Re: [hybi] Frame size

["Thomson, Martin" <[e-mail]>]

You make a good case... for SCTP :p

--Martin

> -----Original Message-----
> From: Jamie Lokier [mailto:[e-mail]]
> Sent: Monday, 19 April 2010 12:01 PM
> To: Thomson, Martin
> Cc: Justin Erenkrantz; Hybi
> Subject: Re: [hybi] Frame size
> 
> Thomson, Martin wrote:
> > Jamie writes:
> > > Alas, there are very good reasons to send non-final frames that
> aren't
> > > the full length.
> >
> > From my perspective, the only very good reason you stop sending one
> thing is because you want to send something else (i.e. multiplexing).
> >
> [...]
> >
> > Addressing shortcomings in the abstraction (graceful close, etc...)
> > is another thing.  It's reasonable to say that the TCP abstraction
> > does not provide graceful close, so providing it at a higher layer is
> > OK.
> 
> If you want to send a graceful close response when you have only sent
> half a message, you cannot do it unless the half message was sent as
> a non-final, non-full-length frame.
> 
> That's why it's useful.
> 
> It's not because you want to send something else _now_, it's so you
> have the _possibility_ to send something else later.  Because you
> can't abort a half-sent frame later.  (You would need TCP urgent data
> or something, and that's not feasible to use.)
> 
> > Aligning with segments isn't a good argument for me.  How the WS
> > layer is aware of layers below it and the way that it interacts with
> > those is orthogonal to frame size.  TCP provides an abstraction of a
> > stream.  If a particular implementation wants to optimize, it can,
> > but that shouldn't affect layers above those key abstraction points.
> 
> Feel free to ignore segment alignment.  It's just an implementation
> tunable, not a requirement and doesn't need to appear in any
> specification.
> 
> The frame size is part of the WS transport layer, and is closely
> related to buffering and forwarding decisions.  (Note that eager
> sending is sub-optimal for latency, in case you're curious.)
> 
> Those decisions have little or nothing to do with TCP segments, but
> the timing of forwarding decisions cause certain kinds of alignment
> anyway.  The same outcome occurs with other types of pipe abstraction.
> 
> You said "it shouldn't affect layers above those key abstraction
> points".  Note that frame size is *not* present in the layer above WS
> buffering and forwarding decisions.  (If it is, it is misdesigned.)
> 
> You might be thinking of the message size, which is.
> 
> -- Jamie