[nanomsg] Re: Fanout pattern

  • From: Holger Winkelmann <hw@xxxxxxxxxxxxxx>
  • To: Martin Sustrik <sustrik@xxxxxxxxxx>
  • Date: Sat, 19 Jan 2013 21:15:56 +0100 (CET)

Sounds reasonable ... And smell like routing protocols at least a bit as BGP 
tables are getting larger and need memory at the router. I rather compare it to 
switches where TCAMs are expensive and small,  At Least today.

Thanks for the clarification, seems network principles can arrive in 
applications. 

--

Holger Winkelmann
Travelping GmbH
+49-171-5594745

### Sent from a mobile device. Sorry for brevity and typos... ###

On 19.01.2013, at 17:17, Martin Sustrik <sustrik@xxxxxxxxxx> wrote:

> On 18/01/13 16:37, Holger Winkelmann [TP] wrote:
> 
>> Sounds OK. But whats the difference of the Subscriber Filter algorithm
>> at subscriber compared to the router? I suspect the Router need to combine
>> the Filter of all Subscribers which have messages passing the router? And 
>> this
>> will either lead to huge tables or to a algorithm which may is not perfect,
>> right?
> 
> Yes. Exactly. Assuming that memory on the router is limited, we have to do 
> imperfect filtering. (E.g. creating a hash, where hash collisions result in 
> non-matching messages passing the filter.) AFAIK that's exactly how IP 
> multicast group filtering is implemented in most network switches and NICs.
> 
>> My question would be: has this any consequences to the behavior or is
>> it "just" adding some costs for the transport and bandwidth?
> 
> The point of the architecture is that there's no impact on behaviour. Perfect 
> message filtering is done on the terminal subscriber, so the end user always 
> gets perfectly filtered messages. Forwarding subscriptions to intermediate 
> nodes is strictly an optimisation and has effect only on bandwidth usage, 
> never on semantics of the application.
> 
> Martin

Other related posts: