Skip to content

Latest commit

 

History

History
487 lines (364 loc) · 17.4 KB

README.md

File metadata and controls

487 lines (364 loc) · 17.4 KB

Microbench

Microbench: Benchmarking and reproducibility metadata capture for Python

Microbench is a small Python package for benchmarking Python functions, and optionally capturing extra runtime/environment information. It is most useful in clustered/distributed environments, where the same function runs under different environments, and is designed to be extensible with new functionality. In addition to benchmarking, this can help reproducibility by e.g. logging the versions of key Python packages, or even all packages loaded into the global environment. Other captured metadata can include CPU and RAM usage, environment variables, and hardware specifications.

Requirements

Microbench by default has no dependencies outside of the Python standard library, although pandas is recommended to examine results. However, some mixins (extensions) have specific requirements:

  • The line_profiler package needs to be installed for line-by-line code benchmarking.
  • MBInstalledPackages requires setuptools, which is not a part of the standard library, but is usually available.
  • The CPU cores, total RAM, and telemetry extensions require psutil.
  • The NVIDIA GPU plugin requires the nvidia-smi utility, which usually ships with the NVIDIA graphics card drivers. It needs to be on your PATH.

Installation

To install using pip:

pip install microbench

Usage

Microbench is designed for benchmarking Python functions. These examples will assume you have already defined a Python function myfunction that you wish to benchmark:

def myfunction(arg1, arg2, ...):
    ...

Minimal example

First, create a benchmark suite, which specifies the configuration and information to capture.

Here's a minimal, complete example:

from microbench import MicroBench

basic_bench = MicroBench()

To attach the benchmark to your function, simply use basic_bench as a decorator, like this:

@basic_bench
def myfunction(arg1, arg2, ...):
    ...

That's it! When myfunction() is called, metadata will be captured into a io.StringIO() buffer, which can be read as follows (using the pandas library):

import pandas as pd
results = pd.read_json(basic_bench.outfile, lines=True)

The above example captures the fields start_time, finish_time, run_durations (of each function call, in seconds by default), function_name, timestamp_tz (timezone name, see Timezones section of this README), and duration_counter (the name of the function used to calculate durations, see Duration timings section of this README).

Microbench can capture many other types of metadata from the environment, resource usage, and hardware, which are covered below.

Extended examples

Here's a more complete example using mixins (the MB prefixed class names) to extend functionality. Note that keyword arguments can be supplied to the constructor (in this case some_info=123) to specify additional information to capture. We also specify iterations=3, which means that the called function with be executed 3 times (the returned result will always be from the final run) with timings captured for each run. We specify a custom duration counter, time.monotonic instead of the default time.perf_counter (see Duration timings section later in this README for explanation). This example also specifies the outfile option, which appends metadata to a file on disk.

from microbench import *
import numpy, pandas, time

class MyBench(MicroBench, MBFunctionCall, MBPythonVersion, MBHostInfo):
    outfile = '/home/user/my-benchmarks'
    capture_versions = (numpy, pandas)  # Or use MBGlobalPackages/MBInstalledPackages
    env_vars = ('SLURM_ARRAY_TASK_ID', )

benchmark = MyBench(some_info=123, iterations=3, duration_counter=time.monotonic)

The env_vars option from the example above specifies a list of environment variables to capture as env_<variable name>. In this example, the slurm array task ID will be stored as env_SLURM_ARRAY_TASK_ID. Where the environment variable is not set, the value will be null.

To capture package versions, you can either specify them individually (as above), or you can capture the versions of every package in the global environment. In the following example, we would capture the versions of microbench, numpy, and pandas automatically.

from microbench import *
import numpy, pandas

class Bench2(MicroBench, MBGlobalPackages):
    outfile = '/home/user/bench2'

bench2 = Bench2()

If you want to go even further, and capture the version of every package available for import, there's a mixin for that:

from microbench import *

class Bench3(MicroBench, MBInstalledPackages):
    pass

bench3 = Bench3()
Mixin Fields captured
(default) start_time
finish_time
function_name
MBGlobalPackages package_versions, with entry for every package in the global environment
MBInstalledPackages package_versions, with entry for every package available for import
MBCondaPackages conda_versions, with entry for every conda package in the environment
MBFunctionCall args (positional arguments)
kwargs (keyword arguments)
MBReturnValue Wrapped function's return value
MBPythonVersion python_version (e.g. 3.6.0) and python_executable (e.g. /usr/bin/python, which should indicate any active virtual environment)
MBHostInfo hostname
operating_system
MBHostCpuCores cpu_cores_logical (number of cores, requires psutil)
MBHostRamTotal ram_total (total RAM in bytes, requires psutil)
MBNvidiaSmi Various NVIDIA GPU fields, detailed in a later section
MBLineProfiler line_profiler containing line-by-line profile (see section below)

Examine results

Each result is a JSON object. When using the outfile option, a JSON object for each @benchmark call is stored on a separate line in the file. The output from the minimal example above for a single run will look similar to the following:

{"start_time": "2018-08-06T10:28:24.806493+00:00", "finish_time": "2018-08-06T10:28:24.867456+00:00", "run_durations": [0.60857599999999999], "function_name": "my_function", "timestamp_tz": "UTC", "duration_counter": "perf_counter"}

Start and finish times are given as timestamps in ISO-8601 format, in the UTC timezone by default (see Timezones section of this README).

Run_durations are given in seconds, captured using the time.perf_counter function by default, but this can be overridden (see Duration timings section of this README).

The simplest way to examine results in detail is to load them into a pandas dataframe:

# Read results directly from active benchmark suite
benchmark.get_results()

# Or, equivalently when using a file, read it using pandas directly
import pandas
results = pandas.read_json('/home/user/my-benchmarks', lines=True)

Pandas has powerful data manipulation capabilities. For example, to calculate the average runtime by Python version:

# Calculate overall runtime
results['runtime'] = results['finish_time'] - results['start_time']

# Average overall runtime by Python version
results.groupby('python_version')['runtime'].mean()

Many more advanced operations are available. The pandas tutorial is recommended.

Line profiler support

Microbench also has support for line_profiler, which shows the execution time of each line of Python code. Note that this will slow down your code, so only use it if needed, but it's useful for discovering bottlenecks within a function. Requires the line_profiler package to be installed (e.g. pip install line_profiler).

from microbench import MicroBench, MBLineProfiler
import pandas

# Create our benchmark suite using the MBLineProfiler mixin
class LineProfilerBench(MicroBench, MBLineProfiler):
    pass

lpbench = LineProfilerBench()

# Decorate our function with the benchmark suite
@lpbench
def my_function():
    """ Inefficient function for line profiler """
    acc = 0
    for i in range(1000000):
        acc += i

    return acc

# Call the function as normal
my_function()

# Read the results into a Pandas DataFrame
results = lpbench.get_results()

# Get the line profiler report as an object
lp = MBLineProfiler.decode_line_profile(results['line_profiler'][0])

# Print the line profiler report
MBLineProfiler.print_line_profile(results['line_profiler'][0])

The last line of the previous example will print the line profiler report, showing the execution time of each line of code. Example:

Timer unit: 1e-06 s

Total time: 0.476723 s
File: /home/user/my_test.py
Function: my_function at line 12

Line #      Hits         Time  Per Hit   % Time  Line Contents
==============================================================
    12                                               @lpbench
    13                                               def my_function():
    14                                                   """ Inefficient function for line profiler """
    15         1          2.0      2.0      0.0          acc = 0
    16   1000001     217874.0      0.2     45.7          for i in range(1000000):
    17   1000000     258846.0      0.3     54.3              acc += i
    18
    19         1          1.0      1.0      0.0          return acc

NVIDIA GPU support

Attributes about NVIDIA GPUs can be captured using the MBNvidiaSmi plugin. This requires the nvidia-smi utility to be available in the current PATH.

By default, the gpu_name (model number) and memory.total attributes are captured. Extra attributes can be specified using the class or object-level variable nvidia_attributes. To see which attributes are available, run nvidia-smi --help-query-gpu.

By default, all installed GPUs will be polled. To limit to a specific GPU, specify the nvidia_gpus attribute as a tuple of GPU IDs, which can be zero-based GPU indexes (can change between reboots, not recommended), GPU UUIDs, or PCI bus IDs. You can find out GPU UUIDs by running nvidia-smi -L.

Here's an example specifying the optional nvidia_attributes and nvidia_gpus fields:

from microbench import MicroBench, MBNvidiaSmi

class GpuBench(MicroBench, MBNvidiaSmi):
    outfile = '/home/user/gpu-benchmarks'
    nvidia_attributes = ('gpu_name', 'memory.total', 'pcie.link.width.max')
    nvidia_gpus = (0, )  # Usually better to specify GPU UUIDs here instead

gpu_bench = GpuBench()

Telemetry support

We use the term "telemetry" to refer to metadata which is captured periodically during the execution of a function by a thread which runs in parallel. For example, this may be useful to see how memory usage changes over time.

Telemetry support requires the psutil library.

Microbench launches and cleans up the monitoring thread automatically. The end user only needs to define a telemetry static method, which accepts a psutil.Process object and returns the telemetry data as a dictionary.

The default telemetry collection interval is every 60 seconds, which can be customized if needed using the telemetry_interval class variable.

A minimal example to capture memory usage every 90 seconds is shown below:

from microbench import MicroBench

class TelemBench(MicroBench):
    telemetry_interval = 90

    @staticmethod
    def telemetry(process):
        return process.memory_full_info()._asdict()

telem_bench = TelemBench()

Extending microbench

Microbench includes a few mixins for basic functionality as described in the extended example, above.

You can also add functions to your benchmark suite to capture extra information at runtime. These functions must be prefixed with capture_ for them to run automatically before the function starts, or capturepost_ for them to run automatically when the function completes. They take a single argument, bm_data, a dictionary to be extended with extra data. Care should be taken to avoid overwriting existing key names.

Here's an example to capture the machine type (i386, x86_64 etc.):

from microbench import MicroBench
import platform

class Bench(MicroBench):
    outfile = '/home/user/my-benchmarks'

    def capture_machine_platform(self, bm_data):
        bm_data['platform'] = platform.machine()

benchmark = Bench()

Extending JSONEncoder

Microbench encodes data in JSON, but sometimes Microbench will encounter data types (like custom objects or classes) that are not encodable as JSON by default (usually meaning they don't have a way to be represented as a string, list, or dictionary). For example, when using the MBFunctionCall and MBReturnValue, a warning will be shown if any argument or return value (respectively) is not encodable as JSON, and the value will be replaced with a placeholder to allow the metadata capture to continue, and a warning will be shown.

If you wish to actually capture those values, you will need to specify a way to convert the object to JSON. This is done using by extending microbench.JSONEncoder with a test for the object type and implementing a conversion to a string, list, or dict.

For example, to capture a Graph object from the igraph package using str(graph) as the representation, we could do the following (note that we could use any representation we want, e.g. if we wanted to capture the object in a more or less detailed way):

import microbench as mb
from igraph import Graph

# Extend the JSONEncoder to encode Graph objects
class CustomJSONEncoder(mb.JSONEncoder):
    def default(self, o):
        # Encode igraph.Graph objects as strings
        if isinstance(o, Graph):
            return str(o)

        # Add further isinstance(o, ...) cases here
        # if needed

        # Make sure to call super() to handle
        # default cases
        return super().default(o)

# Define your benchmark class as normal
class Bench(mb.MicroBench, mb.MBReturnValue):
    pass

# Create a benchmark suite with the custom JSON
# encoder from above
bench = Bench(json_encoder=CustomJSONEncoder)

# Attach the benchmark suite to our function
@bench
def return_a_graph():
    return Graph(2, ((0, 1), (0, 2)))

# This should now work without warnings or errors
return_a_graph()

Redis support

By default, microbench appends output to a file, but output can be directed elsewhere, e.g. redis - an in-memory, networked data source. This option is useful when a shared filesystem is not available.

Redis support requires redis-py.

To use this feature, inherit from MicroBenchRedis instead of MicroBench, and specify the redis connection and key name as in the following example:

from microbench import MicroBenchRedis

class RedisBench(MicroBenchRedis):
    # redis_connection contains arguments for redis.StrictClient()
    redis_connection = {'host': 'localhost', 'port': 6379}
    redis_key = 'microbench:mykey'

benchmark = RedisBench()

To retrieve results, the redis package can be used directly:

import redis
import pandas

# Establish the connection to redis
rconn = redis.StrictRedis(host=..., port=...)

# Read the redis data from 'myrediskey' into a list of byte arrays
redis_data = redis.lrange('myrediskey', 0, -1)

# Convert the list into a single string
json_data = '\n'.join(r.decode('utf8') for r in redis_data)

# Read the string into a pandas dataframe
results = pandas.read_json(json_data, lines=True)

Runtime impact considerations

The runtime impact varies depending on what information is captured and by platform. Broadly, capturing environment variables, Python package versions, and timing information for a function has a negligible impact. Capturing telemetry and invoking external programs (like nvidia-smi for GPU information) has a larger impact, although the latter is a one-off per invocation and typically less than one second. Telemetry capture intervals should be kept relatively infrequent (e.g., every minute or two, rather than every second) to avoid significant runtime impacts.

Duration timings

By default, run_durations are given in seconds using the time.perf_counter function, which should be sufficient for most use cases. You can use any function that returns a float or integer number for time. Some examples would be time.perf_counter_ns if you want time in nanoseconds, or time.monotonic for a monotonic clock impervious to clock adjustments (ideal for very long-running code). Use the duration_counter=... argument when creating a benchmark suite object, as seen in the Extended examples section of this README, above.

Timezones

Microbench captures start_time and finish_time in the UTC timezone by default. This can be overriden by passing a tz=... argument when creating a benchmark class, where the value is a timezone object (e.g. created using the pytz library).

Feedback

Please note this is a recently created, experimental package. Please let me know your feedback or feature requests in Github issues.