Synchronization Primitives with std::atomic

Welcome to the next step in our exploration of concurrency in C++. In the previous lesson, we established a foundation by understanding the C++ Memory Model, focusing on concepts like visibility, atomicity, and memory consistency. Now, we are venturing into synchronization primitives, with a spotlight on std::atomic. Synchronization is at the heart of concurrent programming, ensuring that threads interact with shared data predictably and safely. This lesson will equip you with the tools to manage these interactions effectively.

In this lesson, we will dissect the synchronization capabilities offered by std::atomic:

• Understanding std::atomic: We will explore what std::atomic ensures, why it is essential for concurrency, and how it differs from regular variables.

• Lock-Free Programming: You'll learn about the benefits and limitations of lock-free programming, harnessing the power of atomic operations to improve performance in multi-threaded applications.

Before moving to the code example, let's understand what std::atomic is and why it is crucial for concurrent programming.

std::atomic is a template class in the C++ Standard Library that provides atomic operations on shared data. It ensures, that when multiple threads access the same data concurrently, the operations are performed atomically, without interference from other threads. This means, that if thread 1 is modifying a shared variable, thread 2 will not read or write to it until thread 1 has completed its atomic operation.

To illustrate this, let's revisit a piece of code that emphasizes these concepts:

#include <atomic>
#include <iostream>
#include <thread>

class SynchronizedCounter {
public:
    void increment() {
        count_.fetch_add(1, std::memory_order_relaxed);
    }

    int getCount() const {
        return count_.load(std::memory_order_relaxed);
    }

private:
    std::atomic<int> count_{0};
};

int main() {
    SynchronizedCounter counter;
    std::thread t1([&counter]() { for (int i = 0; i < 1000; ++i) counter.increment(); });
    std::thread t2([&counter]() { for (int i = 0; i < 1000; ++i) counter.increment(); });
    t1.join();
    t2.join();
    std::cout << "Final count: " << counter.getCount() << std::endl; // Expected output: 2000
    return 0;
}

Let's break down the code:

• We define a SynchronizedCounter class with two member functions: increment and getCount.
• The increment function increments the counter atomically using the fetch_add method.
  • The fetch_add method atomically increments the counter by 1 and returns the previous value. Note, that the fetch_add method is an atomic operation, ensuring that in the middle of the operation, no other thread can access the shared data.
  • The std::memory_order_relaxed parameter specifies the memory ordering constraints for the operation, ensuring that the operation is performed atomically without any specific ordering guarantees. We'll delve deeper into memory ordering later.
• The getCount function reads the counter value atomically using the load method.
  • The load method atomically reads the counter value and returns it. The std::memory_order_relaxed parameter specifies the memory ordering constraints for the operation.
• In the main function, we create two threads, t1 and t2, that increment the counter 1000 times each.

You might ask, why not simply use a regular int and the increment operation count_ += 1? The answer lies in the atomicity of the operation.

If count_ were a standard int, the increment count_ += 1 would not be atomic. Instead, it would consist of three distinct CPU instructions:

• Load: Reading the current value from memory into a CPU register.
• Increment: Adding 1 to the value within that register.
• Store: Writing the updated value back from the register to memory.

In a multi-threaded environment, these steps can be interleaved. For instance, if thread 1 loads the value as 5, and before it can write the result (6) back to memory, thread 2 also loads the value as 5. Both threads will then increment their local copies to 6 and write them back. Consequently, one of the increments is "lost," and the final count becomes 6 instead of 7.

By using std::atomic<int>, we ensure that this entire "read-modify-write" sequence is performed as a single, indivisible operation. No other thread can see the variable in an intermediate state or interfere until the operation is complete. While std::atomic does support the += operator (which is also atomic), we use the fetch_add method in our example because it allows us to explicitly specify a memory_order. The += operator always defaults to the most restrictive ordering (std::memory_order_seq_cst).

The importance of atomic operations becomes evident in scenarios where the operations are more complex, involving multiple steps. By using atomic operations, we can ensure that these operations are performed atomically, without interference from other threads.