Exceptional Actors (Hakkell)

This project explores the use of asynchronous exceptions as an inter-thread communication mechanism in GHC Haskell. We implemented the actor model of programming on top of asynchronous exceptions as a small framework: a send function and a message-receipt mainloop (to be run on each thread). The paper (main.lhs) is written as a Literate-Haskell program and was accepted at Haskell Symposium 2023. To read the paper, we recommend rendering it first or visiting https://doi.org/10.1145/3609026.3609728.

Build the program

1. NixOS users

   We developed this program on NixOS, and recommend using nix-shell to enter an environment where the makefile does the rest.

   • nix-shell --run 'make' or nix-build -- Build the paper
   • nix-shell --run 'make main.bench.elf' or nix-build -- Build the benchmark executable
   • nix-shell --run 'make bench' -- Build and run the benchmark (first read all of Benchmark the program)
   • nix-shell --run 'make prof' -- Obtain an eventlog

2. Cabal users

   For users not on NixOS, we include a cabal-file and a cabal-freeze-file. We built the project on GHC 9.0.2 with base-4.15.1.0 using Cabal 3.10.1.0.

   $ cabal update
   ...
   $ cabal v2-build
   ...
   Linking /.../hakkell.paper/dist-newstyle/build/x86_64-linux/ghc-9.0.2/hakkell-paper-0.0.0/x/main/build/main/main ...
   $ cabal v2-exec make main.bench.elf
   ...
   [1 of 1] Compiling Main             ( main.noprint.lhs, main.noprint.o )
   Linking main.bench.elf ...
   rm main.noprint.lhs

   The makefile isn't aware of the cabal exec environment and some targets have additional dependencies, so not all the targets will work under cabal exec (the clean targets do not work, for example).

Import the program

Make sure to use send and run in your programs. The other functions (sendStatic, runStatic, and runDyn) aren't meant to be used directly.

Run the program

The main method supports two modes of operation. For either mode, set the environment variable RING_SIZE to determine how many nodes will participate in the election.

1. Run demo mode

   Set the environment variable MODE to actors, channels, or control to run benchActors, benchChannels, or benchControl once, on the specified RING_SIZE nodes.

   This is the mode we used for measuring total bytes allocated.

2. Run benchmark mode

   If the environment variable MODE is not set, then the criterion benchmark will run. It sets up a benchmark group for the specified RING_SIZE containing a becnhmark of benchControl, benchActors, and benchChannels.

   This is the mode we used for measuring running time. We used a shell script that provided command-line arguments interpreted by criterion to select and run only one of those functions at a time.

Benchmark the program

An extra step is required to obtain a consistent result from benchmark mode (above). Printlines must be stripped from the source code. The included makefile has targets which attempt to control for this and other factors that may confound benchmark results.

• make main.bench.elf

  • This target uses a python script to remove printlines from main.lhs and compiles an executable which defaults to +RTS -N4.
    • You may override the capabilities default with an environment variable. For example, an +RTS -N default is obtained this way:
      env CAPABILITIES= make main.bench.elf or make main.bench.elf CAPABILITIES=
  • Running executables produced by this target is sufficient to get a consistent result on Amazon AWS EC2 instances.

• make bench

  • This target builds an executable and performs additional steps before running benchmark mode.
    • It attempts to control for various factors on a personal laptop running linux that may confound benchmark results.
    • It executes benchprep.sh to set your CPU to 1.6GHz, turns off frequency scaling, stops TLP & ThermalD, and disables Intel Pstate.
    • If you run make bench, first audit benchprep.sh or zero it out.
  • The output from make bench is make.bench.csv and make.bench.html.