Skip to content

Latest commit

 

History

History
56 lines (44 loc) · 3.99 KB

README.md

File metadata and controls

56 lines (44 loc) · 3.99 KB
Raw

Benchmarks for Saturn library.

This repository put together benchmarks for Saturn's data structures.

Running the Benchmarks

To run the benchmarks, use the following command:

dune exec -- ./src/saturn_benchmarks.exe -budget 1

To print the help information, use the -help option:

dune exec -- ./src/saturn_benchmarks.exe -help

This command will also display the names of all available benchmarks. You can filter the benchmarks to run by including part of their name at the end of the command line:

dune exec -- ./src/saturn_benchmarks.exe -budget 1 Parallel Stdlib

This will run all benchmarks with either Parallel or Stdlib in their names.

Queue implementations

Here is a list of queue implementations that can be compared :

  • the Stdlib queue (with one domain only),
  • the Stdlib queue protected with a mutex,
  • the lock-free Michael-Scott queue from Saturn,
  • a Michael-Scott two-stack-based queue (currently in this PR in Saturn).

Stacks implementations

The stack implementations benchmarked are

  • the Stdlib stack (with one domain only),
  • the Stdlib stack protected with a mutex,
  • a concurrent stack implemented with an atomic list (with and without padding and backoff mechanism)
  • a lock-free Treiber stack from Saturn.

Single-consumer single-producer queues : different optimizations

The following command line runs benchmarks for a single-consumer single-producer queue with different optimizations.

dune exec -- ./src/saturn_benchmarks.exe -budget 1 Spsc

The different benchmarked optimizations are (with results on Intel computer 12th Gen Intel® Core™ i7-1270P × 16):

Implementations Notes Results on intel (Millons message/s)
Spsc_queue Original Original spsc implementation with no optimzation 14 M/s
Spsc_queue Original + padding Same than previous with padding to avoid false sharing 14 M/s
Spsc_queue Saturn spsc queue - padding Differences from original : add caches tail and head (but no padding). Differences from Saturn.Single_cons_single_prod_queue : no padding 18 M/s
Spsc_queue Saturn Saturn.Single_cons_single_prod_queue 60 M/s
Spsc_queue Saturn + padding Saturn.Single_cons_single_prod_queue with full padding using Multicore_magic.copy_as_padded 63 M/s
Spsc_queue Saturn + padding + relaxed read Same than previous with relaxed atomic read. 56 M/s
Spsc_queue Saturn + padding + relaxed read + no indirection Same than previous with no option in array (using Obj.magic) to avoid Option indirection. 93 M/s
Spsc_queue Saturn_unsafe Saturn.Single_cons_single_prod_queue_unsafe. Differences from previous line : avoid float array optimisation. 102 M/s