[MITgcm-devel] [MITgcm-support] asyncio

[Ryan Abernathey]

Dimitris, thanks for your explanation. It sounds like an extremely useful
addition to the model!

I second Martin's suggestion enthusiastically. It would be great to have
this become a permanent, stable feature of the model. I assume the
roadblock is finding the developer time to incorporate and test the new
code. The big LLC simulations are exciting enough scientifically and
computationally that it is feasible to write proposals to develop these
capabilities further. I have no idea what long-term roadmap you have in
mind, but I hope this is being considered. I would be happy to collaborate
if so.

Cheers,
Ryan


On Wed, Oct 19, 2016 at 1:03 AM, Martin Losch <martin.losch at awi.de> wrote:

> Hi Dimitris,
> changing to devel-list:
>
> does it make sense to add this to the main trunk? We all are and will be
> running large scale setups with lots of I/O (for me it’s not a bottle-neck,
> but I start to feel it), so this asynchronous I/O would be great. How much
> work is it and how specific is it to the file system that you are using?
>
> Martin
>
> > On 18 Oct 2016, at 19:45, Dimitris Menemenlis <dmenemenlis at gmail.com>
> wrote:
> >
> > Hi Ryan, in theory code-async should work on any lustre file system that
> is similar to
> > the one used on pleiades.  In practice it has only been tested on
> pleiades, so there
> > will probably be a little trial and error in porting.
> >
> > For LLC_1080 and a decent file system, you probably don’t need
> code-async.
> > I have run LLC_1080 and also LLC_2160 on pleiades with standard
> useSingleCpuIO
> > and it run OK for 1080 and OKish for 2160.  For LLC_4320, the I/O got
> bogged down.
> >
> > useSingleCpuIO uses mpi routines to do all the I/O from the head node.
> >
> > code-async reserves a number of additional cores (~3 to 10% on top of
> those used
> > by MITgcm computational cores) for I/O only.  As the code runs, these I/O
> > cores collect the output from MITgcm asynchronously and then take care of
> > writing it to lustre file system.  so the I/O does not add any
> additional time to computation.
> >
> > Dimitris Menemenlis
> >
> >> On Oct 18, 2016, at 9:30 AM, Ryan Abernathey <ryan.abernathey at gmail.com>
> wrote:
> >>
> >> Hi,
> >>
> >> What is the deal with the code-async folders in the llc_hires setups?
> >> http://mitgcm.org/viewvc/MITgcm/MITgcm_contrib/llc_hires
> >>
> >> I am working on setting up an LLC_1080 run. The examples in the readme
> files all include building with this extra code directory.
> >>
> >> What does it do?
> >>
> >> Does it only work on pleiades?
> >>
> >> Cheers,
> >> Ryan
> >>
> >>
> >> _______________________________________________
> >> MITgcm-support mailing list
> >> MITgcm-support at mitgcm.org
> >> http://mitgcm.org/mailman/listinfo/mitgcm-support
> >
> > _______________________________________________
> > MITgcm-support mailing list
> > MITgcm-support at mitgcm.org
> > http://mitgcm.org/mailman/listinfo/mitgcm-support
>
>
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mitgcm.org/pipermail/mitgcm-devel/attachments/20161019/d8377018/attachment.htm>