My Conclusion: Summation of a Vector in three Variants

After I've calculated in three different ways the sum of a std::vector I want to draw my conclusions.

The three strategies

At first, all numbers in an overview. First, the single threaded variant; second, the multiple threads with a shared summation variable; last, the multiple threads with minimal synchronization. I have to admit that I was astonished be the last variant.

Single threaded (1)

Multiple threads with a shared summation variable (2)

Multiple threads with minimal synchronization (3)

My observations

For simplicity reasons I will only reason about Linux. Thanks to Andreas Schäfer (https://plus.google.com/u/0/+AndreasSch%C3%A4fer_gentryx) who gave me deeper insight. 

Single threaded

The range-based for-loop and the STL algorithm std::accumulate are in the same league. This observation holds for the maximal optimized and non-optimized program. It's very interesting that the maximal optimized versions is about 30 times faster than the non-optimized version. The compiler uses for the summation in case of the optimized version vectorized instruction (SSE or AVX). Therefore, the loop counter will be increased by 2 (SSE) or 4 (AVC).

Multiple threads with a shared summation variable

The synchronization on each access to the shared variable (2) shows on point: Synchronization is expensive. Although I break the sequential consistency with the relaxed semantic the program is about 40 times slower than the pendants (1) or (3). Not only out of performance reasons it must be our goal to minimize the synchronization of the shared variable.

Multiple threads with minimal synchronization

The summation with minimal synchronized threads (4 atomic operations or locks) (3) is hardly faster as the range-based for-loop or std::accumulate (1). That holds although in the multithreading variant where four threads can work independently on four cores. That surprised me because I was expecting a nearly fourfold improvement. But what surprised me even more, was that my four cores were not fully utilized.

The reason is simple. The cores can't get the data fast enough from the memory. Or to say it the other way around. The memory slows down the cores.

My conclusion

My conclusion from the performance measurements is to use for such a simple operation std::accumulate. That's for two reasons. First, the performance boost of variant (3) doesn't justify the expense; second, C++ will have in C++17 a parallel version of std::accumulate. Therefore, it's very easy to switch from the sequential to the parallel version. 

What's next?

The time library does not belong to the multithreading library but it's an important component of the multithreading capabilities of C++. For example you have to wait for an absolute time for a lock or put your thread for a relative time to sleep. So in the next post I write about time.

.