C++ Core Guidelines: More Traps in the Concurrency

Concurrency provides many ways to shoot yourself in the foot. The rules for today help you to know these dangers and to overcome them.

First, here are three rules for this post.

• CP.31: Pass small amounts of data between threads by value, rather than by reference or pointer
• CP.32: To share ownership between unrelated threads use shared_ptr
• CP.41: Minimize thread creation and destruction

They are more rules which I ignore because they have no content.

CP.31: Pass small amounts of data between threads by value, rather than by reference or pointer

This rule is quite apparent; therefore, I can make it short. Passing data to a thread by value gives you immediately two benefits:

1. There is no sharing and, therefore, no data race is possible. The requirements for a data race are mutable, shared state. Read the details here: C++ Core Guidelines: Rules for Concurrency and Parallelism.
2. You have not to care about the lifetime of the data. The data stays alive for the lifetime of the created thread. This is in particular important when you detach a thread: C++ Core Guidelines: Taking Care of your Child.

Of course, the crucial question is: What does a small amount of data mean? The C++ core guidelines is not clear about this point. In rule F.16 For “in” parameters, pass cheaply-copied types by value and others by reference to const to functions, the C++ core guidelines states that 4 * sizeof(int) is a rule of thumb for functions. Meaning, smaller than 4 * sizeof(int) should be passed by value; bigger than 4 * sizeof(int) by reference or pointer.

In the end, you have to measure the performance if necessary.

CP.32: To share ownership between unrelated threads use shared_ptr

Imagine, you have an object which you want to share between unrelated threads. The key question is, who is the owner of the object and, therefore, responsible for releasing the memory? Now you can choose between a memory leak if you don't deallocate the memory or undefined behavior because you invoked delete more than once. Most of the time, the undefined behavior ends in a runtime crash.

// threadSharesOwnership.cpp

#include <iostream>
#include <thread>

using namespace std::literals::chrono_literals;

struct MyInt{
  int val{2017};
  ~MyInt(){                                 // (4)
    std::cout << "Good Bye" << std::endl;   
  }
};

void showNumber(MyInt* myInt){
    std::cout << myInt->val << std::endl;
}

void threadCreator(){
    MyInt* tmpInt= new MyInt;             // (1)
    
    std::thread t1(showNumber, tmpInt);   // (2)
    std::thread t2(showNumber, tmpInt);   // (3)
    
    t1.detach();
    t2.detach();
}

int main(){

    std::cout << std::endl;

    threadCreator();
    std::this_thread::sleep_for(1s);

    std::cout << std::endl; 

}

Bear with me. The example is intentionally so easy. I let the main thread sleep for one second to be sure that it outlives the lifetime of the child thread t1 and t2. This is, of course, no appropriate synchronization, but it helps me to make my point. The vital issue of the program is: Who is responsible for the deletion of tmpInt (1)? Thread t1 (2), thread t2 (3), or the function (main thread) itself. Because I can not forecast how long each thread runs, I decided to go with a memory leak. Consequentially, the destructor of MyInt (4) is never called:

The lifetime issues are quite easy to handle if I use a std::shared_ptr.

// threadSharesOwnershipSharedPtr.cpp

#include <iostream>
#include <memory>
#include <thread>

using namespace std::literals::chrono_literals;

struct MyInt{
  int val{2017};
  ~MyInt(){
    std::cout << "Good Bye" << std::endl;
  }
};

void showNumber(std::shared_ptr<MyInt> myInt){    // (2)
    std::cout << myInt->val << std::endl;
}

void threadCreator(){
    auto sharedPtr = std::make_shared<MyInt>();    // (1)
    
    std::thread t1(showNumber, sharedPtr);
    std::thread t2(showNumber, sharedPtr);
    
    t1.detach();
    t2.detach();
}

int main(){
    
    std::cout << std::endl;
    
    threadCreator();
    std::this_thread::sleep_for(1s);
    
    std::cout << std::endl;
    
}

Two small changes to the source code were necessary. First, the pointer in (1) became a std::shared_ptr and second, the function showNumber takes a smart pointer instead of a plain pointer.

CP.41: Minimize thread creation and destruction

How expensive is a thread? Quite expensive! This is the issue behind this rule. Let me first talk about the usual size of a thread, and then about the costs of its creation.

Size

A std::thread is a thin wrapper around the native thread. This means I'm interested in the size of a Windows thread and a POSIX thread.

• Windows systems: the post Thread Stack Size gave me the answer: 1 MB.
• POSIX systems: the pthread_create man-page provides me with the answer: 2MB. This is the sizes for the i386 and x86_64 architectures. If you want to know the sizes for further architectures that support POSIX, here are they:

Creation

I didn't find numbers who much time it takes to create a thread. To get a gut feeling, I made a simple performance test on Linux and Windows.

I used GCC 6.2.1 on a desktop and cl.exe on a laptop for my performance tests. The cl.exe is part of the Microsoft Visual Studio 2017. I compiled the programs with maximum optimization. This means on Linux the flag O3 and on Windows Ox.

Here is my small test program.

// threadCreationPerformance.cpp

#include <chrono>
#include <iostream>
#include <thread>

static const long long numThreads= 1000000;

int main(){

    auto start = std::chrono::system_clock::now();

    for (volatile int i = 0; i < numThreads; ++i) std::thread([]{}).detach();  // (1)

    std::chrono::duration<double> dur= std::chrono::system_clock::now() - start;
    std::cout << "time: " << dur.count() << " seconds" << std::endl;

}

The program creates 1 million threads that execute an empty lambda function (1). These are the numbers for Linux and Windows:

Linux:

This means that the creation of a thread took about 14.5 sec / 1000000 = 14.5 microseconds on Linux.

Windows:

It took about 44 sec / 1000000 = 44 microseconds on Windows.

To put it the other way around. You can create about 69 thousand threads on Linux and 23 thousand threads on Windows in one second.

What's next?

What is the easiest way to shoot yourself in the foot? Use a condition variable! You don't believe it? Wait for the next post!