Skip to main content

SLURM scheduling functionality and a collection of scripts to process the resulting data

Project description

MiniZinc Bench

This is a small collection of scripts that allow you to run benchmarks on a set of MiniZinc instance using MiniZinc Python. The process is split into several steps to be easily customisable to different kinds of possible benchmarks.

Currently, the only supported way of running benchmarks is through SLURM. Other methods may become available in the future.

Preparation

  1. Create a CSV file for the MiniZinc instances containing problem, model, data_file. If you store the instances in the MiniZinc benchmarks repository structure, then you can use the mzn-bench collect-instances command:

    mzn-bench collect-instances <directory> > instances.csv
    
  2. Instantiate a benchmarking environment. This environment should at least contain a Python virtual environment with mzn-bench and your benchmarking scripts, but you can also set up environmental variables, like PATH, and load cluster modules. The following script, bench_env.sh, provides an example environment that can be loaded using source bench_env.sh:

    if [[ "${BASH_SOURCE[0]}" = "${0}" ]]; then
        >&2 echo "Remember: you need to run me as 'source ${0}', not execute it!"
        exit
    fi
    
    # Create or activate Python virtual environment
    if [ -d venv ]; then
        source venv/bin/activate
    else
       python3 -m venv venv
        source venv/bin/activate
        python3 -m pip install mzn-bench
    fi
    
    # Set other environment variables and load cluster modules
    module load MiniZinc/2.4.3
    
  3. Create a benchmarking script. This script will contain the configuration of where the instance file is located, what MiniZinc/Solver configurations to run for every instance, and how the benchmark runner itself should be configured. The script mainly consists of a call to schedule in mzn_bench. For example a benchmarking script that runs Gecode and Chuffed for 20 minutes might look like this:

    from datetime import timedelta
    from pathlib import Path
    
    import minizinc
    
    from mzn_bench import Configuration, schedule
    
    schedule(
        instances=Path("./instances.csv"),
        timeout=timedelta(minutes=20),
        configurations=[
            Configuration(name="Gecode", solver=minizinc.Solver.lookup("gecode")),
            Configuration(name="Chuffed", solver=minizinc.Solver.lookup("chuffed")),
        ],
        nodelist=["critical001"],
    )
    

These are all the possible arguments to schedule:

  • instances: Path - The path to the instances file.
  • timeout: timedelta - The timeout set for the MiniZinc process.
  • configurations: Iterable[Configuration] - MiniZinc solving configurations (see below for details).
  • nodelist: Optional[Iterable[str]] - A list of nodes on which SLURM is allowed to schedule the tasks. If None, mzn-bench will sequentially solve the instances locally.
  • output_dir: Path = Path.cwd() / "results" - The directory in which the raw results will be placed. This directory will be created if it does not yet exist.
  • job_name: str = "MiniZinc Benchmark" - The SLURM job name.
  • cpus_per_task: int = 1 - The number of CPU cores required for each task.
  • memory: int = 4096 - The maximum memory used for each task.
  • debug: bool = False - Directly capture the output of individual jobs and store them in a ./logs/ directory.
  • wait: bool = False - The scheduling process will wait for all jobs to finish.

A Configuration object has the following attributes:

  • name: str - Configuration name used in the output.
  • solver: minizinc.Solver - MiniZinc Python solver configuration.
  • minizinc: Optional[Path] = None - Path to a specific MiniZinc executable. If None is provided, then the first minizinc executable on the PATH will be used.
  • processes: Optional[int] = None - Number of processes to be used by the solver.
  • random_seed: Optional[int] = None - Random seed to be used by the solver.
  • free_search: bool = False - Solver can determine its own search heuristic.
  • optimisation_level: Optional[int] = None - MiniZinc compilation optimisation level, e.g., -O3.
  • other_flags: Dict[str, Any] = field(default_factory=dict) - A mapping of flag name to value of other flags to be provided to the compiler/solver
  • extra_data: Dict[str, Any] = field(default_factory=dict) - Extra data to be added when using a specific Configuration. Internally this will be used by MiniZinc Python's __setitem__ method on the generated instances. If data needs the value of an identifier internal to MiniZinc, then please use an DZNExpression object (e.g., {"preferred_encoding": DZNExpression("UNARY")}).

Schedule SLURM jobs

The job now has to be started on the cluster with the right number of tasks (one for every instance/solver combination). Luckily, the benchmarking script that you've created in the previous step should take care of all of this. So once we ensure that our environment is ready for ou benchmark, we can execute our script and our job will be scheduled.

For example, if we had created a script bench_env.sh with our benchmarking environment and a script start_bench.py with our schedule call, then the following code should schedule our job:

source bench_env.sh
python start_bench.py

You can keep track of the status of your job using the squeue command.

WARNING: Once the job has started the CSV file containing the instances and the instance files themselves should not be changed or moved until the full benchmark is finished. This could causes error or, even worse, inconsistent results.

Note: If you find a mistake after you have scheduled your job, then you should cancel the job as soon as possible. This can be done by using the scancel command. This command will take the job_id, shown when your job is scheduled, as an argument.

Collect information

Once the job is finished, it is time to get your data wrangling pants on! This repository contains some scripts that might be helpful in locating and formatting the information that you need. Some scripts might be used directly while other might need some customising to fit your purpose. Note that these scripts might require some extra dependencies. For this reason, these scripts are not expected to work unless this package is installed as pip install mzn-bench[scripts]. This allows us to install a minimal version on the running cluster and this more complete version locally while processing the data.

General aggregation

The following scripts can help gather the raw *_stats.yml/*_sol.yml files and combine them for further use:

  • mzn-bench collect-objectives <result_dir> <objectives.csv> - This script gathers all objective value information given by MiniZinc and the used solvers and combines it into a single CSV file.
  • mzn-bench collect-statistics <result_dir> <statistics.csv> - This script gathers all statistical information given by MiniZinc and the used solvers and combines it into a single CSV file.

Tabulation

The following scripts filter and tabulate specific statistics.

  • mzn-bench report-status <statistics.csv> - This command will report the number of occurrences of the various solving status of your MiniZinc tasks. Note that the number of satisfied instances is reported as A + B, where A is the number of optimisation instances that reach a solution not proven optimal and B is the number of satisfaction instance finding a solution. Please consult the -h flag to display all options.
  • mzn-bench compare-configurations <statistics.csv> <before_conf> <after_conf> - This command reports on the differences of the achieved results between two configurations (differences in status, runtime, and objective). You can adjust the changes deemed significant with the --time-delta and --objective-delta flag. You can use the --output-mode json option to ensure the output can be easily parsed by other programs.

Solution checking

The mzn-bench check-solutions command takes the solutions output during run and feeds them back into the model to check that the result is satisfiable. It also stores the objective and satisfiability information to be used when checking statuses. The -c option can be used to set how many solutions to check (zero to check all solutions).

# Check three solutions from each instance
mzn-bench check-solutions -c 3 ./results

This requires the problem .mzn and .dzn files from the benchmark run to be available in order to run the checker. The --base-dir <DIR> option can be used to specify a root directory relative to which the file names in the *_sol.yml files are resolved.

Status checking

The mzn-bench check-statuses command takes the results from check-solutions command above (which must be run first) and then checks for any solvers which have either

  • Falsely claimed optimality - where optimality was found by a solver, but a better objective was found elsewhere and verified to be correct.
  • Falsely claimed unsatisfiability - where unsatisfiability was found by a solver, but another solver has given a correct solution for the instance.

Graph generation

There are a number of plotting helper functions available in mzn_bench.analysis.plot. In order to use these, you must enable the plotting features with pip install mzn-bench[plotting]. These use the Bokeh visualisation library to provide interactive plots.

The read_csv function returns a tuple of pandas data frames containing objective and statistics data for plotting or further data analysis.

from mzn_bench.analysis.collect import read_csv
from mzn_bench.analysis.plot import plot_all_instances
from bokeh.plotting import show

# Read CSVs generated by mzn-bench collect-result as pandas dataframes
objs, stats = read_csv("objectives.csv", "statistics.csv")

# Grid plot giving objective values over time, or time to solve
# (depending on instance type)
show(plot_all_instances(objs, stats))

Testing

Currently, this library is tested using a single end-to-end test which runs most of the pipeline locally (without SLURM).

pytest

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

mzn_bench-0.2.0.tar.gz (22.7 kB view hashes)

Uploaded Source

Built Distribution

mzn_bench-0.2.0-py3-none-any.whl (22.5 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page