On 20 juin, 22:04, c00p <cdevi...@gmail.com> wrote:
> On Jun 20, 6:22 am, patrice <quet.patr...@neuf.fr> wrote:
>
>
>
>
>
> > Hello everybody,
> > I have some AIX boxes connected on EMC Symmetrix disk subsystems and
> > am wondering what value for queue_depth I may define on each hdisk.
> > By default, queue_depth for each hdisk is set to 16 by the driver.
> > Saying that a hdisk is a striped meta-volume on N physical disks
> > inside the disk subsystem, the value of 16 seems to me very poor.
> > IBM DS8000 disks may support up to 256 as the queue_depth value.
> > After asking it to EMC, they said, without particular clear
> > explanations, that the default is 16 and it has not to be increased.
> > I know that it is quite not recommended to bypass manufacturer advices
> > on such types of parameter.
> > But I would like to understand why an hdisk viewed by AIX, which is a
> > striped meta volume composed of 16 HVE on 16 different axes is limited
> > to only 16 I/Os in queue_depth, even if I understand that there are
> > other meta-volumes on the same EMC physical volumes accessed by others
> > nodes.
> > A queue_depth value equivalent to 16 on a meta volume composed of 16
> > HVE seems to me that there is a maximum of 1 I/O per HVE !
> > Am I right or did I miss something.
> > Many thanks in advance for your help, explanations, advices and ideas.
> > Patrice.
>
> I agree that 16 seems low...I don't see why you'd have any issues
> going with 2or3*N on a N+P array. Worst case you SHOULD get I/O
> errors in the errpt when you have high I/O on the system indicating
> the setting may be too high, but the driver should initiate retries to
> prevent actual failures. As long as you increment the value
> reasonably and don't get greedy I don't think you'll have any problems.- Masquer le texte des messages précédents -
>
> - Afficher le texte des messages précédents -

Thanks for your advices.
For me it should not be a big problem to increase those values, even
if EMC doesn't like that.
I would prefer having a kind of calculation formula, depending i.e. on
the way EMC subsystem is configured, how many LUNs are on the physical
drives, type of raid used, ... to approach as near as possible the
maximum value for our configuration without going into driver retries.
When a disk driver has to perform a lot of retries, the result could
be unpredictable, especially on I/O heavy loaded servers. Big hang
situations could hapen, which are definitely uncomfortable on
production servers.
If somebody has something like a calculation formula, coming from EMC
or not, I would be very interested in.
Thanks in advance to all.