다음을 통해 공유


Walkthrough: Creating a Dataflow Agent

This document demonstrates how to create agent-based applications that are based on dataflow, instead of control flow.

Control flow refers to the execution order of operations in a program. Control flow is regulated by using control structures such as conditional statements, loops, and so on. Alternatively, dataflow refers to a programming model in which computations are made only when all required data is available. The dataflow programming model is related to the concept of message passing, in which independent components of a program communicate with one another by sending messages.

Asynchronous agents support both the control-flow and dataflow programming models. Although the control-flow model is appropriate in many cases, the dataflow model is appropriate in others, for example, when an agent receives data and performs an action that is based on the payload of that data.

Prerequisites

Read the following documents before you start this walkthrough:

Sections

This walkthrough contains the following sections:

  • Creating a Basic Control-Flow Agent

  • Creating a Basic Dataflow Agent

  • Creating a Message-Logging Agent

Creating a Basic Control-Flow Agent

Consider the following example that defines the control_flow_agent class. The control_flow_agent class acts on three message buffers: one input buffer and two output buffers. The run method reads from the source message buffer in a loop and uses a conditional statement to direct the flow of program execution. The agent increments one counter for non-zero, negative values and increments another counter for non-zero, positive values. After the agent receives the sentinel value of zero, it sends the values of the counters to the output message buffers. The negatives and positives methods enable the application to read the counts of negative and positive values from the agent.

// A basic agent that uses control-flow to regulate the order of program  
// execution. This agent reads numbers from a message buffer and counts the  
// number of positive and negative values. 
class control_flow_agent : public agent
{
public:
   explicit control_flow_agent(ISource<int>& source)
      : _source(source)
   {
   }

   // Retrieves the count of negative numbers that the agent received.
   size_t negatives() 
   {
      return receive(_negatives);
   }

   // Retrieves the count of positive numbers that the agent received.
   size_t positives()
   {
      return receive(_positives);
   }

protected:
   void run()
   {
      // Counts the number of negative and positive values that 
      // the agent receives.
      size_t negative_count = 0;
      size_t positive_count = 0;

      // Read from the source buffer until we receive 
      // the sentinel value of 0. 
      int value = 0;      
      while ((value = receive(_source)) != 0)
      {
         // Send negative values to the first target and 
         // non-negative values to the second target. 
         if (value < 0)
            ++negative_count;
         else
            ++positive_count;
      }

      // Write the counts to the message buffers.
      send(_negatives, negative_count);
      send(_positives, positive_count);

      // Set the agent to the completed state.
      done();
   }
private:
   // Source message buffer to read from.
   ISource<int>& _source;

   // Holds the number of negative and positive numbers that the agent receives.
   single_assignment<size_t> _negatives;
   single_assignment<size_t> _positives;
};

Although this example makes basic use of control flow in an agent, it demonstrates the serial nature of control-flow-based programming. Each message must be processed sequentially, even though multiple messages might be available in the input message buffer. The dataflow model enables both branches of the conditional statement to evaluate concurrently. The dataflow model also enables you to create more complex messaging networks that act on data as it becomes available.

[Top]

Creating a Basic Dataflow Agent

This section shows how to convert the control_flow_agent class to use the dataflow model to perform the same task.

The dataflow agent works by creating a network of message buffers, each of which serves a specific purpose. Certain message blocks use a filter function to accept or reject a message on the basis of its payload. A filter function ensures that a message block receives only certain values.

To convert the control-flow agent to a dataflow agent

  1. Copy the body of the control_flow_agent class to another class, for example, dataflow_agent. Alternatively, you can rename the control_flow_agent class.

  2. Remove the body of the loop that calls receive from the run method.

    void run()
    {
       // Counts the number of negative and positive values that 
       // the agent receives.
       size_t negative_count = 0;
       size_t positive_count = 0;
    
    
       // Write the counts to the message buffers.
       send(_negatives, negative_count);
       send(_positives, positive_count);
    
       // Set the agent to the completed state.
       done();
    }
    
  3. In the run method, after the initialization of the variables negative_count and positive_count, add a countdown_event object that tracks the count of active operations.

    // Tracks the count of active operations.
    countdown_event active;
    // An event that is set by the sentinel. 
    event received_sentinel;
    

    The countdown_event class is shown later in this topic.

  4. Create the message buffer objects that will participate in the dataflow network.

    // 
    // Create the members of the dataflow network. 
    // 
    
    // Increments the active counter.
    transformer<int, int> increment_active(
       [&active](int value) -> int {
          active.add_count();
          return value;
       });
    
    // Increments the count of negative values.
    call<int> negatives(
       [&](int value) {
          ++negative_count;
          // Decrement the active counter.
          active.signal();
       },
       [](int value) -> bool {
          return value < 0;
       });
    
    // Increments the count of positive values.
    call<int> positives(
       [&](int value) {
          ++positive_count;
          // Decrement the active counter.
          active.signal();
       },
       [](int value) -> bool {
          return value > 0;
       });
    
    // Receives only the sentinel value of 0.
    call<int> sentinel(
       [&](int value) {            
          // Decrement the active counter.
          active.signal();
          // Set the sentinel event.
          received_sentinel.set();
       },
       [](int value) { 
          return value == 0; 
       });
    
    // Connects the _source message buffer to the rest of the network.
    unbounded_buffer<int> connector;
    
  5. Connect the message buffers to form a network.

    // 
    // Connect the network. 
    // 
    
    // Connect the internal nodes of the network.
    connector.link_target(&negatives);
    connector.link_target(&positives);
    connector.link_target(&sentinel);
    increment_active.link_target(&connector);
    
    // Connect the _source buffer to the internal network to  
    // begin data flow.
    _source.link_target(&increment_active);
    
  6. Wait for the event and countdown event objects to be set. These events signal that that the agent has received the sentinel value and that all operations have finished.

    // Wait for the sentinel event and for all operations to finish.
    received_sentinel.wait();
    active.wait();
    

The following diagram shows the complete dataflow network for the dataflow_agent class:

The dataflow network

The following table describes the members of the network.

Member

Description

increment_active

A concurrency::transformer object that increments the active event counter and passes the input value to the rest of the network.

negatives, positives

concurrency::call objects that increment the count of numbers and decrements the active event counter. The objects each use a filter to accept either negative numbers or positive numbers.

sentinel

A concurrency::call object that accepts only the sentinel value of zero and decrements the active event counter.

connector

A concurrency::unbounded_buffer object that connects the source message buffer to the internal network.

Because the run method is called on a separate thread, other threads can send messages to the network before the network is fully connected. The _source data member is an unbounded_buffer object that buffers all input that is sent from the application to the agent. To make sure that the network processes all input messages, the agent first links the internal nodes of the network and then links the start of that network, connector, to the _source data member. This guarantees that messages do not get processed as the network is being formed.

Because the network in this example is based on dataflow, rather than on control-flow, the network must communicate to the agent that it has finished processing each input value and that the sentinel node has received its value. This example uses a countdown_event object to signal that all input values have been processed and a concurrency::event object to indicate that the sentinel node has received its value. The countdown_event class uses an event object to signal when a counter value reaches zero. The head of the dataflow network increments the counter every time that it receives a value. Every terminal node of the network decrements the counter after it processes the input value. After the agent forms the dataflow network, it waits for the sentinel node to set the event object and for the countdown_event object to signal that its counter has reached zero.

The following example shows the control_flow_agent, dataflow_agent, and countdown_event classes. The wmain function creates a control_flow_agent and a dataflow_agent object and uses the send_values function to send a series of random values to the agents.

// dataflow-agent.cpp 
// compile with: /EHsc 
#include <windows.h>
#include <agents.h>
#include <iostream>
#include <random>

using namespace concurrency;
using namespace std;

// A basic agent that uses control-flow to regulate the order of program  
// execution. This agent reads numbers from a message buffer and counts the  
// number of positive and negative values. 
class control_flow_agent : public agent
{
public:
   explicit control_flow_agent(ISource<int>& source)
      : _source(source)
   {
   }

   // Retrieves the count of negative numbers that the agent received.
   size_t negatives() 
   {
      return receive(_negatives);
   }

   // Retrieves the count of positive numbers that the agent received.
   size_t positives()
   {
      return receive(_positives);
   }

protected:
   void run()
   {
      // Counts the number of negative and positive values that 
      // the agent receives.
      size_t negative_count = 0;
      size_t positive_count = 0;

      // Read from the source buffer until we receive 
      // the sentinel value of 0. 
      int value = 0;      
      while ((value = receive(_source)) != 0)
      {
         // Send negative values to the first target and 
         // non-negative values to the second target. 
         if (value < 0)
            ++negative_count;
         else
            ++positive_count;
      }

      // Write the counts to the message buffers.
      send(_negatives, negative_count);
      send(_positives, positive_count);

      // Set the agent to the completed state.
      done();
   }
private:
   // Source message buffer to read from.
   ISource<int>& _source;

   // Holds the number of negative and positive numbers that the agent receives.
   single_assignment<size_t> _negatives;
   single_assignment<size_t> _positives;
};

// A synchronization primitive that is signaled when its  
// count reaches zero. 
class countdown_event
{
public:
   countdown_event(unsigned int count = 0L)
      : _current(static_cast<long>(count)) 
   {
      // Set the event if the initial count is zero. 
      if (_current == 0L)
         _event.set();
   }

   // Decrements the event counter. 
   void signal() {
      if(InterlockedDecrement(&_current) == 0L) {
         _event.set();
      }
   }

   // Increments the event counter. 
   void add_count() {
      if(InterlockedIncrement(&_current) == 1L) {
         _event.reset();
      }
   }

   // Blocks the current context until the event is set. 
   void wait() {
      _event.wait();
   }

private:
   // The current count. 
   volatile long _current;
   // The event that is set when the counter reaches zero. 
   event _event;

   // Disable copy constructor.
   countdown_event(const countdown_event&);
   // Disable assignment.
   countdown_event const & operator=(countdown_event const&);
};

// A basic agent that resembles control_flow_agent, but uses uses dataflow to  
// perform computations when data becomes available. 
class dataflow_agent : public agent
{
public:
   dataflow_agent(ISource<int>& source)
      : _source(source)
   {
   }

   // Retrieves the count of negative numbers that the agent received.
   size_t negatives() 
   {
      return receive(_negatives);
   }

   // Retrieves the count of positive numbers that the agent received.
   size_t positives()
   {
      return receive(_positives);
   }

protected:
   void run()
   {
      // Counts the number of negative and positive values that 
      // the agent receives.
      size_t negative_count = 0;
      size_t positive_count = 0;

      // Tracks the count of active operations.
      countdown_event active;
      // An event that is set by the sentinel. 
      event received_sentinel;

      // 
      // Create the members of the dataflow network. 
      // 

      // Increments the active counter.
      transformer<int, int> increment_active(
         [&active](int value) -> int {
            active.add_count();
            return value;
         });

      // Increments the count of negative values.
      call<int> negatives(
         [&](int value) {
            ++negative_count;
            // Decrement the active counter.
            active.signal();
         },
         [](int value) -> bool {
            return value < 0;
         });

      // Increments the count of positive values.
      call<int> positives(
         [&](int value) {
            ++positive_count;
            // Decrement the active counter.
            active.signal();
         },
         [](int value) -> bool {
            return value > 0;
         });

      // Receives only the sentinel value of 0.
      call<int> sentinel(
         [&](int value) {            
            // Decrement the active counter.
            active.signal();
            // Set the sentinel event.
            received_sentinel.set();
         },
         [](int value) { 
            return value == 0; 
         });

      // Connects the _source message buffer to the rest of the network.
      unbounded_buffer<int> connector;

      // 
      // Connect the network. 
      // 

      // Connect the internal nodes of the network.
      connector.link_target(&negatives);
      connector.link_target(&positives);
      connector.link_target(&sentinel);
      increment_active.link_target(&connector);

      // Connect the _source buffer to the internal network to  
      // begin data flow.
      _source.link_target(&increment_active);

      // Wait for the sentinel event and for all operations to finish.
      received_sentinel.wait();
      active.wait();

      // Write the counts to the message buffers.
      send(_negatives, negative_count);
      send(_positives, positive_count);

      // Set the agent to the completed state.
      done();
   }

private:
   // Source message buffer to read from.
   ISource<int>& _source;

   // Holds the number of negative and positive numbers that the agent receives.
   single_assignment<size_t> _negatives;
   single_assignment<size_t> _positives;
};

// Sends a number of random values to the provided message buffer. 
void send_values(ITarget<int>& source, int sentinel, size_t count)
{
   // Send a series of random numbers to the source buffer.
   mt19937 rnd(42);
   for (size_t i = 0; i < count; ++i)
   {
      // Generate a random number that is not equal to the sentinel value. 
      int n;
      while ((n = rnd()) == sentinel);

      send(source, n);      
   }
   // Send the sentinel value.
   send(source, sentinel);   
}

int wmain()
{
   // Signals to the agent that there are no more values to process. 
   const int sentinel = 0;
   // The number of samples to send to each agent. 
   const size_t count = 1000000;

   // The source buffer that the application writes numbers to and  
   // the agents read numbers from.
   unbounded_buffer<int> source;

   // 
   // Use a control-flow agent to process a series of random numbers. 
   //
   wcout << L"Control-flow agent:" << endl;

   // Create and start the agent.
   control_flow_agent cf_agent(source);
   cf_agent.start();

   // Send values to the agent.
   send_values(source, sentinel, count);

   // Wait for the agent to finish.
   agent::wait(&cf_agent);

   // Print the count of negative and positive numbers.
   wcout << L"There are " << cf_agent.negatives() 
         << L" negative numbers."<< endl;
   wcout << L"There are " << cf_agent.positives() 
         << L" positive numbers."<< endl;  

   // 
   // Perform the same task, but this time with a dataflow agent. 
   //
   wcout << L"Dataflow agent:" << endl;

   // Create and start the agent.
   dataflow_agent df_agent(source);
   df_agent.start();

   // Send values to the agent.
   send_values(source, sentinel, count);

   // Wait for the agent to finish.
   agent::wait(&df_agent);

   // Print the count of negative and positive numbers.
   wcout << L"There are " << df_agent.negatives() 
         << L" negative numbers."<< endl;
   wcout << L"There are " << df_agent.positives() 
         << L" positive numbers."<< endl;
}

This example produces the following sample output:

Control-flow agent:
There are 500523 negative numbers.
There are 499477 positive numbers.
Dataflow agent:
There are 500523 negative numbers.
There are 499477 positive numbers.

Compiling the Code

Copy the example code and paste it in a Visual Studio project, or paste it in a file that is named dataflow-agent.cpp and then run the following command in a Visual Studio Command Prompt window.

cl.exe /EHsc dataflow-agent.cpp

[Top]

Creating a Message-Logging Agent

The following example shows the log_agent class, which resembles the dataflow_agent class. The log_agent class implements an asynchronous logging agent that writes log messages to a file and to the console. The log_agent class enables the application to categorize messages as informational, warning, or error. It also enables the application to specify whether each log category is written to a file, the console, or both. This example writes all log messages to a file and only error messages to the console.

// log-filter.cpp 
// compile with: /EHsc 
#include <windows.h>
#include <agents.h>
#include <sstream>
#include <fstream>
#include <iostream>

using namespace concurrency;
using namespace std;

// A synchronization primitive that is signaled when its  
// count reaches zero. 
class countdown_event
{
public:
    countdown_event(unsigned int count = 0L)
        : _current(static_cast<long>(count)) 
    {
        // Set the event if the initial count is zero. 
        if (_current == 0L)
        {
            _event.set();
        }
    }

    // Decrements the event counter. 
    void signal()
    {
        if(InterlockedDecrement(&_current) == 0L)
        {
            _event.set();
        }
    }

    // Increments the event counter. 
    void add_count()
    {
        if(InterlockedIncrement(&_current) == 1L)
        {
            _event.reset();
        }
    }

    // Blocks the current context until the event is set. 
    void wait()
    {
        _event.wait();
    }

private:
    // The current count. 
    volatile long _current;
    // The event that is set when the counter reaches zero. 
    event _event;

    // Disable copy constructor.
    countdown_event(const countdown_event&);
    // Disable assignment.
    countdown_event const & operator=(countdown_event const&);
};

// Defines message types for the logger. 
enum log_message_type
{
    log_info    = 0x1,
    log_warning = 0x2,
    log_error   = 0x4,
};

// An asynchronous logging agent that writes log messages to  
// file and to the console. 
class log_agent : public agent
{
    // Holds a message string and its logging type. 
    struct log_message
    {
        wstring message;
        log_message_type type;
    };

public:
    log_agent(const wstring& file_path, log_message_type file_messages, log_message_type console_messages)
        : _file(file_path)
        , _file_messages(file_messages)
        , _console_messages(console_messages)
        , _active(0)
    {
        if (_file.bad())
        {
            throw invalid_argument("Unable to open log file.");
        }
    }

    // Writes the provided message to the log. 
    void log(const wstring& message, log_message_type type)
    {  
        // Increment the active message count.
        _active.add_count();

        // Send the message to the network.
        log_message msg = { message, type };
        send(_log_buffer, msg);
    }

    void close()
    {
        // Signal that the agent is now closed.
        _closed.set();
    }

protected:

    void run()
    {
        // 
        // Create the dataflow network. 
        // 

        // Writes a log message to file.
        call<log_message> writer([this](log_message msg)
        {
            if ((msg.type & _file_messages) != 0)
            {
                // Write the message to the file.
                write_to_stream(msg, _file);
            }
            if ((msg.type & _console_messages) != 0)
            {
                // Write the message to the console.
                write_to_stream(msg, wcout);
            }
            // Decrement the active counter.
            _active.signal();
        });

        // Connect _log_buffer to the internal network to begin data flow.
        _log_buffer.link_target(&writer);

        // Wait for the closed event to be signaled.
        _closed.wait();

        // Wait for all messages to be processed.
        _active.wait();

        // Close the log file and flush the console.
        _file.close();
        wcout.flush();

        // Set the agent to the completed state.
        done();
    }

private:
    // Writes a logging message to the specified output stream. 
    void write_to_stream(const log_message& msg, wostream& stream)
    {
        // Write the message to the stream.
        wstringstream ss;

        switch (msg.type)
        {
        case log_info:
            ss << L"info: ";
            break;
        case log_warning:
            ss << L"warning: ";
            break;
        case log_error:
            ss << L"error: ";
        }

        ss << msg.message << endl;
        stream << ss.str();
    }

private:   
    // The file stream to write messages to.
    wofstream _file;

    // The log message types that are written to file.
    log_message_type _file_messages;

    // The log message types that are written to the console.
    log_message_type _console_messages;

    // The head of the network. Propagates logging messages 
    // to the rest of the network.
    unbounded_buffer<log_message> _log_buffer;

    // Counts the number of active messages in the network.
    countdown_event _active;

    // Signals that the agent has been closed. 
    event _closed;
};

int wmain()
{
    // Union of all log message types.
    log_message_type log_all = log_message_type(log_info | log_warning  | log_error);

    // Create a logging agent that writes all log messages to file and error  
    // messages to the console.
    log_agent logger(L"log.txt", log_all, log_error);

    // Start the agent.
    logger.start();

    // Log a few messages.

    logger.log(L"===Logging started.===", log_info);

    logger.log(L"This is a sample warning message.", log_warning);
    logger.log(L"This is a sample error message.", log_error);

    logger.log(L"===Logging finished.===", log_info);

    // Close the logger and wait for the agent to finish.
    logger.close();
    agent::wait(&logger);
}

This example writes the following output to the console.

error: This is a sample error message.

This example also produces the log.txt file, which contains the following text.

info: ===Logging started.===
warning: This is a sample warning message.
error: This is a sample error message.
info: ===Logging finished.===

Compiling the Code

Copy the example code and paste it in a Visual Studio project, or paste it in a file that is named log-filter.cpp and then run the following command in a Visual Studio Command Prompt window.

cl.exe /EHsc log-filter.cpp

[Top]

See Also

Other Resources

Concurrency Runtime Walkthroughs