I was flummoxed by the lack of the nonblocking flags on sockets on microsoft windows, until I realized that inside a nonblocking accept(2) call, accept(2) is just going to have to do something like a select(2).

So a non-blocking accept IS NEVER REQUIRED because you can either

• just accept one connection per iteration -- the other ones will still be there on the next iteration or
• When you have a hot listening socket, accept the first connection AND THEN SELECT AGAIN ON IT AND ONLY IT.

So instead of

  foreach (@Listeners){
   vec($rout,fileno($_),1) or next;
   # listener is nonblocking, goes until
   # expected accept failure
   while (accept(my $NewServer, $_)){
      push @Clients, $NewServer
[download]

you just do do

  foreach (@Listeners){
   vec($rout,fileno($_),1) or next;
   # listener is blocking, but we
   # know this listener is hot
   if (accept(my $NewServer, $_)){
      push @Clients, $NewServer
   }else{ log "accept: $!" }
[download]

Or, mock up the accept-em-all-NOW method without clumsily making accept fail:

  foreach (@Listeners){
   vec($rout,fileno($_),1) or next;
   # listener is blocking, but we
   # know this listener is hot
   acc:
   accept(my $NewServer, $_) and
      push @Clients, $NewServer;
   # select again to see if there's another
   my $rvec;
   vec($rvec,fileno($_),1) = 1;
   select($rvec,undef,undef,0);
   vec($rvec,fileno($_),1) and goto acc;
[download]

The difference is

by accepting all connections immediately we will possibly have more connections going, more suddenly. If we have a limit on our number of open connections, we only need to check it once per loop to keep from overrunning it.

How much can it affect throughput? It's like asking is it better for an office building to have a one-person revolving door that spins fast or a family-size one that spins slow.