Choosing Thread Object

• We now have three different ways to execute a task
  • Create an std::thread object
  • Create an std::packaged_task object
  • Call std::async()

Advantages of std::async()

• The simplest way to execute a task

  • Easy to obtain the return value from a task
  • Or to catch any exception thrown in the task
  • Choice of running the task synchronously or asynchronously

• Higher level abstraction than std::thread

  • The library manages the threads for the programmer
  • And the inter-thread communication
  • No need to use shared data

Disadvantages of async()

• Cannot detach tasks
• A task executed with std::launch::async is "implicitly joined"

{
  auto fut = std::async(std::launch::async, hello);
} // Calls ~fut()

• The returned future's destructor will block
  • Until the task completes

Advantages of std::packaged_task

• The best choice if we want to represent tasks as objects

  • e.g. to create a container of tasks

• A lower level abstraction than std::async()

  • Can control when a task is executed
  • Can control on which thread it is executed

Advantages of std::thread

• The most flexible
  • Allows access to the underlying software thread
  • Useful for features not supported by standard C++
  • Can be detached

Recommendations

• For starting a new thread in general

  • Use std::async()

• For containers of thread objects

  • Use std::packaged_task

• For starting a detachable thread

• For more specialized requirements

  • Use std::thread

C++ Asynchronous Programming Limitations

• Lacks a number of important features

  • Continuations - "do this task, then do that task"
  • Only supports waiting on one future at a time
  • Waiting on multiple threads has to be done sequentially
  • Concurrent queue

• These were planned for C++20, but were not included

  • Were to be implemented in C++23 using "executors"
  • Perhaps in C++26?

Third Party Libraries

• Microsoft Parallel Patterns library
  • Windows only
• Apple Grand Central Dispatch
  • Open Source, runs on Linux and Android
• Intel oneAPI Thread Building Blocks
  • Open Source, runs on many platforms
• HPX
  • Open Source, runs on many platforms