生产者-消费者的速度是否应该随着线程的增加而增加？

Question

I'm trying to implement a producer-consumer problem program in C++. One of the threads fills a queue with vectors of different numbers, while the other threads take the vectors out of the queue (synchronized using conditional variables/unique locks) and perform a simple for loop over the vector, doing some operations over the numbers. The problem is, the program's speed doesn't seem to be increasing if I use more than 2 threads. Here are some things I have found out working on it:

• The producer thread is faster than the consumers, meaning the vector will be filled way faster than the consumers are able to process the data
• Processing a single vector from the queue takes a very short time, meaning the consumers are constantly asking for data from the queue (I see this as a possible bottleneck due to the synchronization, but am not sure)

In such a program, is it expected that more threads would make the program faster, or is the constant speed independent of thread count? Thanks for any answers or explanations!

Answer 1

Welcome to Amdahl's law: https://en.wikipedia.org/wiki/Amdahl%27s_law

If your synchronization takes significant time compared to the computation, you cannot expect much speedup because the critical section is effectively single-threaded. Also memory allocation / deallocation is effectively single-threaded in your scenario because your producer allocates from its own memory arena and the consumers need to deallocate the vectors into the same arena.

A good way around this is to increase the size of the work items. Don't take single vectors but multiple ones. The exact size will need some benchmarking. A good starting point would be to take ca. L2 cache size, meaning vectors with an accumulative size of something around 64-256 kiB.