On Tue, Apr 15, 2014 at 2:51 PM, Amila Jayasekara
<[email protected]>wrote:

> On Tue, Apr 15, 2014 at 2:38 PM, Lahiru Gunathilake <[email protected]
> >wrote:
>
> > On Tue, Apr 15, 2014 at 2:29 PM, Marlon Pierce <[email protected]> wrote:
> >
> > > When is a resource bound to an experiment?  In at least some cases,
> > > max_walltime, cpu count, etc can be resource specific for a given
> > > application.
> > >
> > During launching in orchestrator. Currently these are read from
> application
> > description and its resource specific. Each resource has its own
> > application descriptor.
> >
>
> Is that so ?
> i.e. user specifies the resource when launching the application ? not when
> creating the experiment ?
>
When user create experiment it gives the host name and during launching the
actual binding happens by loading the appropriate application descriptor.

>
> Thanks
> Amila
>
>
> >
> > Lahiru
> >
> >
> > >
> > > Marlon
> > >
> > > On 4/15/14 2:18 PM, Amila Jayasekara wrote:
> > > > Hi Lahiru,
> > > >
> > > > Some thoughts below;
> > > >
> > > > Regarding 1
> > > >
> > > > Parameters such as max_wall time, cpu count, queue name etc ...
> should
> > be
> > > > validate at the experiment creating stage.
> > > > Reason : We create experiment in one call and we launch the
> experiment
> > > in a
> > > > different call. So if user specifies an invalid or out of range value
> > for
> > > > walltime or cpucount, he/she will get an error at experiment launch
> > time,
> > > > not during creation time. Thats going to be tedious to user.
> Therefore
> > we
> > > > need to validate those job specific parameters during experiment
> > > creation.
> > > > How to implement : Use gsissh to login to resource and execute a
> > command
> > > > similar to "qstat -q <queue name>" and parse the output. It gives
> most
> > of
> > > > above parameters. Also there are other commands you could use to get
> > such
> > > > information (But commands will depend on the job scheduler you are
> > > using).
> > > >
> > > > Regarding 3
> > > >
> > > > It doesnt sound correct to copy what ever the status when cloning the
> > > > experiment. This could break some of the logic based on experiment
> > status
> > > > and may give wrong information to user.
> > > >
> > > > Thanks
> > > > Amila
> > > >
> > > >
> > > > On Tue, Apr 15, 2014 at 1:47 PM, Lahiru Gunathilake <
> [email protected]
> > > >wrote:
> > > >
> > > >> Hi Eroma,
> > > >>
> > > >> Please find my answers below.
> > > >>
> > > >>
> > > >> On Mon, Apr 14, 2014 at 10:04 AM, Eroma Abeysinghe <
> > > >> [email protected]> wrote:
> > > >>
> > > >>> Hi All,
> > > >>>
> > > >>> I was preparing test cases for Airavata API and few
> > > >>> questions/clarifications;
> > > >>> 1. Do we have a max walltime and CPU count?
> > > >>>
> > > >> Not sure what did you mean by do we have. Yes we have those two
> > > properties
> > > >> and its used in most of the unit tests and createLaunchExperiment
> > > sample.
> > > >> You can have a look at the code. Currently its mostly a static
> > > >> configuration where gateway administrator set when he/she save the
> > > >> application context but these properties can be configured for each
> > > request
> > > >> (job) rather using the static configuration.
> > > >>
> > > >>> 2. We don't define ppn for resource node at creation of
> experiments.
> > Is
> > > >> it
> > > >>> set internally for each experiment?
> > > >>>
> > > >> Yes but we set this in application Descriptor currently and its been
> > > read
> > > >> in gfac, but we can set in Experiment. To clarify this I have
> answered
> > > your
> > > >> last question and please refer that.
> > > >>
> > > >>> 3. When we clone an experiment what is the status that the new
> cloned
> > > >>> experiment get saved? Is it CREATED ?
> > > >>>
> > > >> We don't save specific state, we simply clone the experiment with a
> > new
> > > >> experiment ID, whatever the status get copied.
> > > >>
> > > >>> 4. Do we have  a defined mapping between experiment status and task
> > > >> status?
> > > >> Currently they are independent statuses. if you can have a look
> > > >> experimentModel.thrift you can see that. but I think we can have
> some
> > > >> constraint and keep a mapping to avoid conflicting states between
> > > >> experiments, tasks and Jobs. Currently we don't have any rules
> defined
> > > >> between these statuses.
> > > >>
> > > >>>
> > > >>> --
> > > >>> Thank You,
> > > >>> Best Regards,
> > > >>> Eroma
> > > >>>
> > > >>
> > > >>
> > > >> --
> > > >> System Analyst Programmer
> > > >> PTI Lab
> > > >> Indiana University
> > > >>
> > >
> > >
> >
> >
> > --
> > System Analyst Programmer
> > PTI Lab
> > Indiana University
> >
>



-- 
System Analyst Programmer
PTI Lab
Indiana University

Reply via email to