airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Saminda Wijeratne <samin...@gmail.com>
Subject Re: Persisting GFac job data
Date Tue, 21 May 2013 21:04:00 GMT
But I thought the providers are part of the GFac (not as a separate
service). If not then the providers should report to GFac. Orelse there is
no way the GFac knows what status to update which data to update etc. Does
the current GFac implementation support this?


On Tue, May 21, 2013 at 4:47 PM, Amila Jayasekara
<thejaka.amila@gmail.com>wrote:

> I think that should be handled at a more upper layer like Workflow
> Interpretter or GFac. In FT perspective it is better if providers are
> stateless. One reason is we dont have control over some providers and and
> there will be many places writing to disk if we implement the persistence
> logic at provider level.
>
> Thanks
> Amila
>
>
> On Tue, May 21, 2013 at 4:39 PM, Saminda Wijeratne <samindaw@gmail.com
> >wrote:
>
> > On Tue, May 21, 2013 at 4:36 PM, Amila Jayasekara
> > <thejaka.amila@gmail.com>wrote:
> >
> > > On Tue, May 21, 2013 at 3:51 PM, Saminda Wijeratne <samindaw@gmail.com
> > > >wrote:
> > >
> > > > Thanks for the feedback Amila. a few comments inline
> > > >
> > > >
> > > > On Tue, May 21, 2013 at 12:29 PM, Amila Jayasekara
> > > > <thejaka.amila@gmail.com>wrote:
> > > >
> > > > > Hi Saminda,
> > > > >
> > > > > Great suggestion. Also +1 for Dhanushka's proposal to have
> > > > > serialize/de-serilized data.
> > > > > Few suggestions,
> > > > > 1. In addition to successful/error statuses we need other status
> for
> > > > nodes
> > > > > & workflows
> > > > > and workflows.
> > > > > E . g :-
> > > > >    node - started, submitted, in-progress, failed, successful etc
> ...
> > > > >
> > > > Sorry if I was too vague. Yes we have more fine-grain statuses for
> > > workflow
> > > > and node[1]. We will have a much fine-grained level of granuality
> for a
> > > > GFacJob status.
> > > >     public static enum GFacJobStatus{
> > > >         SUBMITTED, //job is submitted, possibly waiting to start
> > > executing
> > > >         EXECUTING, //submitted job is being executed
> > > >         CANCELLED, //job was cancelled
> > > >         PAUSED, //job was paused
> > > >         WAITING_FOR_DATA, // job is waiting for data to continue
> > > executing
> > > >         FAILED, // error occurred while job was executing and the job
> > > > stopped
> > > >         FINISHED, // job completed successfully
> > > >         UNKNOWN // unknown status. lookup the metadata for more
> > details.
> > > >     }
> > > >
> > > >
> > > > 2. This data will be useful in implementing FT and Load Balancing in
> > each
> > > > > component. Sometime back we had discussions to make GFac stateless.
> > So
> > > > who
> > > > > is going to populate this data structure and persist it ?
> > > > >
> > > > That is a very good question... :). This summer is going to be a long
> > > > one... ;)
> > > >
> > >
> > > What I meant is which component is doing persistence ? (GFac or WF
> > > Interpretter). Not the actual person who is going to implement it :).
> > >
> > hih hih....
> > Well its going to be whatever the provider respondible for managing the
> job
> > lifecycle. For example GRAMProvider should be responsible for recording
> all
> > the data relating to the GRAM jobs its working with.
> >
> > >
> > >
> > > >
> > > > 1.
> > > >
> > > >
> > >
> >
> https://svn.apache.org/repos/asf/airavata/trunk/modules/workflow-model/workflow-model-core/src/main/java/org/apache/airavata/workflow/model/graph/Node.java
> > > >
> > > > >
> > > > > Thanks
> > > > > Amila
> > > > >
> > > > >
> > > > > On Tue, May 21, 2013 at 11:39 AM, Saminda Wijeratne <
> > > samindaw@gmail.com
> > > > > >wrote:
> > > > >
> > > > > > Thats is an excellent idea. We can have the job data field to
be
> > the
> > > > > > designated GFac job serialized data. The whatever GFacProvider
> > should
> > > > > > adhere to it.
> > > > > >
> > > > > > I'm still inclined to have the rest of the fields to ease of
> > querying
> > > > for
> > > > > > the required data. For example if we wanted all attempts on
> > executing
> > > > > for a
> > > > > > particular node of a workflow or if we wanted to know which
> > > application
> > > > > > descriptions are faster in execution or more reliable etc. we
can
> > let
> > > > the
> > > > > > query language deal with it. wdyt?
> > > > > >
> > > > > >
> > > > > > On Tue, May 21, 2013 at 11:24 AM, Danushka Menikkumbura <
> > > > > > danushka.menikkumbura@gmail.com> wrote:
> > > > > >
> > > > > > > Saminda,
> > > > > > >
> > > > > > > I think the data container does not need to have a generic
> > format.
> > > We
> > > > > can
> > > > > > > have a base class that facilitate object
> > > > serialization/deserialization
> > > > > > and
> > > > > > > let specific meta data structure implement them as required.
We
> > get
> > > > the
> > > > > > > Registry API to serialize objects and save them in a meta
data
> > > table
> > > > > > (with
> > > > > > > just two columns?) and to deserialize as they are loaded
off
> the
> > > > > > registry.
> > > > > > >
> > > > > > > Danushka
> > > > > > >
> > > > > > >
> > > > > > > On Tue, May 21, 2013 at 8:34 PM, Saminda Wijeratne <
> > > > samindaw@gmail.com
> > > > > > > >wrote:
> > > > > > >
> > > > > > > > It has being apparent more and more that saving the
data
> > related
> > > to
> > > > > > > > executing a jobs from the GFac can be useful for many
reasons
> > > such
> > > > > as,
> > > > > > > >
> > > > > > > > debugging
> > > > > > > > retrying
> > > > > > > > to make smart decisions on reliability/cost etc.
> > > > > > > > statistical analysis
> > > > > > > >
> > > > > > > > Thus we thought of saving the data related to GFac
jobs in
> the
> > > > > registry
> > > > > > > in
> > > > > > > > order to facilitate feature such as above in the future.
> > > > > > > >
> > > > > > > > However a GFac job is potentially any sort of computing
> > resource
> > > > > access
> > > > > > > > (GRAM/UNICORE/EC2 etc.). Therefore we need to come
up with a
> > > > > > generalized
> > > > > > > > data structure that can hold the data of any type
of
> resource.
> > > > > > Following
> > > > > > > > are the suggested data to save for a single GFac job
> execution,
> > > > > > > >
> > > > > > > > *experiment id, workflow instance id, node id* - pinpoint
the
> > > node
> > > > > > > > execution
> > > > > > > > *service, host, application description ids *- pinpoint
the
> > > > > descriptors
> > > > > > > > responsible
> > > > > > > > *local job id* - the unique job id retrieved/generated
per
> > > > execution
> > > > > > > > [PRIMARY KEY]
> > > > > > > > *job data* - data related executing the job (eg: the
rsl in
> > GRAM)
> > > > > > > > *submitted, completed time*
> > > > > > > > *completed status* - whether the job was successfull
or ran
> in
> > to
> > > > > > errors
> > > > > > > > etc.
> > > > > > > > *metadata* - custom field to add anything user wants
> > > > > > > >
> > > > > > > > Your feedback is most welcome. The API related changes
will
> > also
> > > be
> > > > > > > > discussed once we have a proper data structure. We
are hoping
> > to
> > > > > > > implement
> > > > > > > > this within next few days.
> > > > > > > >
> > > > > > > > Thanks,
> > > > > > > > Saminda
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message