README.md

Weld benchmarks

This repository tries to run benchmarks for the Weld project in an easy-to-reproduce way.

Key requirements

• Each benchmark is in a separate directory under benchmarks/.

• All benchmark binaries have to to be named bench in the corresponding directory (if compilation is required, the output binary produced by make should be called bench).

• In addition, each benchmark binary is responsible for its own timing; all timing information needs to be printed to stdout in the following format,

  <Experiment description>: <Time> <Other metadata>
  

  Each benchmark binary can print multiple pieces of timing information.

  To make this more concrete, consider the output of the matrix multiplication benchmark,

  $ ./bench
  Transposed: 0.028259 (result=38238491904)
  Unblocked: 0.029373 (result=38288406288)
  Blocked: 0.028529 (result=38228352573)

• All parameters need to be passed into the bench binary in the form -<parameter name> <parameter value>.

• Benchmark directories must include a configuration file named "config.json".

  A sample configuration file looks like this:

  {
      "compile": true,
      "params":
      {
        "b": [64, 128]
      },
      "scaled_params":
      {
        "n": [512, 1024]
      }
  }

  The params field specifies the parameters that need to be swept over.

  The compile field is a true/false field and specifies whether workloads need to be compiled beforehand using make or not.

Running instructions

The main script is run_benchmarks.py in the root directory. It takes the following arguments:

• -n / --num_iterations: Specifies the number of trials for each benchmark.
• -s / --scale_factor: Specifies the factor by which scaled parameters need to be scaled.
• -f / --csv_filename: Specifies the output file for dumped experiment results.
• -b / --benchmarks: Comma-separated list of benchmarks that should be run (must be a subset of benchmarks listed in the configuration file).
• -v / --verbose: A flag specifying whether to print verbose statistics.

Sample output looks like this:

$ python run_benchmarks.py -b crime_index -n 5 -f results.csv -v
++++++++++++++++++++++++++++++++++++++
crime_index
++++++++++++++++++++++++++++++++++++++
s=1, f=data/us_cities_states_counties_sf=%d.csv
Grizzly: 1.4256 +/- 0.0102 seconds
Pandas: 0.0229 +/- 0.0009 seconds

s=10, f=data/us_cities_states_counties_sf=%d.csv
Grizzly: 2.1567 +/- 0.2303 seconds
Pandas: 0.0725 +/- 0.0066 seconds

s=100, f=data/us_cities_states_counties_sf=%d.csv
Grizzly: 6.8954 +/- 0.2109 seconds
Pandas: 0.6040 +/- 0.0640 seconds