[DRBD-user] peer-max-bio-size 1M

[Ben RUBSON]

Le 11 avr. 2014 à 19:10, Lutz Vieweg <lvml at 5t9.de> a écrit :
> On 04/11/2014 09:02 AM, Lars Ellenberg wrote:
> >> peer-max-bio-size out of range (0...128k)
> >>
> >> Did I miss something ?
> >
> > That's apparently an oversight (Bug),
> > there is still a value of 128k hardcoded in drbdmeta.
> 
> What a timing - I was just about to submit a report
> that a block device stack "dm-crypt on drbd" delivers
> only half the throughput of a "drbd on dm-crypt" setup
> on the same underlying physical storage when I read this.
> 
> But not it seems the relevant reason was that
> I benchmarked the performance before connecting a
> second node.
> 
> It's kind of counter-intuitive that running a benchmark
> on only one local disk does not avoid but actually introduce
> a subtle additional reason for losing performance...
> 
> >> Will both nodes discuss about the value to use up to 1M, even if I don't
> >> use "--peer-max-bio-size 1M" in drbdmeta ?
> >
> > If you have two nodes,
> > don't use that option to drbdmeta.
> 
> But that doesn't answer the question - which IMHO is a really
> important one, as lower values of queue/max_hw_sectors_kb can
> dramatically reduce performance, especially when using fast SSDs.
> And 128k _is_ a too low value for usage with contemporary SSDs.

From what I have understood, you can simply ignore this parameter,
as both nodes once connected will discuss together about the best value
to work with (up to 1M).
128k limitation is on manual drbdmeta input only.
However, I may be wrong with my understanding :)

> > For any normal use case, using that option to drbdmeta is simply wrong.
> > So don't.
> 
> I would recommend to put a big fat caveat emptor on the
> "peer-max-bio-size" / max_hw_sectors_kb issue in the documentation
> where the intial setup of a DRBD device is explained, and not
> hide it in the drbdmeta man-page where people are unlikely
> to ever spot it and realize its consequences.
> 
> Regards,
> 
> Lutz Vieweg