Skip to content

Latest commit

 

History

History
212 lines (138 loc) · 13.4 KB

GlobalRunner.md

File metadata and controls

212 lines (138 loc) · 13.4 KB
Raw

Global Runner

The global runner described in Architecture.md is implemented as a PowerShell script ADBench/run-all.ps1. This document describes its exact functionality, interfaces with other components of ADBench, and provides a user guide.

Functionality

When invoked, the script will run a set of benchmarks on a set of testing modules. Both sets have default values, that can be overridden by passing specific arguments to the script.

The default set of benchmarks includes:

  • All GMM problems with 1k and 10k points defined in /data/gmm folder
  • 5 (out of 20 present in /data/ba) smallest BA problems
  • 5 (out of 12 present in /data/hand) Hand Tracking problems of every kind (simple-small, simple-big, complicated-small, complicated-big)
  • All LSTM problems defined in /data/lstm folder

The default set of testing modules includes all testing modules registered with the script.

Unless excluded, run-all.ps1 will run benchmarks on the Manual module first in order to produce the golden Jacobians, which will be used to check the correctness of the Jacobians produced by all other modules. See Jacobian Correctness Verification for the details and the justification.

For every specified testing module run-all.ps1 will run every specified benchmark that this module supports. So, e.g. if BA benchmarks were requested, but some of the requested modules don't support BA, then BA benchmarks won't be run on these modules and such behavior won't be considered an error.

The script will run benchmarks on testing modules using the corresponding benchmark runners. It will enforce the hard timeout (default - 5 minutes, overridable in the corresponding argument) and RAM consuming control (it isn't performed by default, overridable in the corresponding argument) by terminating the runner processes as necessary. It will check that the runner produces a file with timings and a file with the result Jacobian. Then it will compare the result Jacobian to the golden one and output the result of the comparison in JSON format. Unless overridden in the corresponding argument, run-all.ps1 will delete files with Jacobians it considered correct during comparison.

The script checks guaranteed timeouts. This means that the benchmark for the bigger test will not be run if the run test for the smaller size finishes with timeout. Thus, a GMM objective test will not be run in any case where the d and K values are both greater than or equal to the d and K values of a previously timed-out test (of the same count of points). For LSTM such a timeout checking is the same (but for l and c instead of d and K respectively). BA and Hand tests with the bigger order number will not be run if any test with the less number is terminated due to timeout. So, for them tests are expected to be sorted respective their sizes. Guaranted out of memory checking, that is also performed by the script, has the same scenario.

Timeouts, out of memory, missing output files, and failed comparisons to the golden Jacobians are all considered to be non-fatal errors. They cause a warning to be printed when the script finishes and make script's exit code to become non-zero, but don't prevent the execution of other benchmarks.

Usage

Run from PowerShell command prompt. Syntax:

    ./run-all.ps1 [[-buildtype] <String>] [[-minimum_measurable_time] <Double>] [[-nruns_f] <Int32>] [[-nruns_J] <Int32>] [[-time_limit] <Double>] [[-timeout] <Double>] [[-max_memory_amount_in_gb] <Double>] [[-tmpdir] <String>] [-repeat] [-repeat_failures] [[-tools] <String[]>] [-keep_correct_jacobians] [[-gmm_d_vals_param] <Int32[]>] [[-gmm_k_vals_param] <Int32[]>] [[-gmm_sizes] <String[]>] [[-hand_sizes] <String[]>] [[-ba_min_n] <Int32>] [[-ba_max_n] <Int32>] [[-hand_min_n] <Int32>] [[-hand_max_n] <Int32>] [[-lstm_l_vals] <Int32[]>] [[-lstm_c_vals] <Int32[]>]

Parameters:

  • -buildtype <String>

    Which build to test. Builds should leave a script file cmake-vars-$buildtype.ps1 in the ADBench directory, which sets $bindir to the build directory. And if only some D,K are valid for GMM, sets $gmm_d_vals, and $gmm_k_vals.

  • -minimum_measurable_time <Double>

    Estimated time of accurate result achievement. A runner cyclically reruns measured function until total time becomes more than that value. Supported only by benchmark the runner-based tools (those with ToolType cpp, dotnet, julia, or python).

  • -nruns_f <Int32>

    Maximum number of times to run the function for timing.

  • -nruns_J <Int32>

    Maximum number of times to run the Jacobian for timing.

  • -time_limit <Double>

    How many seconds to wait before we believe we have accurate timings.

  • -timeout <Double>

    Kill the test after this many seconds.

  • -max_memory_amount_in_gb <Double>

    Kill the test if it consumes more than this amount of gigabytes of RAM. This parameter equals to positive infinity by default, so the out of memory checking isn't performed.

  • -tmpdir <String>

    Where to store the ouput, defaults to tmp/ in the project root.

  • -repeat [<SwitchParameter>]

    Repeat tests, even if output file exists.

  • -repeat_failures [<SwitchParameter>]

    Repeat only failed tests.

  • -tools <String[]>

    List of tools to run.

  • -keep_correct_jacobians [<SwitchParameter>]

    Don't delete produced jacobians even if they're accurate.

  • -gmm_d_vals_param <Int32[]>

    GMM D values to try. Must be a subset of the list of compiled values in ADBench/cmake-vars-$buildtype.ps1.

  • -gmm_k_vals_param <Int32[]>

    GMM K values to run. As above.

  • -gmm_sizes <String[]>

    GMM sizes to try. Must be a subset of @("1k", "10k", "2.5M"). 2.5M currently is not supported.

  • -hand_sizes <String[]>

    Hand problem sizes to try. Must be a subset of @("small", "big").

  • -ba_min_n <Int32>

    Number of the first BA problem to try. Must be between 1 and ba_max_n.

  • -ba_max_n <Int32>

    Number of the last BA problem to try. Must be between ba_min_n and 20.

  • -hand_min_n <Int32>

    Number of the first Hand problem to try. Must be between 1 and hand_max_n.

  • -hand_max_n <Int32>

    Number of the last Hand problem to try. Must be between hand_min_n and 12.

  • -lstm_l_vals <Int32[]>

    Numbers of layers in LSTM to try. Must be a subset of @(2, 4).

  • -lstm_c_vals <Int32[]>

    Sequence lengths in LSTM to try. Must be a subset of @(1024, 4096).

Example:

./run-all.ps1 -buildtype "Release" -minimum_measurable_time 0.5 -nruns_f 10 -nruns_J 10 -time_limit 180 -timeout 600 -tmpdir "C:/path/to/tmp/" -tools @("Finite", "Manual", "PyTorch") -gmm_d_vals_param @(2,5,10,64)

This will:

  • Run only release builds.
  • Loop measured function while total calculation time is less than 0.5 seconds.
  • Aim to run 10 tests of each function, and 10 tests of the derivative of each function.
  • Stop (having completed a whole number of tests) at any point after 180 seconds.
  • Allow each program a maximum of 600 seconds to run all tests.
  • Output results to C:/path/to/tmp/.
  • Not repeat any tests for which there already exist a results file.
  • Run only Finite, Manual, and PyTorch.
  • Try GMM d values of 2, 5, 10, 64.

Interfaces

This section describes, how run-all.ps1 interfaces with other components of ADBench.

Interfacing with Benchmark Runners

run-all.ps1 is aware of and knows how to invoke all benchmark runners.

First, there's an enumeration ToolType that lists names of runners. These names are local to the script. Then, the method [Tool]::run(...) contains an if...elseif... block, that has a clause for every runner, in which that runner is invoked to perform a specific benchmark. These are the two places, one would need to modify to make run-all.ps1 support a new runner.

When the benchmark runner finishes, run-all.ps1 checks that it produced a file with timings (<name of the input>_times_<name of the testing module>.txt) and a file with the result Jacobian (<name of the input>_J_<name of the testing module>.txt) in the specified folder. Any of these files missing is a non-fatal error. Then it checks the correctness of the Jacobian. For that to be possible, the file with the Jacobian must have a format described in FileFormat.md.

Interfacing with Testing Modules

While run-all.ps1 does not interface with the testing modules directly - we have benchmark runners for that, all testing modules still must be listed in this script for it to be aware of their existence, supported objectives, and required benchmark runner.

To add a new testing module to run-all.ps1, find an array $tool_descriptors near the end of the script and add a new [Tool] object to it. The syntax of the Tool's constructor is

[Tool]::new("<Testing module name>", "<Benchmark runner name>", [ObjectiveType] "<Supported objectives>", $true, <tolerance>)

Here

  • <Testing module name> is the name of your module,
  • <Benchmark runner name> is the name of the benchmark runner that can invoke your module as listed in the ToolType enumeration,
  • <Supported objectives> is a comma-separated list of objective function names supported by your module. Possible names are GMM, BA, Hand, and LSTM,
  • <tolerance> is the maximum error, values of the Jacobians produced by your module are allowed to have. Use $default_tolerance unless there's some specific reason for your module to produce results of non-standard accuracy. See JacobianCheck.md for the definition of error.

Output

run-all.ps1 outputs its logs into standard output.

For every benchmark run-all.ps1 performs, it produces a number of file outputs. These files are placed in the following folder

/<path to tmp>/<build config>/<objective>/<objective subtype>/<testing module>/

Here

  • <path to tmp> is the path passed to the script in the tmpdir parameter, which defaults to tmp folder in the root of the repository.
  • <build config> is the configuration of the build passed to the script in the buildtype parameter, which defaults to Release.
  • <objective> is the short name of the objective function in lowercase (so it's one of "gmm", "ba", "hand", and "lstm").
  • <objective subtype> is an optional subtype specific to the objective. For the GMM objective it's the number of points ("1k", "10k", or "2.5M"), for the Hand objective it's <complexity>_<size>, where <complexity> is either "simple" or "complicated", and <size> is either "big" or "small". BA and LSTM have no subtypes.
  • <testing module> is the name of the testing module.

The outputs themselves are:

  • <name of the input>_times_<name of the testing module>.txt - new line-separated timings for the computation of the objective function and the derivative. Produced by the benchmark runner, unless it timed out, in which case run-all.ps1 produces this file by itself (with inf values for both times).
  • <name of the input>_F_<name of the testing module>.txt - new line-separated values of the objective function computed by the module. Produced by the benchmark runner.
  • <name of the input>_J_<name of the testing module>.txt - values of the derivative computed by the module. Exact format is specific to the objective function. See FileFormat for details. Produced by the benchmark runner. Unless explicitly instructed otherwise, run-all.ps1 deletes these files if they pass the correctness test.
  • <name of the input>_correctness_<name of the testing module>.txt - JSON files with the results of correctness checking. Produced by run-all.ps1. These files have the following format:
{
    "Tolerance": <double>,
    "File1": "/path/to/Jacobian/being/checked.txt",
    "File2": "/path/to/golden/Jacobian.txt",
    "DimensionMismatch": <bool>,
    "ParseError": <bool>,
    "MaxDifference": <double>,
    "AvgDifference": <double>,
    "DifferenceViolationCount": <int>,
    "NumberComparisonCount": <int>,
    "Error": "Text of the error, that caused the termination of the comparison, if any",
    "ViolationsHappened": <bool>
}

Here

  • Tolerance is the maximum difference between the values of compared Jacobians that was not considered an error.
  • DimensionMismatch is true, when the compared Jacobians have different sizes, false otherwise.
  • ParseError is true, when the parsing of at least one of the compared files ended in an error, false otherwise.
  • MaxDifference is the maximum difference encountered while comparing the corresponding values of the two Jacobians.
  • AvgDifferenceis the average difference encountered while comparing the corresponding values of the two Jacobians.
  • DifferenceViolationCount is the number of times the difference encountered while comparing the corresponding values of the two Jacobians exceeded the Tolerance.
  • NumberComparisonCount is the number of times the corresponding values of the two Jacobians were compared before the comparison ended (possibly, due to an error).
  • ViolationsHappened is true if DifferenceViolationCount is non-zero or if an error happened during the comparison.