ysbaddaden

Use a rwlock to have MP (multiple producers) access to a Ring?

The use-case is multiple threads trying to submit operations to an io_uring SQ ring, using a rwlock, a local sq_tail to be synchronized with the tail shared with the kernel (sq_ktail).

Maybe we can submit to the SQ ring from multiple threads with reduced contentions using a rwlock instead of a mutex. The read lock would protect incrementing the local sq_tail (updated with a CAS) and preparing the SQE, while the write lock would protect updates to the sq_ktail shared with the kernel (updated with a

ysbaddaden

Prerequisite

Get the current master branch of Crystal. Compile a new crystal compiler:

$ make .build/crystal

Setup

ysbaddaden

require "crystal/spin_lock"
require "crystal/pointer_linked_list"

# Synchronize execution of concurrently running fibers.
#
# This can be used to replace polling with a waiting list that can be resumed
# when resources are available, while still behaving inside a mutually exclusive
# context: when a waiting fiber is resumed, the mutex will be relocked.
#
# Can also be used as a notification system without a mutex.

ysbaddaden

diff --git a/src/compiler/crystal/codegen/ast.cr b/src/compiler/crystal/codegen/ast.cr
index ab4fa76d4..ed3465be1 100644
--- a/src/compiler/crystal/codegen/ast.cr
+++ b/src/compiler/crystal/codegen/ast.cr
@@ -72,6 +72,10 @@ module Crystal
       nil
     end
 
+    def linkage
+      nil

ysbaddaden

# Crystal port of the mutex from the "nsync" library.
#
# Copyright 2016 Google Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#

ysbaddaden

HEAP benchmarks

• CPU: Intel i7 14700K
• RAM: 5600 MT/s
• Ubuntu 22.04 / Linux 6.9

Each test creates N items (not measured) and shuffles them (unless said otherwise) then measures how long the operation(s) take to complete. The test repeats itself 10 times. The average time per operation of all runs is eventually printed.

ysbaddaden

diff --git a/src/compiler/crystal/compiler.cr b/src/compiler/crystal/compiler.cr
index 38880ee9e..966e2f954 100644
--- a/src/compiler/crystal/compiler.cr
+++ b/src/compiler/crystal/compiler.cr
@@ -523,13 +523,17 @@ module Crystal
     private def fork_codegen(units, n_threads)
       workers = fork_workers(n_threads) do |input, output|
         while i = input.gets(chomp: true).presence
+          Crystal::System.print_error "CODEGEN pid=#{LibC.getpid} STDIN => #{i}\n"
           unit = units[i.to_i]

ysbaddaden

class STW
  def initialize
    @running = true
    @started = Atomic(Int32).new(0)
    @count = Atomic(Int32).new(0)
    @threads = [] of Thread
  end

  def start_thread
    @threads << Thread.new do

ysbaddaden

require "json"

def cleanup_constant(type : JSON::Any)
  type.as_h.select! do |key, _|
    {
      "id",
      "name",
      "value",
      "doc",
      "summary",

ysbaddaden

require "c/sys/time"

lib LibC
  KQUEUE_CLOEXEC = O_CLOEXEC

  # todo: constants
  # EV_ADD = ...
  # EV_ENABLE = ...
  # ...