Cpp-Taskflow  2.3.0
Release 2.2.0 (2019/06/15)

Cpp-Taskflow 2.2.0 is the 3rd release in the 2.x line! This release includes several new changes such as tf::ExecutorObserverInterface, tf::Executor, isolation of taskflow graph and executor, benchmarks, and so forth. In particular, this release improve the performance of the work stealing scheduler.


Cpp-Taskflow 2.2.0 can be downloaded from here.

New Features

  • A new executor class to isolate the execution module from a taskflow
  • A new observer interface to inspect the activities of an executor
  • A decomposable taskflow construction interface
  • A new work-stealing algorithm to improve the performance

Breaks and Deprecated Features

In this release, we isolated the executor interface from tf::Taskflow, and merge tf::Framework with tf::Taskflow. This change largely improved the modularity and composability of Cpp-Taskflow in creating clean task dependency graphs and execution flows. Performance is also better. While this introduced some breaks in tf::Taskflow, we have managed to make it as less painful as possible for users to adapt to the new change.

Previously, tf::Taskflow is a hero class that manages both a task dependency graph and the execution of all graphs including frameworks. For example:

// before v2.2.0, tf::Taskflow manages both graph and execution
tf::Taskflow taskflow(4); // create a taskflow object with 4 threads
taskflow.emplace([] () { std::cout << "task A\n"; });
taskflow.wait_for_all(); // dispatch the present graph
tf::Framework framework; // create a framework object
framework.emplace([] () { std::cout << "task B\n"; });
taskflow.run(framework); // run the framework once
taskflow.wait_for_all(); // wait until the framework finishes

However, this design is awkward in many aspects. For instance, calling wait_for_all dispatches the present graph and the graph vanishes when the execution completes. To reuse a graph, users have to create another special graph called framework and mix its execution with the one in a taskflow object. Given the user feedback and lessons we have learned so far, we decided to isolate the executor interface out of tf::Taskflow and merge tf::Framework with tf::Taskflow. All execution methods such as dispatch and wait_for_all have been moved from tf::Taskflow to tf::Executor.

// starting from v2.2.0, tf::Executor manages the execution of graphs
tf::Taskflow taskflow; // create a taskflow to build dependent tasks
tf::Task A = taskflow.emplace([] () { std::cout << "task A\n"; });
tf::Task B = taskflow.emplace([] () { std::cout << "task B\n"; });
tf::Executor executor(4); // create an executor of 4 threads
executor.run(taskflow); // run the taskflow once
executor.run(taskflow, 2); // run the taskflow twice
executor.wait_for_all(); // wait for the three runs to finish

The new design has a clean separation between a task dependency graph builder tf::Taskflow and the execution of graphs tf::Executor. Users are fully responsible for the lifetime of a taskflow, and need to ensure a taskflow is alive during its execution. Besides, all task constructs remain unchanged in tf::Taskflow. In most situations, you will just need to add an executor to your program to run your taskflow graphs.

Again, we apologize this breaking change! I hope you can understand what we did is to make Cpp-Taskflow provide good performance scaling and user experience.