Ongoing Optimization: Acquire-Release Semantic with CppMem

With the acquire-releae semantic, we break the sequential consistency. In the acquire-release semantic the synchronization takes place between atomic operations on the same atomic and not between threads.

Acquire-release semantic

The acquire-release semantic is more lightweight and therefore faster than the sequential consistency, because the synchronization only takes place between atomic operations. But although the intellectual challenge increases.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

// ongoingOptimizationAcquireRelease.cpp

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

std::atomic<int> x{0};
std::atomic<int> y{0};

void writing(){  
  x.store(2000,std::memory_order_relaxed);  
  y.store(11,std::memory_order_release);
}

void reading(){  
  std::cout << y.load(std::memory_order_acquire) << " ";  
  std::cout << x.load(std::memory_order_relaxed) << std::endl;
}

int main(){
  std::thread thread1(writing);
  std::thread thread2(reading);
  thread1.join();
  thread2.join();
};

On the first glance you will notice, that all operations are atomic. So the program is well defined. But the second glance shows more. The atomic operations on y are attached with the flag std::memory_order_release (line 12) and std::memory_order_acquire (line 16). In opposite to that, the atomic operations on x are annotated with std::memory_order_relaxed. So there is no synchronization and ordering constraints for x. The key for the possible values for x and y can only be answered by y.

It holds:

1. y.store(11,std::memory_order_release) synchronizes-with y.load(std::memory_order_acquire)
2. x.store(2000,std::memory_order_relaxed is visible before y.store(11,std::memory_order_release)
3. y.load(std::memory_order_acquire) is visible before x.load(std::memory_order_relaxed)

I will elaborate a little bit more on these three statements. The key idea is, that the store of y in line 10 synchronizes with the load of y in line 16. The reason is, that the operations takes place on the same atomic and they follow the acquire-release semantic. So y uses std::memory_order_release in line 12 and std::memory_order_acquire in line 16. But the pairwise operations on y have another very interesting property. They establish a kind of barrier relative to y. So x.store(2000,std::memory_order_relaxed) can not be executed after y.store(std::memory_order_release), so x.load() can not be executed before y.load().  

The reasoning was in the case of the acquire-release semantic a more sophisticated than in the case of the sequential consistency. But the possible values for x and y are the same. Only the combination y == 11 and x == 0 is no possible.

There are three different interleavings of the threads possible, which produces in the three different combinations of the values x and y.

1. thread1 will be executed before thread2.
2. thread2 will be executed before thread1.
3. thread1 executes x.store(2000), before thread2 will be exectued.

At the end the table.

CppMem

At first, the program once more with CppMem.

int main(){
  atomic_int x= 0; 
  atomic_int y= 0;
  {{{ { 
      x.store(2000,memory_order_relaxed);
      y.store(11,memory_order_release);
      }
  ||| {
      y.load(memory_order_acquire);
      x.load(memory_order_relaxed);
      } 
  }}}
}

We already know, all results except of (y=11, x=0) are possible.

Possible executions

Have a look at the three graphs, with the consistent execution. The graphs show, that there is an acquire-release semantic between the store-release of y and the load-acquire from y. It makes no difference, if the reading of y (rf) takes places in the main thread or in a separate thread. The graphs show the synchronizes-with relation with a sw arrow.

Execution for (y=0, x= 0)

Execution for (y= 0, x= 2000)

Execution for (y=11, x= 2000)

What's next?

But we can do better. Why should x be an atomic? There is no reason. That was my first, but incorrect assumption. Why? You will read in the next post.