MLRsearch 1.2.1

Library for extending and speeding up througput search.

Origins

This library was developed as a speedup for traditional binary search in CSIT (Continuous System and Integration Testing) project of fd.io (Fast Data), one of LFN (Linux Foundation Networking) projects.

In order to make this code available in PyPI (Python Package Index), the setuputils stuff has been added, but after some discussion, the export directory is only a symlink to the original place of tightly coupled CSIT code.

Change log

1.2.1: Updated the readme document.

1.2.0: Changed the output structure to use Goal Result as described in draft-05.

1.1.0: Logic improvements, independent selectors, exceed ratio support, better width rounding, conditional throughput as output. Implementation relies more on dataclasses, code split into smaller files. API changed considerably, mainly to avoid long argument lists.

0.4.0: Considarable logic improvements, more than two target ratios supported. API is not backward compatible with previous versions.

0.3.0: Migrated to Python 3.6, small code quality improvements.

0.2.0: Optional parameter “doublings” has been added.

0.1.1: First officially released version.

Usage

High level description

A complete application capable of testing performance using MLRsearch consists of three layers: Manager, Controller and Measurer. This library provides an implementation for the Controller only, including all the classes needed to define API between Controller and other two components.

Users are supposed to implement the whole Manager layer, and also implement the Measurer layer. The Measurer instance is injected as a parameter when the manager calls the controller instance.

The purpose of Measurer instance is to perform one trial measurement. Upon invocation of measure() method, the controller only specifies the intended duration and the intended load for the trial. The call is done using keyword arguments, so the signature has to be:

def measure(self, intended_duration, intended_load):

Usually, the trial measurement process also needs other values, collectively caller a traffic profile. User (the manager instance) is responsible for initiating the measurer instance accordingly. Also, the manager is supposed to set up SUT, traffic generator, and any other component that can affect the result.

For specific input and output objects see the example below.

Example

This is a minimal example showing every configuration attribute. The measurer does not interact with any real SUT, it simulates a SUT that is able to forward exactly one million packets per second (unidirectional traffic only), not one packet more (fully deterministic). In these conditions, the conditional throughput for PDR happens to be accurate within one packet per second.

This is the screen capture of interactive python interpreter (wrapped so long lines are readable):

>>> import dataclasses
>>> from MLRsearch import (
...     AbstractMeasurer, Config, MeasurementResult,
...     MultipleLossRatioSearch, SearchGoal,
... )
>>>
>>> class Hard1MppsMeasurer(AbstractMeasurer):
...     def measure(self, intended_duration, intended_load):
...         sent = int(intended_duration * intended_load)
...         received = min(sent, int(intended_duration * 1e6))
...         return MeasurementResult(
...             intended_duration=intended_duration,
...             intended_load=intended_load,
...             offered_count=sent,
...             forwarding_count=received,
...         )
...
>>> def print_dot(_):
...     print(".", end="")
...
>>> ndr_goal = SearchGoal(
...     loss_ratio=0.0,
...     exceed_ratio=0.005,
...     relative_width=0.005,
...     initial_trial_duration=1.0,
...     final_trial_duration=1.0,
...     duration_sum=21.0,
...     preceding_targets=2,
...     expansion_coefficient=2,
... )
>>> pdr_goal = dataclasses.replace(ndr_goal, loss_ratio=0.005)
>>> config = Config(
...     goals=[ndr_goal, pdr_goal],
...     min_load=1e3,
...     max_load=1e9,
...     search_duration_max=1.0,
...     warmup_duration=None,
... )
>>> controller = MultipleLossRatioSearch(config=config)
>>> result = controller.search(measurer=Hard1MppsMeasurer(), debug=print_dot)
....................................................................................
....................................................................................
...................>>> print(result)
{SearchGoal(loss_ratio=0.0, exceed_ratio=0.005, relative_width=0.005, initial_trial_
duration=1.0, final_trial_duration=1.0, duration_sum=21.0, preceding_targets=2, expa
nsion_coefficient=2, fail_fast=True): fl=997497.6029392382,s=(gl=21.0,bl=0.0,gs=0.0,
bs=0.0), SearchGoal(loss_ratio=0.005, exceed_ratio=0.005, relative_width=0.005, init
ial_trial_duration=1.0, final_trial_duration=1.0, duration_sum=21.0, preceding_targe
ts=2, expansion_coefficient=2, fail_fast=True): fl=1002508.6747611101,s=(gl=21.0,bl=
0.0,gs=0.0,bs=0.0)}
>>> print(f"NDR conditional throughput: {float(result[ndr_goal].conditional_throughp
ut)}")
NDR conditional throughput: 997497.6029392382
>>> print(f"PDR conditional throughput: {float(result[pdr_goal].conditional_throughp
ut)}")
PDR conditional throughput: 1000000.6730730429
>>>

IETF documents

The currently published IETF draft describes the logic of version 1.2.0, earlier library and draft versions do not match each other that well.