[zeromq-dev] Integrating additional PGM rate limit controls into ØMQ

Martin Sustrik sustrik at 250bpm.com
Thu Dec 2 10:57:40 CET 2010


On 11/28/2010 11:48 AM, Steven McCoy wrote:
> On 27 November 2010 22:31, Martin Sustrik <sustrik at 250bpm.com
> <mailto:sustrik at 250bpm.com>> wrote:
>
>     On 11/27/2010 03:23 PM, Steven McCoy wrote:
>
>         It's complicated when you are constantly sending at full speed,
>         contention on transmission is likely to artificially restrict repair
>         throughput.  It's something that needs explicit testing for each
>         scenario.  The current implementation gives ODATA and RDATA equal
>         priority but a waiting send thread is likely to bias the weighting.
>           Explicitly giving RDATA priority is what is not productive.
>
>
>     Hm. Even "priotity up to N% of the overall bandwidth"?
>
>     My understanding was that the admin sets the RATE to limit the
>     network bandwidth used by particular application. Then he sets
>     REPAIR_RATIO to match the bandwidth needed for repairs under normal
>     circumstances.
>
>     When there is an congestion though, the need for rapair exceeds the
>     allocated capacity and thus some consumers will get unrecoverable
>     data loss -- which is a reasonable behaviour IMO.
>
>
> Due to ØMQ's design this is highly likely what you will see, if instead
> the design used multiple threads on one socket you could start to see
> unfair balancing.

Ok. So we can move on with simple two-parameter setup.

Martin



More information about the zeromq-dev mailing list