Increase FD_SETSIZE

[TankOs]

On Windows, FD_SETSIZE (the maximum number of allowed entries in the fd_set structure) is limited to 64 which is quite low. So when doing a select() with more than 64 you run into trouble.

One could either create groups with FD_SETSIZE entries and call select() multiple times or increase FD_SETSIZE by defining the macro before Winsock2.h is included.

On Linux the default is 1024.

More info: http://msdn.microsoft.com/en-us/library/windows/desktop/ms740141(v=vs.85).aspx

[TankOs]

Do you already have an idea? I'm just thinking about that too any ask myself if

• calling select() multiple times or
• calling select() once and iterating over all possible sockets

is faster.

[LaurentGomila]

How can you call select() multiple times? The first call will block the others.

[TankOs]

Okay, should think before posting. ;) However it still applies for a zero timeout and/or multiple threads, but that only complicates things.

So probably increasing FD_SETSIZE manually is a good choice then.

[LaurentGomila]

Yep :)

[TankOs]

I investigated a little bit further: It seems that redefining FD_SETSIZE on Windows may have serious side-effects, it's really not recommended. Therefore we end up with a 64 sockets limit. There're actually two ways I encountered that solve the problem:<

• Use multiple threads for multiple select()s and use a maximum of 64 sockets per thread for the fd_set. Not possible: The highest file descriptor ID counts, not the amount.
• Use WSAPoll(), however that one has been added in Windows Vista and is therefore not available in XP, which might be a real showstopper (also it's Windows-only, on Linux you still have the 1024 limit (in detail it means the highest file descriptor ID must be less than 1024, not the amount!!))

Seems like wrapping the sockets API isn't enough. SFML Network probably needs another layer of abstraction if you want to keep it simple to use without annoying limitations.

[LaurentGomila]

So far nobody complained about the limitation, I guess that 64 sockets is enough for most applications.

Since it's potentially a complicated problem, and I have other priorities for the near future, I'll leave the implementation as it is with its limitations, and wait for more feedback with real use cases.

Thank you very much for your feedback and investigations!

[TankOs]

You're welcome, but please at least add an assert() or debug message to Selector::add() that warns the user when a file descriptor with an ID >= FD_SETSIZE is being added (+ documentation). I think debugging such a bug is not so much fun.

[LaurentGomila]

please at least add an assert() or debug message to Selector::add() that warns the user when a file descriptor with an ID >= FD_SETSIZE is being added (+ documentation)

Sure.