[zeromq-dev] Proposal regarding PGM and ZMQ

[Martin Sustrik]

Martin Sustrik wrote:

> Yes. Something like this should be done. The question is what's the 
> optimal algorithm given that on faulty links the portion on bandwidth 
> needed for RDATA may be higher than 5% while with say IB it would be 
> minimal...

Thinking about it a bit more:

1. RDATA cannot take precedence to ODATA because that way slow/faulty 
consumers could consume considerable portion of bandwidth and eventually 
kill the actual data transfer.

2. ODATA cannot take precedence to RDATA because that way transferring 
at full speed would disable any repairs (Emmanuel's case).

So it looks like the bandwidth should be split in advance into ODATA 
bandwith and RDATA bandwidth. That means that even if there are no 
repairs part of the bandwidth would be unused, which is suboptimal but 
better than any of the above.

Maybe two socket options would be appropriate (ZMQ_RATE and 
ZMQ_REPAIR_RATE)?

Thoughts?
Martin