arrowbench

This R package contains tools for defining benchmarks, running them across a range of parameters, and reporting their results in a standardized form. It also contains some benchmark code for measuring performance of Apache Arrow and other projects one might compare it to.

The purpose of the package is to provide developers with better tools for creating, parametrizing, and reproducing benchmarks across a range of library versions, variables, and machines, as well as to facilitate continuous monitoring. While this package could be used for microbenchmarking, it is designed specially for "macrobenchmarks": workflows that real users do with real data that take longer than microseconds to run.

It builds on top of existing R benchmarking tools, notably the bench package. Among the features that this package adds are

Setup designed with parametrization in mind so you can test across a range of variables, which may not all be valid in combination
Isolation of benchmark runs in separate processes to prevent cross-contamination such as global environment changes and previous memory allocation
Tools for bootstrapping package versions and known data sources to facilitate running the same code on different machines

User guide

Installation

The quickest and easiest way to install is to run remotes::install_github("ursacomputing/arrowbench", dependencies = TRUE) in R. If you need to install remotes you can install.packages("remotes").

If you've downloaded the source, or you're making changes to arrow bench you should make sure that you have the dependencies with remotes::install_deps(".", dependencies = TRUE) in R (this will also install the arrow package along with other packages that can be benchmarked with arrowbench. And then running R CMD INSTALL . in a terminal (for both, you should do this in the root directory of arrowbench, or pass the path to arrowbench instead of .).

Running benchmarks

Pass a Benchmark to run_benchmark() and it will run it across the range of parameters specified. For parameters specified in bm$setup that are omitted when calling run_benchmark(bm), it will test across all combinations of them (what we call a benchmark matrix). If some parameter combinations are not valid, define a bm$valid_params(params) function that will filter that expanded data.frame of parameters down to the valid set.

For example,

library(arrowbench)

run_benchmark(write_file, source = "nyctaxi_2010-01")

will run the write_file benchmark with "nyctaxi_2010-01" source file on the Cartesian product of the other function parameters--format, compression, and input--along with cpu_counts of c(1, Ncpus).

Another example:

library(arrowbench)

run_benchmark(write_file, source = "nyctaxi_2010-01", writer = "feather", 
  input = "data.frame", cpu_count = c(1, 4, 8))

will run only the Feather writing tests with the two valid compression variants, each one done for 1, 4, and 8 threads, for a total of 6 runs.

If lib_path is not provided to run_benchmark(), it will use the default .libPath and whatever is installed there. You can also indicate a subset of released x.y Arrow version numbers, or lib_path = "all" to test all past releases of arrow plus "latest".

Run options

run_benchmark() handles executing benchmarks across a range of parameters. After determining the valid parameters, it calls run_one() on each and collects the results. run_one() generates an R script and then shells out to a separate R process to execute the benchmark, then collects the results from it.

You may call run_one() directly. It takes some options, which may be passed from run_benchmark() (both default FALSE):

n_iter: the number of iterations to run the specific benchmark (default: 3)
dry_run: logical, returns the R script instead of executing it. Useful for debugging, though you probably don't want to execute the script yourself in order to do the benchmarking: run_script(), which run_one() calls when dry_run = FALSE, has some useful wrapping for caching and collecting results.
profiling: logical, allows you you instrument the R code and collect profiling data. You don't want to do this if you're collecting benchmark data because the profiler can add to the run time, but if you see something slow and want to explore why, this can be a good start. Note that this doesn't do profiling at the C++ level, so if something is slow inside an arrow C++ function, this won't tell you what exactly, but it can help rule things out. If TRUE, the result data will contain a prof_file field, which you can read in with profvis::profvis(prof_input = file).
read_only: don't actually run benchmarks, but read any results that are in the results directory.

Defining benchmarks

Benchmarks are constructed by Benchmark(), which takes expressions that handle setup, teardown, and the actual work that we want to measure. See its documentation for details, and see read_file and write_file for examples.

A (contrived) example

Here we create a new kind of csv benchmark that uses arrow::read_csv_arrow() and varies the arguments as_data_frame and skip_empty_rows. This is a bit of a contrived example, see read_csv for how we actually test csv reading. There are comments in the code block explaining what each section does.

new_csv_benchmark <- Benchmark(
  "new_csv_benchmark",
  # This setup block will be run before the benchmark is started. This is run
  # before each case / single item in the benchmark matrix, so is a good time
  # to setup case- or source-specific properties (see `result_dim` below).
  setup = function(source = names(known_sources),
                   as_data_frame = c(TRUE, FALSE),
                   skip_empty_rows = TRUE)) {
    # Validate the parameters
    # For our benchmark: as_data_frame defaults to TRUE and FALSE (so if it is 
    # unspecified you will get both TRUE and FALSE in the benchmark matrix)
    as_data_frame <- match.arg(as_data_frame)

    # For our benchmark: skip_empty_rows defaults to TRUE (so if it is 
    # unspecified you will only get TRUE in the benchmark matrix), but it can 
    # accept TRUE or FALSE, so we validate that it is one of those.
    skip_empty_rows <- match.arg(skip_empty_rows, c(TRUE, FALSE))

    # Ensure the file exists as an uncompressed csv
    input_file <- ensure_format(source, "csv", "uncompressed")

    # Extract the dim attribute from a data source for validation later
    result_dim <- get_source_attr(source, "dim")

    # Finally we return a `BenchEnvironment` with the parameters we defined 
    # above that are needed
    BenchEnvironment(
      input_file = input_file,
      result_dim = result_dim,
      as_data_frame = as_data_frame,
      skip_empty_rows = skip_empty_rows
    )
  },
  # This is run before each iteration. 
  before_each = {
    # Make sure the result is cleared
    result <- NULL
  },
  # This is the only part of the code that is actually measured when the benchmark 
  # is run. It should include all and only the code you are interested in benchmarking.
  run = {
    result <- read_csv_arrow(
      input_file, 
      as_data_frame = as_data_frame, 
      skip_empty_rows = skip_empty_rows
    )
  },
  # This is run after each iteration. This is a good time to validate that the
  # benchmark ran correctly.
  after_each = {
    stopifnot(
      "The dimensions do not match" = all.equal(dim(result), result_dim)
    )
    result <- NULL
  },
  # This defines if the parameters are valid. If there are certain combinations
  # that are not valid, add them to drop and they will be excluded from the 
  # benchmark matrix
  valid_params = function(params) {
  
    # Do not allow both skip_empty_rows == FALSE and as_data_frame == TRUE at 
    # the same time
    drop <- ( params$skip_empty_rows == FALSE & params$as_data_frame == TRUE ) 
    
    params[!drop,]
  },
  # This lists any packages that are used by this benchmark so that they can 
  # be installed prior to starting the run. Typically this will be simply "arrow"
  packages_used = function(params) {
    "arrow"
  }
)

And now we could run our benchmark with the following for the default matrix, using 4 cpu cores and 5 iterations per case.

run_benchmark(
  new_csv_benchmark,
  cpu_count = 4,
  n_iter = 5
)

Or specify parameters (including non-default parameters) with:

run_benchmark(
  new_csv_benchmark,
  as_data_frame = c(TRUE, FALSE),
  skip_empty_rows = c(TRUE, FALSE),
  cpu_count = 4,
  n_iter = 5
)

Enabling benchmarks to be run on conbench

Conbench is a service that runs benchmarks continuously on a repo. We have a conbench service setup to run benchmarks on the apache/arrow repository (and pull requests, if requested).

Before a benchmark can be run on conbench, one must add a (or extend an existing) benchmark in the benchmarks python package. If you are adding a new benchmark see the R-only example external benchmarks in benchmarks. An example of adding an R-only benchmark is benchmarks#14

Known data sources and versions

The package knows about certain large data files to use in benchmarks. These are registered in a known_sources object, which specifies where they can be downloaded and how to read them, as well as optional attributes about them (e.g. dim()) that can be used to validate that they've been read in correctly.

To use them in benchmarks, use the ensure_source() function to take a source identifier and mapping it to a file path, downloading and extracting the file if it isn't found. Pass the result to read_source() load the data with the source's provided reader function.

Source files are cached in a data directory and are only downloaded if not present. This speeds up repeat benchmark runs on the same host. By default, data is assumed to be relative to the current working directory, but you can set the environment variable ARROWBENCH_DATA_DIR to point to another (permanent) base directory.

Similarly, there is an ensure_lib() function called in the global_setup() that supports a list of known arrow package versions, which are mapped to daily snapshots of CRAN hosted by Microsoft. If you specify lib_path = "0.17", for example, ensure_lib() will use a .libPath for this version and install all Suggested packages into that directory using the MRAN snapshot for "2020-05-29", a date when 0.17 was the arrow version on CRAN. This lets you test against old versions of the code and to backfill benchmark results.

These versioned R package libraries are cached in an r_libs directory, like data relative to the directory specified by the environment variable ARROWBENCH_LOCAL_DIR.

Results and caching

run_benchmark() returns a list of benchmark results, which may be massaged, JSON-serialized, and uploaded to the conbench service. Within an R process, you can call as.data.frame() on it to get a more manageable view, which can be passed to plotting functions.

In addition to timings, parameter values, and the versions of loaded packages, the benchmark results contain some extra data on memory usage and garbage collection. gc() can add significant time to large operations, and while we can't prevent it, we can at least be aware of when it is happening.

Individual benchmark results (the output of run_one()) are cached in a results directory. This way, if the main process running run_benchmark() fails or is interrupted in the middle, you can restart. Note however that if you are using the default lib_path and are updating the package versions installed there between benchmark runs, you should clear the cache before starting a new run (at least deleting the cached .json files containing "latest" in the file name). The location of this cache is the directory specified by the environment variable ARROWBENCH_LOCAL_DIR. If no environment variable is given, this will default to the current working directory.

ahadnagy / arrowbench