fxmozart / disruptor-cpp Goto Github PK

Performance can be GREATLY improved by not using std::vectors. After I modified 
code to use arrays for both Consumer lists and actual RingBuffer performance 
increased dramatically.

Just submitting it so that other people are aware - I've considered not using 
the library at all for performance reasons, until I've made the change.

Brief description of the program:

1. Producer thread will mark entries with the fd to be processed as they are 
received from epoll_wait()
2. Consumer thread will call read(2) on descriptor from the entry

Internal reporting numbers are great, but when combined with tcpdump to measure 
wall-to-wall latency we see that messages spend *very* long time in TCP buffer.

This is on 64-bit Linux 2.6.18-164.20.1.el5 #1 SMP

When read() is done in the same thread as epoll, there's no issue.

What steps will reproduce the problem?
1. WaitStrategy.h refers to a Util.h, but there is no Util.h in the src 
directory, there is only a util.h

What steps will reproduce the problem?

Try to compile under vc2010 and got compile error in function.

int ceilingNextPowerOfTwo(const int & x)
{
  return 1 << fls(x - 1);
}

Can you consider to replace it to :
int ceilingNextPowerOfTwo(int x)
{
    assert(x>0);
    x--;
    x |= x >> 1;
    x |= x >> 2;
    x |= x >> 4;
    x |= x >> 8;
    x |= x >> 16;

    return ++x;
}