airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Shameera Rathnayaka <shameerai...@gmail.com>
Subject Re: Airavata Output Processing Doubt
Date Tue, 17 Jan 2017 13:19:28 GMT
Yes this seems to be a good solution. Yes we can assume job writers to pre
configured file with key value pair.

On Mon, Jan 16, 2017 at 12:58 PM Ajinkya Dhamnaskar <adhamnas@umail.iu.edu>
wrote:

Shameera,

Now, I got your concern. This is an inherent problem. I was actually
referring to STDOUT and STDERR which job generates after execution.
I was wondering if we can write this output as key value pair in some text
file under specific directory for particular JOB.

So, basically every experiment will have output staging task, but in case
of output types (INTEGER, STRING etc) the file which will be staged would
have output in key-value pairs.
For example, lets consider the case that you mentioned with two string
outputs, we can probably generate file on remote server as str1-value2 \n
str2-value2 and stage the same.
But, I understand this needs to be implemented in job scripts and requires
much more than this.

How would you think through this idea? Do you think this has some potential?


On Mon, Jan 16, 2017 at 6:55 AM, Shameera Rathnayaka <shameerainfo@gmail.com
> wrote:

Hi Ajinkya,


As and when job gets completed we save output for the same. I was
wondering, if we can get that information knowing process_id and task_id.
Job table has process_id and task_id, possibly we can fetch output stored
in Job table.

If you check the JobModel we don't associate Job output with it. So no job
outputs in JobTable. We have to fetch output and save it in registry(I
think that is what you refering to log output in our previous replies)
problem is how to fetch and save if the job is completed and output is not
a file. If we can find a solution to this we can extend this to support
multiple outputs as well.



Also, in case of multiple outputs, each task would know its output name and
possibly we can use that name alongside process_id for fetching correct
value. Workflow would know which output to use as an input for other
application. I hope, I understood your concern correctly.


On Sun, Jan 15, 2017 at 11:59 PM, Shameera Rathnayaka <
shameerainfo@gmail.com> wrote:

This approach sounds promising to me, if we have more than one non-file
outputs then we will have more than OutputLogginTasks. But how exactly this
OutputLogginTask(please think of a new name for this task) read the data,
because by the time OutputLogginTask is invoked, the actual job in target
computer resource is completed and output is where? if we have more than
one OutputLogginTasks how it reads the value associated with it. eg: if my
job output two Strings "str1", "str2" and I am using "str2" for downstream
application in my workflow how we can guarantee downstream application
always get the correct value?

On Sun, Jan 15, 2017 at 1:02 PM Ajinkya Dhamnaskar <adhamnas@umail.iu.edu>
wrote:

Hi Shameera,

If you check
org.apache.airavata.orchestrator.cpi.impl.SimpleOrchestratorImpl#createAndSaveOutputDataStagingTasks()
method, we entertain output staging task only when output data type is
STDOUT, STDERR and URI. I am suggesting, in default case we will create
different task which points to OutputLogginTask.

OutputLogginTask is nothing but yet another implementation of task, similar
to SCPDataStageTask where we stage files and log output as well. But,
in OutputLogginTask
we need not to stage any data, we would just log data as it come.

I am assuming TaskId and ProcessId is sufficient to fetch output. (please
correct me if you don't think so)

Thanks Shameera, this discussion is helping me a lot.

On Sun, Jan 15, 2017 at 9:38 AM, Shameera Rathnayaka <shameerainfo@gmail.com
> wrote:

Hi Ajinkya,

It is not clear to me how this "OutputLogginTask" knows about job output
data. It would be helpful if you can explain your suggested approach bit
more.

Best,
Shameera.

On Sat, Jan 14, 2017 at 1:07 PM Ajinkya Dhamnaskar <adhamnas@umail.iu.edu>
wrote:

Hi Shameera,

One possible solution would be to introduce OutputLoggingTask. We can
create output task irrespective of output data type and if there isn't any
file to stage we can call OutputLoggingTask.
Its sole purpose is to log data, that way we can justify each output type.

Please suggest, in case you think of any better solution.

Thanks in anticipation.


On Sat, Jan 14, 2017 at 9:46 PM, Shameera Rathnayaka <shameerainfo@gmail.com
> wrote:

Hi Ajinkya,

Yes, that is the case. how would you plan to solve it?

Regards,
Shameera.

On Fri, Jan 13, 2017 at 6:37 AM, Ajinkya Dhamnaskar <adhamnas@umail.iu.edu>
wrote:

Amila,

Thanks Amila for explaining. It really explains how things are mapped. I
could see output against JOB but could not figure out from where exactly we
are logging output for a process.

Shameera,

Yeah, that's true. So basically, if application does not have output
staging task, it would not log output for respective process.
Which means if output data type is not URI, we are not logging output
against process.(Please correct me if I am wrong).

Probably, here we have an opportunity to improve.

Thanks in anticipation

On Fri, Jan 13, 2017 at 8:58 AM, Shameera Rathnayaka <shameerainfo@gmail.com
> wrote:

Hi Ajinkya,

If you check
here org.apache.airavata.gfac.impl.task.SCPDataStageTask#outputDataStaging
you will see that we are saving process outputs to database(through
registry). You probably testing with local job submission
with org.apache.airavata.gfac.impl.task.DataStageTask as data staging task
implementation. There we don't save process outputs. First thing is to fix
this and save the process outputs to the database.

If you know the JobId then you can retrieve processId from Job model. Using
processId you can get all process outputs. see PROCES_OUTPUT case in
org.apache.airavata.registry.core.experiment.catalog.impl.ExperimentCatalogImpl#get(..,..)
method.

Hope this will help you to move forward.

Best,
Shameera.

On Thu, Jan 12, 2017 at 3:48 PM Amila Jayasekara <thejaka.amila@gmail.com>
wrote:

Hi Ajinkya,

I am not familiar with the context of your question but let me try to
answer.

If you are referring to an application deployed in a supercomputer, then
the application should have a job id. In the supercomputer, each
application runs as a separate batch job and each job is distinguished
using the job id (similar to process id in a PC). Usually, the job
scheduler returns this job id and Airavata should be aware about that job
id. Then, you should be able to use this job id to identify the output,
provided job script specify instructions to generate output.

I did not understand what you referred as "process model" and "job model".
I assume these are database tables.

Thanks
-Amila



On Wed, Jan 11, 2017 at 1:17 PM, Ajinkya Dhamnaskar <adhamnas@umail.iu.edu>
wrote:

Hello Dev,

I am trying to fetch application output (type:INTEGER) after experiment
completion. As per my understanding each application runs as a process and
that process should have final output.

So, ideally we should be able to get final output from process id itself
(correct me if I am wrong).
In my case, I am not seeing final output in database. Basically, we are not
updating process model after job completion, we update job model though.

Am I missing anything here?

Any help is appreciated.

-- 
Thanks and regards,

Ajinkya Dhamnaskar
Student ID : 0003469679
Masters (CS)
+1 (812) 369- 5416 <(812)%20369-5416>







-- 
Shameera Rathnayaka





-- 
Thanks and regards,

Ajinkya Dhamnaskar
Student ID : 0003469679
Masters (CS)
+1 (812) 369- 5416 <(812)%20369-5416>







-- 
Best Regards,
Shameera Rathnayaka.

email: shameera AT apache.org , shameerainfo AT gmail.com
Blogs : https://shameerarathnayaka.wordpress.com ,
http://shameerarathnayaka.blogspot.com/







-- 

-- 
Shameera Rathnayaka

Mime
View raw message