MPLLang / delayed-seq

(PPoPP'22) Experiments with parallel delayed sequences

delayed-seq

This repo contains all code necessary for reproducing the experiments in the following paper:

Parallel Block-Delayed Sequences. Sam Westrick, Mike Rainey, Daniel Anderson, and Guy E. Blelloch. PPoPP 2022

It includes code for both MPL and C++, including both the delayed-sequences library and all benchmarks.

An artifact was published with the paper, and is available on Zenodo. The version of this repo used to generate the artifact is preserved on the ppopp22-artifact branch. If you are interested in reproducing experimental results from the paper, we recommending using the artifact directly.

The main branch includes improvements made to the libraries and benchmarks after the artifact was published. These changes include:

• For MPL, a cleaned-up library implementation which is easier to understand and maintain. The new library code is available at ml/lib/NewDelayedSeq.sml. When compiling a benchmark, use new-delay for the cleaned-up library. (For example, cd ml && make bfs.new-delay.mpl-latest.bin.)
• Small performance improvements to a few benchmarks.

Setup

First, install mpl-switch.

Next, run the following to install versions of mpl needed for experiments. This will take a long time.

$ ./init

Run it

The files ml-exp.json and cpp-exp.json define benchmark parameters. Do the following to run both:

$ ./run

This creates a directory results/ named with the current time. Each line of a results file is a JSON object with various info about the run.

To see a summary of results, do:

$ ./report

Notes

• In cpp/ and ml/, there needs to be a Makefile that has targets of the form BENCH.IMPL.CONFIG.bin, for example mcss.array.mpl.bin and primes.delay.cpp.bin.

• The ml-exp.json and cpp-exp.json files are passed to the scripts/gencmds utility. Ask Sam about documentation. Basically, gencmds produces a bunch of "rows" where each row has key/value pairs that define one shell command. These rows are then passed to scripts/runcmds, which runs each command and writes out the results. Important key/value pairs in the rows include:

  • cmd: the shell command to run the experiment
  • cwd: the subdirectory in which to run this command
  • tag: a unique name for this experiment
  • procs: how many processors to run on
  • bench: the name of the benchmark
  • config: the configuration used to compile the benchmark