=1 --ntasks-per-node=<#tasks> ./sub_gather.bash <"mpi" | "pthread"> <#size> <#process> <#iterations>
• Job script to run stencil programs
• runs "gather-data.py" for

---

Project Usage and How to Guides

Part 1: Data Gathering

[Local Machine]

---

Manual Method:

1. Compile programs with Makefile

• Run: make | make all

2. Create a matrix with make-2D

• Run: ./make-2d <rows> <cols> <outfile>

3. Run pthreads-stencil-2d

• Run: ./pth-stencil-2d <num_iterations> <infile> <outfile> <debug_level[0-2]> <num_threads> <all_stacked_file(optional)>

4. Run mpi-stencil-2d

• Run: mpirun -np <num processes> ./mpi-stencil-2d <num_iterations> <infile> <outfile> <debug_level[0-2]> <all_stacked_file(optional)>

Script Method:

1. Compile programs with Makefile

• Run: make | make all

2. Create matrices with generate-matrix.py

• Run: python3 generate-matrix.py <#rows/cols>
• Script will automatically generate output file names as ./mat-[size].dat
• All files are stored in project files directory

3. Run gather-data.py for pthreads

• Run: python3 gather-data.py pthread [size_string_list] [threads_list] [iterations]
• Script will create directory ./data if it doesn't exist
• Saves all output matrices with name ./data/pthread-[size]-[#p].dat and captured results to ./data/pthread_stencil_data.txt

4. Run gather-data.py for OpenMPI

• Run: python3 gather-data.py mpi [size_string_list] [threads_list] [iterations]
• Script will create directory ./data if it doesn't exist
• Saves all output matrices with name ./data/mpi-[size]-[#p].dat and captured results to ./data/mpi_stencil_data.txt

Note: Python Scripts only support square matrices, must use manual method to compute non square matrices Addionally gather-data.py does not support stacked raw file creation, use manual method or create-video.py

[SLURM Environments]

---

1. Setup python script parameters and sbatch flags in "submit_all.bash"

• set matrix size string list param (generate-matrix.py)
  • MAT="5000,10000,20000,40000"
• set number of iterations param (gather-data.py)
  • NI=12
• set proccess/threads list (gather-data.py)
  • procs=(1 2 4 8 16)
• set matrix size list (gather-data.py)
  • sizes=(5000 10000 20000 40000)
• set memory per size list (gather-data.py)
  • mems=(8GB 16GB 32GB 64GB)

Note: The values show in the examples above are already set in "submit_all.bash"

3. run submit-jobs.bash

• $ bash ./submit-all.bash

Note: This will submit individual jobs for both pthreads and mpi automatically

Part 2: Data Analysis

[Local Machine Only]

---

1. Pthreads Performance

• Ensure ./data/pthread_stencil_data.txt generated from "gather_data.py" exists
• Run: python3 ./analyze-data.py mpi
• The following graphs will save in ./plots/ subfolder
  • pthread-Time.png
  • pthread-Speedup.png
  • pthread-Efficiency.png
  • pthread-e.png

2. OpenMPI Performance

• Ensure ./data/mpi_stencil_data.txt generated from "gather_data.py" exists
• Run: python3 ./analyze-data.py pthread
• The following graphs will save in ./plots/ subfolder
  • mpi-Time.png
  • mpi-Speedup.png
  • mpi-Efficiency.png
  • mpi-e.png

2. Create mp4 video of all iterations

• Run: python3 ./create-video.py <stackedfile>
• Alt: python3 <rows> <columns> <iterations>
  • Script will create a raw file named [rows]-[cols]-[iterations].raw
• The generated video will save in ./videos/ subfolder
• Video name is auto generated as heatmap-[rows]x[cols]x[iterations].mp4

Note: create-video.py uses pth-stencil-2d with 2 threads to generate raw files If the above is not preferred, please used the manual method to create raw file