Skip to content

Experiment execution and result management for empirical evaluations of algorithms in Python.

License

Notifications You must be signed in to change notification settings

d-krupke/AlgBench

Repository files navigation

AlgBench: A Python-util to run benchmarks for the empirical evaluation of algorithms.

There are a number of challenges when performing benchmarks for (long-running) algorithms.

  • Saving all information requires a lot of boilerplate code and often you forget something.
  • If you add some further instances or want to compare an additional parameter, you have to check which data is already available to skip existing entries. Same if you need to interrupt the benchmark.
  • Just piping the results into a file can create a huge amount of data, no longer fitting into memory.
  • Proper benchmarks often take days or even weeks to run, such that parallelization is necessary (e.g., with slurm) which requires a thread-safe database.
  • Many file formats and databases are difficult to access or impossible to repair once corrupted.

AlgBench tries to ease your life by

  • saving a lot of the information and data (function arguments, return values, runtime, environment information, stdout, etc.) automatically with a single line of code
  • remembering which function arguments have already run and skipping those
  • providing a compressible database to save memory, and saving highly redundant information, e.g., of the environment, only once
  • providing an NFS-compatible parallel database and compatibility to distribution libraries, such as slurminade
  • using a simple format based on JSON and Zip to allow simple parsing and even repairing broken databases by hand

There is a predecessor project, called AeMeasure. AeMeasure made saving the data easy, but required more boilerplate code and reading the data was more difficult and less efficient.

Other things you should know about for empirical/experimental evaluations

The following tools I consider essential for empirical evaluations (of algorithms):

  • pandas: Simple and powerful tool for working with data tables. Do your experiments and parse the important data into a pandas DataFrame.
  • seaborn and matplotlib: Creating beautiful plots from pandas DataFrames with little work.
  • JupyterLab: Interactive Python+Markdown documents. Great for analyzing data and sharing the insights. Works great with pandas and seaborn.

AlgBench essentially takes over the part of saving the information from the runs and allowing you to easily extract pandas DataFrames from it. For very simple studies, you could also directly save your data into a Pandas DataFrame but even for nearly every serious experiment, you run into the problems mentioned in the beginning.

Note that the actual algorithms can also be written in another, more efficient programming language. It is reasonably easy to create Python-bindings, e.g., for C++ with pybind11, or just call the binaries with Python.

Publishable evaluations often require extensive experiments that are best performed on a cluster of shared workstations. Many institutes and companies are using slurm to schedule and distribute the workloads. The data is usually shared via a network file system (NFS), for which AlgBench is designed. While you usually also have databases available, they are not made for just dumping all the data you may need for analyzis and potentially debugging into. We developed an additional tool slurminade that allows you to distribute your experiments with just a few additional lines. You can see this in an example: original script vs script with slurminade.

Let me further recommend the books A Guide To Experimental Algorithmics by Catherine McGeoch here that gives a good introduction into the big picture of performing empirical evaluations for algorithms. If you want to know more about actually implementing complex algorithms for difficult problems, I recommend to read In Pursuit of the Traveling Salesman by Bill Cook or The Traveling Salesman Problem: A Computational Study by Appelgate et al. to really go into details. The Traveling Salesman Problem is an excellent example for this because it is probably had gotten the most attention of any NP-hard combinatorial problems. However, it can also be intimidating as you probably won’t have the funds to look into any problem as deep as the Travelings Salesman Problem has been looked at. Maybe you want to read some papers from the SIAM Symposium on Algorithm Engineering and Experiments (ALENEX) to see how smaller studies can be performed (though, for most papers you will find aspects that could be improved).

Before you submit any paper (or thesis) with an empirical analysis, I also recommend to first go through this checklist.

Installation

You can install AlgBench using pip

pip install -U algbench

Usage

There is one important class Benchmark to run the benchmark, and two important functions describe and read_as_pandas to analyze the results.

  1. Create a function that creates an entry in the database. Name all arguments that should be saved and used for identifying entries without _ in the front. They should be JSON-compatible. Name all arguments that provide higher objects, such as the instance database, with an _ in the front to tell algbench not to try to save or compare them. Return everything you want to be saved for the benchmark, best as a dictionary.
def create_benchmark_entry(
    instance_name: str,  # instance identifier for the database
    alg_parameters: dict,  # readable parameters for the algorithm
    _instance,  # the parsed instance (not to be added to the database)
):
    solution = alg(_instance, **alg_parameters)
    return {"objective_value": solution.obj()}
  1. Create a Benchmark-object by passing it a path for the database.
from algbench import Benchmark

benchmark = Benchmark("./my_benchmark")

# Optionally (if logging is used):
import logging

# Configure which logger should be captured and with which level
benchmark.capture_logger("my_alg", logging.INFO)
benchmark.capture_logger("my_alg.submodule", logging.WARNING)
  1. Use Benchmark.add to the function for all missing entries.
for instance_name, instance in instance_db:
    for params in params_to_compare:
        benchmark.add(
            create_benchmark_entry,  # function (could also be a lambda)
            # arguments for function
            instance_name=instance_name,
            alg_parameters=params,
            _instance=instance,
        )
benchmark.compress()  # reduce the size of the database by file compression
  1. Use a for loop to iterate over all raw entries
benchmark = Benchmark("./my_benchmark")
for entry in benchmark:
    print(entry)  # dictionary

or read_as_pandas to extract a simple pandas table

t = read_as_pandas(
    "./my_benchmark/",
    lambda result: {
        "instance": result["parameters"]["args"]["instance_name"],
        "alg_params": result["parameters"]["args"]["alg_params"],
        "obj": result["result"]["objective_value"],
        "runtime": result["runtime"],  # automatically saved
    },
)

You can use describe("./my_benchmark") to get an overview of the available entries.

The Benchmark class provides further functionality, e.g., for deleting selected entries or reparing a broken database.

You can find an example for graph coloring. The important parts are shown below.

Running a benchmark

from _utils import InstanceDb
from algbench import Benchmark
import networkx as nx

benchmark = Benchmark("03_benchmark_data")
instances = InstanceDb("./01_instances.zip")


def load_instance_and_run(instance_name: str, alg_params):
    # load the instance outside the actual measurement
    g = instances[instance_name]

    def eval_greedy_alg(instance_name: str, alg_params, _instance: nx.Graph):
        # arguments starting with `_` are not saved.
        coloring = nx.coloring.greedy_coloring.greedy_color(_instance, **alg_params)
        return {  # the returned values are saved to the database
            "num_vertices": _instance.number_of_nodes(),
            "num_edges": _instance.number_of_edges(),
            "coloring": coloring,
            "n_colors": max(coloring.values()) + 1,
        }

    benchmark.add(eval_greedy_alg, instance_name, alg_params, g)


alg_params_to_evaluate = [
    {"strategy": "largest_first", "interchange": True},
    {"strategy": "largest_first", "interchange": False},
    {"strategy": "random_sequential", "interchange": True},
    {"strategy": "random_sequential", "interchange": False},
    {"strategy": "smallest_last", "interchange": True},
    {"strategy": "smallest_last", "interchange": False},
    {"strategy": "independent_set"},
    {"strategy": "connected_sequential_bfs", "interchange": True},
    {"strategy": "connected_sequential_bfs", "interchange": False},
    {"strategy": "connected_sequential_dfs", "interchange": True},
    {"strategy": "connected_sequential_dfs", "interchange": False},
    {"strategy": "saturation_largest_first"},
]

if __name__ == "__main__":
    for instance_name in instances:
        print(instance_name)
        for conf in alg_params_to_evaluate:
            load_instance_and_run(instance_name, conf)
    benchmark.compress()

Analyzing the data

from algbench import describe, read_as_pandas, Benchmark

describe("./03_benchmark_data/")

Output:

 result:
| num_vertices: 68
| num_edges: 697
| coloring:
|| 0: 7
|| 1: 8
|| 2: 2
|| 3: 5
|| 4: 3
|| 5: 7
|| 6: 7
|| 7: 6
|| 8: 5
|| 9: 4
|| 10: 5
|| 11: 4
|| 12: 0
|| 13: 6
|| 14: 0
|| 15: 3
|| 16: 5
|| 17: 5
|| 18: 7
|| 19: 0
|| ...
| n_colors: 9
 timestamp: 2023-05-25T21:58:39.201553
 runtime: 0.002952098846435547
 stdout:
 stderr:
 env_fingerprint: 53ad3b5b29d082d7e2bca6881ec9fe35fe441ae1
 args_fingerprint: 10ce65b7a61d5ecbfcb1f4e390d72122f7a1f6ec
 parameters:
| func: eval_greedy_alg
| args:
|| instance_name: graph_0
|| alg_params:
||| strategy: largest_first
||| interchange: True
 argv: ['02_run_benchmark.py']
 env:
| hostname: workstation-r7
| python_version: 3.10.9 (main, Jan 11 2023, 15:21:40) [GCC 11.2.0]
| python: /home/krupke/anaconda3/envs/mo310/bin/python3
| cwd: /home/krupke/Repositories/AlgBench/examples/graph_coloring
| environment: [{'name': 'virtualenv', 'path': '/home/krupke/.local/lib/python3.10/site-pack...
| git_revision: 5357426feb4b49174c313ffa33e2cadf6a83e226
| python_file: /home/krupke/Repositories/AlgBench/examples/graph_coloring/02_run_benchmark.py
# we can also see the raw data of the first entry using `front`
Benchmark("./03_benchmark_data/").front()

Output:

{'result': {'num_vertices': 68,
  'num_edges': 697,
  'coloring': {'0': 7,
   '1': 8,
   '2': 2,
   '3': 5,
   '4': 3,
   '5': 7,
   '6': 7,
   '7': 6,
   '8': 5,
   '9': 4,
   '10': 5,
   '11': 4,
   '12': 0,
   '13': 6,
   '14': 0,
   '15': 3,
   '16': 5,
   '17': 5,
   '18': 7,
   '19': 0,
   '20': 2,
   '21': 3,
    ...},
  'n_colors': 9},
 'timestamp': '2023-05-25T21:58:39.201553',
 'runtime': 0.002952098846435547,
 'stdout': '',
 'stderr': '',
 'env_fingerprint': '53ad3b5b29d082d7e2bca6881ec9fe35fe441ae1',
 'args_fingerprint': '10ce65b7a61d5ecbfcb1f4e390d72122f7a1f6ec',
 'parameters': {'func': 'eval_greedy_alg',
  'args': {'instance_name': 'graph_0',
   'alg_params': {'strategy': 'largest_first', 'interchange': True}}},
 'argv': ['02_run_benchmark.py'],
 'env': {'hostname': 'workstation-r7',
  'python_version': '3.10.9 (main, Jan 11 2023, 15:21:40) [GCC 11.2.0]',
  'python': '/home/krupke/anaconda3/envs/mo310/bin/python3',
  'cwd': '/home/krupke/Repositories/AlgBench/examples/graph_coloring',
  'environment': [{'name': 'virtualenv',
    'path': '/home/krupke/.local/lib/python3.10/site-packages',
    'version': '20.14.1'},
   {'name': 'cfgv',
    'path': '/home/krupke/.local/lib/python3.10/site-packages',
    'version': '3.3.1'},
  ...],
  'git_revision': '5357426feb4b49174c313ffa33e2cadf6a83e226',
  'python_file': '/home/krupke/Repositories/AlgBench/examples/graph_coloring/02_run_benchmark.py'}}
# we can extract a full pandas tables using `read_as_pandas`
t = read_as_pandas(
    "./03_benchmark_data/",
    lambda result: {
        "instance": result["parameters"]["args"]["instance_name"],
        "strategy": result["parameters"]["args"]["alg_params"]["strategy"],
        "interchange": result["parameters"]["args"]["alg_params"].get(
            "interchange", None
        ),
        "colors": result["result"]["n_colors"],
        "runtime": result["runtime"],
        "num_vertices": result["result"]["num_vertices"],
        "num_edges": result["result"]["num_edges"],
    },
)
print(t)

Output:

       instance                  strategy interchange  colors   runtime ...
0       graph_0             largest_first        True       9  0.002952
1       graph_0             largest_first       False      10  0.000183
2       graph_0         random_sequential        True       9  0.003562
3       graph_0         random_sequential       False      12  0.000173
4       graph_0             smallest_last        True       9  0.003813
...         ...                       ...         ...     ...       ...
5995  graph_499  connected_sequential_bfs        True       3  0.000216
5996  graph_499  connected_sequential_bfs       False       3  0.000132
5997  graph_499  connected_sequential_dfs        True       3  0.000231
5998  graph_499  connected_sequential_dfs       False       4  0.000132
5999  graph_499  saturation_largest_first        None       3  0.000202


[6000 rows x 7 columns]

Which information is saved?

The following information is saved automatically:

  • function name
  • all arguments that do not begin with “_” (use this to pass parsed instances etc.)
  • the returned values
  • runtime
  • current date and time
  • hostname
  • Python version
  • Python binary path
  • current working directory
  • stdout and stderr
  • all installed modules and their versions
  • git revision
  • path of the python file

Things to be aware of

  • Only function name and arguments not starting with “_” are used to compare entries. If an argument (or part of it) is not JSON-compatible, the string of it is used.
  • Arguments and return values that cannot be translated to json are converted to string in the database. The default string conversion may not be very useful.
  • The stdout/strerr capturing only works if Python’s stdout/stderr are used. E.g., C++ write by default to the system’s stdout/stderr and cannot be captured (if you have been wondering, why C++-modules have a bad output it Jupyter-notebooks: this is the reason). PyBind11 allows you to change that behavior.
  • Global variables are not saved. Try to pass all important parameters as function arguments, as they can also alter the benchmark and are important to distinguish entries (e.g., you would want to recompute an entry if the timelimit has been changed. This is only possible if you tell algbench this by making it an argument).
  • ‘sys.argv’ and the filename are saved, but not used for distinguishing entries.

On doing good empirical evaluations of algorithms

To get a feeling on the interesting instances and parameters, or generally on where to look deeper, you should first perform an explorative study. For such an explorative study, you should select some random parameters and instances, and just look how the numbers look. Iteratively change the parameters and instances, until you know what to evaluate properly. At that point, you can state some research questions and design corresponding workhorse studies to answer them.

Here are some general hints:

  • Do not mix algorithm code and experiment code, even if it saves you rebuilding your package after every change. Such a mixed setup may save you a command line, but it is harder to log and many problems may remain unnoticed until you try to publish your algorithm. The little overhead is worth it in the long run.
  • Create a separate folder for every study. Don’t mix too much because you want to reduce redundancies: Once things become complicated, you may draw conclusions from the wrong data without noticing.
  • Add a README.md into each folder that describes the study. At least describe in a sentence, who created this study when in which context.
  • Have separated, numerated files for preparing, running, processing, checking, and evaluating the study.
  • Extract a simplified pandas table from the database with only the important data (e.g., stdout or environment information are only necessary for debugging and don’t need to be shared for evaluation). You can save pandas tables as .json.zip such that they are small and can simply be added to your Git, even when the full data is too large.
  • The file for checking the generated data should also describe it.
  • Use a separate Jupyter-notebook for each family of plots you want to generate.
  • Save the plots into files whose name you can easily trace back to the generating notebook. You will probably copy them later into some paper and half a year later, when you receive the reviews and want to do some changes, you have to find the code that generated them.

On gaining more insights using logging

If you develop complex algorithms, you often want to not only measure the runtime of the whole algorithm, but also of its parts, as well as other information, such as the number of iterations, the current solution, etc. You can use the Python logging framework for this. The logging framework allows you to create loggers that can be configured individually. You can also create a logger for each module and submodule, and configure them individually. You can further configure handlers for the loggers, e.g., to write them to a file or to the console. The level of the loggers and handlers can also be configured, such that you can easily switch between different levels of logging. AlgBench allows you to capture the loggers and save them to the database. You can then extract and analyze them.

You can also use simple print statements, but they are not as flexible as the logging framework. While AlgBench can actually add the runtime to the print statements, it is not as easy to configure the output as with the logging framework. There is no way to disable the output for individual parts of your algorithm, or to change the level of the output. The logging framework is as easy to use as print statements, but much more flexible. It can be more expensive, but print statements are also not free and should be used with care.

Here is an example for using the logging framework:

import logging


def my_alg():
    logger = logging.getLogger("my_alg")
    logger.info("Starting my_alg")
    # do something
    logger.info("Finished my_alg")


logger = logging.getLogger("my_alg")
logger.setLevel(logging.INFO)
logger.addHandler(logging.StreamHandler())
my_alg()

A further advantage of the logging framework is that you can separate the message structure from the data. This allows you to easily query for specific events and directly extract the data you want to analyze.

logger.info("Submodule X needed %d iterations", 42)

Will be saved as a dictionary with a separate field for the message and the data:

{
    "msg": "Submodule X needed %d iterations",
    "args": [42],
}

A further alternative is to use a dedicated class for stats that you pass around. This is generally a good idea, but takes more work and requires you to change the code. The logging framework is a good compromise between flexibility and ease of use.

If your algorithm may be run in parallel or different contexts, you may want to allow to pass a logger to the algorithm. This allows you to create a separate logger for each context to separate the logs.

Note that AlgBench v2 automatically adds the runtime to print statments and log entries.

Using Git LFS for the data

The data are large binary files. Use Git LFS to add them to your repository more efficiently.

You can find a guide here on how to install Git LFS.

Run

git lfs install

to set up git LFS and

git lfs track "*.zip"

to manage all zips via LFS.

Alternatively, you can also just edit .gitattributes by hand

*.zip filter=lfs diff=lfs merge=lfs -text

Finally, add .gitattributes to git via

git add .gitattributes

Version History

  • 2.4.1 Fixes bug when path ends with /.
  • 2.4.0 Removing information on installed packages due to deprecated pkg_resources. New apply-function.
  • 2.2.2 Fixing problem with Jupyter notebooks, because they may not have a __file__ attribute.
  • 2.2.1 Should be able to deal with corrupt zip files now.
  • 2.2.0 Allowing to skip entries in read_as_pandas by returning a None for the row.
  • 2.1.0 More flexible stream handling. You can now disable the output saving and hidding. The default behavior still is to save the output with time stamps and hide it from the console.
  • 2.0.0 Extensive change of stdout/stderr handling and new logging functionality.
    By default, stdout and stderr will now be saved with the runtime of the function. Additionally, you can now capture loggers of the Python logging framework and save them to the database. This is especially useful if you use a library that uses the logging framework. Prefere logging over print for logging information.
  • 1.1.0 Some changes for efficiency turned out to be less robust in case of, e.g., keyboard interrupt. Fixed that.
  • 1.0.0 Changing the database layout, making it more efficient (breaking change!).
  • 0.2.0 Changing database slightly to contain meta data and doing more caching. Saving some more information.
  • 0.1.3 Fixed bug in arg fingerprint set.
  • 0.1.2 Fixed bug with empty rows in pandas table.
  • 0.1.1 Fixed bug with delete_if.
  • 0.1.0 First complete version

About

Experiment execution and result management for empirical evaluations of algorithms in Python.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages