id,summary,reporter,owner,description,type,status,milestone,component,version,severity,resolution,keywords,cc
12120,Performance improvement in thread/barrier.hpp,Ronald Holthuizen <ronald@…>,viboes,"The current Boost code in thread/barrier.hpp has this code:

---------------------
    bool wait()
    {
      boost::unique_lock < boost::mutex > lock(m_mutex);
      unsigned int gen = m_generation;

      if (--m_count == 0)
      {
        m_generation++;
        m_count = static_cast<unsigned int>(fct_());
        BOOST_ASSERT(m_count != 0);
        m_cond.notify_all();
        return true;
      }

      while (gen == m_generation)
        m_cond.wait(lock);
      return false;
    }
------------------
In the line m_cond.notify_all(); m_mutex is locked while the other threads are notified. This is a performance anti-pattern: the notified threads will first have to wait for the mutex to be released.

By adding the line:
lock.unlock();
before the m_cond.notify_all(); the performance of the barrier class is improved.

More info why this is better, see sample in:
http://en.cppreference.com/w/cpp/thread/condition_variable
",Bugs,closed,Boost 1.62.0,thread,Boost 1.61.0,Optimization,fixed,barrier wait performance,