[zeromq-dev] Does ZMQ "Over Send" Using OpenPGM

[Steven McCoy]

On 21 October 2010 00:07, Martin Sustrik <sustrik at 250bpm.com> wrote:

> Ok. So as far as I understand it makes no sense to use PGM with multiple
> multicast groups. You should choose *one* multicast group and pass all the
> traffic on that multicast group, distinguishing different feeds only by port
> number.
>
> If you fail to do so, you'll get random unexpected behaviour, such as the
> one Bob have experienced.
>
>
Normally you have a IGMP aware network infrastructure that filters the
multicast group for you so you do not receive everything.  If you overload
the switch and turn it into a hub you are going to have more problems than
PGM receiving applications multiplexing incoming packets.

Multiple multicast groups are required for scaling up, i.e. reducing the
recovery burden for each transport, and to take advantage of network
filtering - splitting the application domain across multicast groups such
that only a small number are in use by a receiving node at any one time.

If all receiving nodes are receiving all multicast groups you lose out on
network filtering but you still gain due to less reliability overhead per
group and fair queuing between segments.  You are going to get a big latency
peak when PGM has to handle a retransmit request and so if you have a big
data feed it is going to queue up pretty fast and start dropping data if you
are not careful.

-- 
Steve-o
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20101021/da9c77e4/attachment.htm>