Hi Jody, --- Jody Hagins <jody-boost-011304@atdesk.com> wrote:
Maybe on WAN servers, but not on high performance LAN based servers. Ruunning a simple test on 100 file descriptors, here is what I get.
poll() calls per second: 107906 select() calls per second: 104891 epoll() calls per second: 70923
Granted, it's not a huge difference, and other numbers of FDs show differing results (also, only selecting for 1 FD in epoll() shows lots better results, but you have to make more system calls as well.
I have just been experimenting with modifications to the internal reactor interfaces such that they will attempt a non-blocking operation first, and only add the requested event to epoll if it doesn't complete immediately. In cases where descriptors are chronically ready, this bypasses epoll system calls altogether. Rough indication so far is that it improves performance 10-20%.
Anyway, my point is that there surely are use cases where a user would want control over which underlying implementation is used. The programmer should be allowed to determine which implementation is used under the hood, and should be able to use several different ones for several different uses.
I'm happy to provide "private" methods to allow you to choose the backend, such as a #define, or something that defers the choice to runtime. My point is that it should not be part of the public interface, since: - These implementations are inherently non portable, and portability is the goal. - Adding things to a public interface implies that they are stable and will not be taken away in the future. Suppose some new demuxing technique comes along -- whether it be a new OS-level feature or a better way of utilising existing ones -- or an existing technique becomes obsolete, should asio's interface provide a guarantee that your chosen implementation is still supported? To me, it should not. Exposing these implementation details is likely to lead to onerous backward compatibility and maintenance requirements, and restrict the ability to optimise for the far more common use cases. As I said above, I'm willing to provide a private mechanism for choosing the implementation. I just don't believe that this is something that I can publish public interfaces for. Cheers, Chris