Software Lab: Parallel Programming for Many-Core Architectures with OpenMP

Handouts

Kmeans

Sobel

Merge-sort

spmxv

skeleton

Env scripts

TODOs

Regularly every Year

• Setup thesis compute project
  • Apply for compute resources (and if desired for an advanced reservation which is available during in-person meetings)
  • Collect the students' user-ids on the cluster
  • Add all students to the compute project
• Setup Moodle room
  • Schedule automatic uploads of tar archives
  • Schedule automatic uploads of slides

Once

• Add example of a good dev. diary
• Add improved reference solution for kmeans which:
  • initializes the device(s) before entering the time measurement
  • uses hybrid parallelism i.e. multi-threading on the CPU as well
  • supports multiple devices
• Check why -O1 improves performance for spmxv

Creating New Tasks

A template is provided in the skeleton task, which can also be used as a starting point for creating a new task.

Adding a Reference Implementation

• create a directory under tasks using all lowercase letters, which uniquely identifies the new task
• directories for reference versions should use the naming scheme XX_Name, where XX are two digits and Name is a unique identifier in camelcase
• each newly created task should have two directories 00_Handout and 01_Basic which contain the handout version and a serial implementation of the task respectively

Scripting Integration

• to allow for reuse of benchmarking scripts the following Makefile targets and conventions should be available:
  • build, which creates the task's main executable
  • run-small, which runs the task's main executable utilizing a pre-defined small problem size
  • run-large, which runs the task's main executable utilizing a pre-defined large problem size
  • clean, which removes all temporary build files from previous compilations
  • archive, which creates a tar archive of all files relevant for evaluation i.e. omit files which are shared between all groups
  • following a common style for passing the number of threads for a given execution, the environment variable ${NTHREADS} should be read out to set ${OMP_NUM_THREADS} accordingly
• integrating a new task into the parse-benchmark.sh script is possible by providing a custom result parser as described in this section
• integrating a new task into the parse-competition.sh script is possible by providing a custom prepare scripts as described in this section

Scripts

Required Directory Hierarchy

$ tree .
.
├── .Rprofile
├── renv.lock
├── renv
├── ...
├── scripts
│   ├── benchmark
│   │   └── collect-benchmark.sh
│   ├── parse
│   │   ├── parse-benchmark.sh
│   │   └── result-parsers
│   └── visualize
│       ├── visualize-benchmark.R
│       └── visualize-comparison.R
└── tasks
    ├── merge-sort
    ├── ...
    └── spmxv

Instructions for Dependency Installation

# this module has to be loaded before executing any Rscript in this repository
$ module load MATH r-project/4.0.2
$ R
> renv::init()
# press 1 to install the required libraries

Overview

$ tree scripts
scripts
├── batch                               # directory for batch scripts which are templates for the students
│   ├── slurm.batch.gpu.sh              # --> standard execution on CPU backend nodes with GPUs
│   ├── slurm.batch.knl.sh              # --> standard execution on KNL backend nodes
│   ├── slurm.interactive.gpu.sh        # --> interactive session on CPU backend nodes with GPUs
│   └── slurm.interactive.knl.sh        # --> interactive session on KNL backend nodes
├── benchmark                           # directory for scripts related to benchmarking individual task versions
│   ├── collect-benchmark.sh            # --> main script to collect measurements with multiple thread counts
│   └── setup-environment.sh            # --> example script which can be sourced to setup a reproducible shell environment
├── competition                         # directory for scripts related to collecting measurements of multiple task versions (as done in the final student competition evaluation)
│   ├── collect-competition.sh          # --> main script to collect measurements for one fixed execution configuration
│   └── prepare/*                       # --> per-task scripts to create any task specific files before data collection
├── housekeeping                        # directory for scripts that help manage the software lab
│   ├── make-handout-base.sh            # --> script to create a tar archive containing the base directory structure which can be distributed to students
│   ├── make-handout-task.sh            # --> script to create tar archives of individual tasks which can be distributed to students
│   └── create-reference-competition.sh # --> script to create tar archives for the reference implementations of a task
├── parse                               # directory for scripts related to parsing collected measurements into a csv format
│   ├── parse-benchmark.sh              # --> main script to summarize an individual benchmark measurement
│   ├── parse-competition.sh            # --> main script to summarize a competition measurement
│   └── result-parsers/*                # --> per-task scripts to parse the task output into a csv format
└── visualize                           # directory for scripts related to visualizing measurements from csv results
    ├── visualize-benchmark.R           # --> script to inspect an individual benchmark measurement
    ├── visualize-comparison.R          # --> script to compare multiple benchmark measurements between each other
    └── visualize-competition.R         # --> script to evaluate a competition measurement

Result Parsing Integration of new Tasks

The generic implementation of the parse-benchmark.sh script utilizes scripts under result-parsers to extract task specific fields from result files. Integrating with this process without modifying the source of parse-benchmark.sh directly is possible by following these steps:

• create a shell script under result-parsers using ${task}.sh as the script name
• the provided shell script needs to output the task specific csv header fields, when no input file is given ([ ${#} -eq 0 ])
• otherwise the shell script needs to parse the provided input file and output the extracted performance metrics using a csv style

An example is provided as skeleton.sh, which can also be used as a starting point for writing your own result parser.

Competition Integration of new Tasks

The generic implementation of the collect-competition.sh script utilizes scripts under prepare to prepare task directories considering task specific files. Integrating with this process without modifying the source of collect-competition.sh directly is possible by following these steps:

• create a shell script under prepare using ${task}.sh as the script name
• the shell script needs to create / copy all task specific files, which are not part of the extracted archive, in / into the given directory provided as script argument ${1}

An example is provided as skeleton.sh, which can also be used as a starting point for writing your own prepare script.