How to: Use Exception Handling to Break from a Parallel Loop
This topic shows how to write a search algorithm for a basic tree structure.
The topic Cancellation explains the role of cancellation in the Parallel Patterns Library. The use of exception handling is a less efficient way to cancel parallel work than the use of the concurrency::task_group::cancel and concurrency::structured_task_group::cancel methods. However, one scenario where the use of exception handling to cancel work is appropriate is when you call into a third-party library that uses tasks or parallel algorithms but does not provide a task_group or structured_task_group object to cancel.
Example
The following example shows a basic tree type that contains a data element and a list of child nodes. The following section shows the body of the for_all method, which recursively performs a work function on each child node.
// A simple tree structure that has multiple child nodes.
template <typename T>
class tree
{
public:
explicit tree(T data)
: _data(data)
{
}
// Retrieves the data element for the node.
T get_data() const
{
return _data;
}
// Adds a child node to the tree.
void add_child(tree& child)
{
_children.push_back(child);
}
// Performs the given work function on the data element of the tree and
// on each child.
template<class Function>
void for_all(Function& action);
private:
// The data for this node.
T _data;
// The child nodes.
list<tree> _children;
};
Example
The following example shows the for_all method. It uses the concurrency::parallel_for_each algorithm to perform a work function on each node of the tree in parallel.
// Performs the given work function on the data element of the tree and
// on each child.
template<class Function>
void for_all(Function& action)
{
// Perform the action on each child.
parallel_for_each(begin(_children), end(_children), [&](tree& child) {
child.for_all(action);
});
// Perform the action on this node.
action(*this);
}
Example
The following example shows the search_for_value function, which searches for a value in the provided tree object. This function passes to the for_all method a work function that throws when it finds a tree node that contains the provided value.
Assume that the tree class is provided by a third-party library, and that you cannot modify it. In this case, the use of exception handling is appropriate because the for_all method does not provide a task_group or structured_task_group object to the caller. Therefore, the work function is unable to directly cancel its parent task group.
When the work function that you provide to a task group throws an exception, the runtime stops all tasks that are in the task group (including any child task groups) and discards any tasks that have not yet started. The search_for_value function uses a try-catch block to capture the exception and print the result to the console.
// Searches for a value in the provided tree object.
template <typename T>
void search_for_value(tree<T>& t, int value)
{
try
{
// Call the for_all method to search for a value. The work function
// throws an exception when it finds the value.
t.for_all([value](const tree<T>& node) {
if (node.get_data() == value)
{
throw &node;
}
});
}
catch (const tree<T>* node)
{
// A matching node was found. Print a message to the console.
wstringstream ss;
ss << L"Found a node with value " << value << L'.' << endl;
wcout << ss.str();
return;
}
// A matching node was not found. Print a message to the console.
wstringstream ss;
ss << L"Did not find node with value " << value << L'.' << endl;
wcout << ss.str();
}
Example
The following example creates a tree object and searches it for several values in parallel. The build_tree function is shown later in this topic.
int wmain()
{
// Build a tree that is four levels deep with the initial level
// having three children. The value of each node is a random number.
mt19937 gen(38);
tree<int> t = build_tree<int>(4, 3, [&gen]{ return gen()%100000; });
// Search for a few values in the tree in parallel.
parallel_invoke(
[&t] { search_for_value(t, 86131); },
[&t] { search_for_value(t, 17522); },
[&t] { search_for_value(t, 32614); }
);
}
This example uses the concurrency::parallel_invoke algorithm to search for values in parallel. For more information about this algorithm, see Parallel Algorithms.
Example
The following complete example uses exception handling to search for values in a basic tree structure.
// task-tree-search.cpp
// compile with: /EHsc
#include <ppl.h>
#include <list>
#include <iostream>
#include <algorithm>
#include <sstream>
#include <random>
using namespace concurrency;
using namespace std;
// A simple tree structure that has multiple child nodes.
template <typename T>
class tree
{
public:
explicit tree(T data)
: _data(data)
{
}
// Retrieves the data element for the node.
T get_data() const
{
return _data;
}
// Adds a child node to the tree.
void add_child(tree& child)
{
_children.push_back(child);
}
// Performs the given work function on the data element of the tree and
// on each child.
template<class Function>
void for_all(Function& action)
{
// Perform the action on each child.
parallel_for_each(begin(_children), end(_children), [&](tree& child) {
child.for_all(action);
});
// Perform the action on this node.
action(*this);
}
private:
// The data for this node.
T _data;
// The child nodes.
list<tree> _children;
};
// Builds a tree with the given depth.
// Each node of the tree is initialized with the provided generator function.
// Each level of the tree has one more child than the previous level.
template <typename T, class Generator>
tree<T> build_tree(int depth, int child_count, Generator& g)
{
// Create the tree node.
tree<T> t(g());
// Add children.
if (depth > 0)
{
for(int i = 0; i < child_count; ++i)
{
t.add_child(build_tree<T>(depth - 1, child_count + 1, g));
}
}
return t;
}
// Searches for a value in the provided tree object.
template <typename T>
void search_for_value(tree<T>& t, int value)
{
try
{
// Call the for_all method to search for a value. The work function
// throws an exception when it finds the value.
t.for_all([value](const tree<T>& node) {
if (node.get_data() == value)
{
throw &node;
}
});
}
catch (const tree<T>* node)
{
// A matching node was found. Print a message to the console.
wstringstream ss;
ss << L"Found a node with value " << value << L'.' << endl;
wcout << ss.str();
return;
}
// A matching node was not found. Print a message to the console.
wstringstream ss;
ss << L"Did not find node with value " << value << L'.' << endl;
wcout << ss.str();
}
int wmain()
{
// Build a tree that is four levels deep with the initial level
// having three children. The value of each node is a random number.
mt19937 gen(38);
tree<int> t = build_tree<int>(4, 3, [&gen]{ return gen()%100000; });
// Search for a few values in the tree in parallel.
parallel_invoke(
[&t] { search_for_value(t, 86131); },
[&t] { search_for_value(t, 17522); },
[&t] { search_for_value(t, 32614); }
);
}
This example produces the following sample output.
Found a node with value 32614.
Found a node with value 86131.
Did not find node with value 17522.
Compiling the Code
Copy the example code and paste it in a Visual Studio project, or paste it in a file that is named task-tree-search.cpp and then run the following command in a Visual Studio Command Prompt window.
cl.exe /EHsc task-tree-search.cpp
See Also
Cancellation
Exception Handling
Task Parallelism
Parallel Algorithms
task_group Class
structured_task_group Class
parallel_for_each Function