> Well, the default is 16k for both (see net.inet.tcp.sendspace and
> net.inet.tcp.rcvspace).  You certainly want it to be at least
> as large as the default MSS, I would think, although I guess that's
> not strictly necessary, either.  Certainly 1k or 2k as a receive window
> size would not break TCP (and if it did, it's a bug).

I seem to remember reading a couple of papers about TCP performance 
quirks when the window/buffer size was under 3*MTU.  That was several 
years ago.  It didn't break anything in the sense of crashing or 
connections timing out.  It just didn't go very fast.  It triggered 
the 200 ms timeouts. 

I'd run some experiments before I depended on tiny buffers doing 
the right thing.