Re: [Anima-signaling] GRASP and unicast UDP

[Brian E Carpenter <[e-mail]>]

On 18/08/2016 08:15, Toerless Eckert wrote:
> Thanks, Brian
> 
> I think your self-experience with UDP should help to rule that out.
> Not sure how much of that experience summary could go for example
> into explanatory draft text. I am always a fan of that, but most people
> on the IETF seemingly not.

I will tune the section on Transport Layer Usage accordingly, and
maybe add a note to the Implementation Status section.

> Wrt to CoAP: I would hope that similar to bootstrap, we could define
> CoAP as an extension, so that we can continue to keep the main GRASP
> draft on a fast track to WGLC.

Yes, I agree. I will also add a note that other transports are "out
of scope for this document."

    Brian

> 
> Cheers
>     Toerless
> 
> On Thu, Aug 18, 2016 at 08:10:23AM +1200, Brian E Carpenter wrote:
>> On 17/08/2016 18:23, Toerless Eckert wrote:
>>> Thanks a lot for the insight. How about trying CoAP ?
>>
>> Two comments on that.
>>
>> 1) I woke up before the alarm went off this morning, and had exactly that
>> thought. But I'd have to study CoAP in detail to understand whether it
>> helps. Maybe Carsten can advise on this.
>>
>> (In GRASP, you may get multiple simultaneous requests or multiple simultaneous
>> discovery responses. The problem in a nutshell is that with TCP you have
>> accept(), which fires up a new socket for each incoming transaction, but with
>> UDP you have to manage sharing the socket yourself.)
>>
>> 2) I also realised that we can't leave it open in GRASP whether discovery
>> uses TCP or UDP for the discovery response. Otherwise discovery might not work
>> between two random implementations.
>>
>>    Brian
>>
>>>
>>> On Wed, Aug 17, 2016 at 04:42:40PM +1200, Brian E Carpenter wrote:
>>>> To entertain myself while waiting for comments on other matters, I spent
>>>> quite some time (and swear words) on hacking the prototype GRASP to use
>>>> UDP instead of TCP for unicast operations (i.e. discovery responses,
>>>> synchronization responses, and negotiation).
>>>>
>>>> My conclusion is that it's a fairly silly thing to do, except possibly
>>>> for discovery, which only involved a normal amount of debugging. I'm sure
>>>> this is something that DNS server developers are very familiar with:
>>>> writing a robust server handling things over UDP is a great deal harder
>>>> than doing it with TCP. In fact I gave up when I realised that for
>>>> the negotiation case (where there are an unpredictable number of messages
>>>> in both directions) I was basically going to have to rewrite half of TCP,
>>>> and still not have a robust solution. Without a lot more work, the UDP
>>>> solution is riddled with timing-dependent issues. (At one point, adding
>>>> a print statement for debugging fixed the bug, so that I simply couldn't
>>>> debug it...)
>>>>
>>>> Of course, all this could be solved with more work, but I'm really not
>>>> sure we should recommend UDP at all, except for discovery responses.
>>>> I can't really imagine a real autonomic node that doesn't have TCP code.
>>>>
>>>> Comments?
>>>>
>>>>     Brian
>>>>
>>>> _______________________________________________
>>>> Anima-signaling mailing list
>>>> [e-mail]
>>>> https://www.ietf.org/mailman/listinfo/anima-signaling
>>>
>