Whatever proves to be a good measure of "server is busy". If requests are relatively constant in resource cost, then you might have a hard ceiling of "N max simultaneous requests" and a semaphore would be a good fit (and you'd need to pre-fork upto N+1 children, of course).

But you might have a different "busy" metric you want to use like "system load average" or "idle CPU percentage" or "virtual memory in use" or whatever.

But, even if you do implement some "busy" metric other than "N simultaneous requests", you'll probably still have some "max N children" configuration and so a N-1 semaphore is probably still a good idea.

If it's worth for the client to reconnect to another box if the server is busy, I suppose the jobs take relatively long to process, otherwise the client could expect a server process to become free soonish when it enters the listen queue.

Are you sure you want preforked servers for that? Because the preforking is usually done to avoid the fork() overhead if it takes a significant fraction of the whole request processing time.

Otherwise you could have one master process that accepts all connections and can optionally have a little dialog with the client. If it can down a nonblocking semaphore, it just forks and lets the child do the rest, otherwise it tells the client to check another box (you could even have the servers inform each other of their current load so it could give a hint to the client) and hangs up.

See my other post with "option 2". If you want to share a bit of "go no/go information" between processes, consider using the file lock mechanism. This is a common way to do this. The "busy/not busy decider" manages a write lock.. more specifically an exclusive lock to a zero length file. Other processes test this file to see: "if I wanted to get a write lock, could I do it?". If the answer is "no" then close the active connection.

The file lock control table is a memory resident thing and checking this is fast but you have to open the file first, however this is usually fast compared with the network delays just to get to your box to begin with. Have the "decider" function be the only one who actually locks this "I_am_too_busy" file.

If the "decider" process crashes for some reason, its lock is released - no clean-up required. A semaphore can potentially have problems.

There are other ways to share information between processes on *nix systems. I would think about the easiest ways and only get more complicated when needed. Stay with the *nix forking server model if you can. I would also consider the post by mbethke.

I don't see any specs on how fast this has to be, nor benchmarks that show why a particular implementation is too slow. The fastest implementation would be a "super server" which is the most complicated model because it involved both select based and fork based code. I would write that in C if this level of complexity and performance is needed.

Anyway, this thread started with a fairly simple question and it appears that things are getting more and more complicated.