GSA-RG
Session #1
29.01.07 - 9:00-10:30am

Lead: Philipp Wieder
notes by Ramin Yahyapour

- Philipp gives intro on progress since OGF18
  o Update on progress after OGF18
  o Goal: short-term feasibility study with actual
    implementation and show-case of scheduler interoperability
    = a job is submitted to an existing scheduler in a certain Grid middleware
      and this scheduler interacts with another scheduler covering a different
      Grid/Grid middleware to execute a job there.
      (not just submission as in HPC-P, but scheduling consideration)
    Involved partners/middlewares:
       - VIOLA-MSS/UNICORE
       - GRMS
       - D-Grid Scheduler (Globus-based; gLite based)
       - GridWay (status not clear, need to query)
    Time-line: Within the next twelve months

- Q: Clarification on what are scheduling domains:
  o different independent Grids with different Grid middlewares
  o therefore, we do not consider a single information source.
  o bridge only through agreement management (retrieve templates,
     parallel negotiation, final committments)


- Job Attributes
------------------

- only a very small subset of JSDL attributes have been identified
 which are supported by all reviewed systems
 o however, these systems could be extended to support more JDSL entries, e.g.
   VIOLA-MSS expressed their willigness

Q. Ali: why just subset and not the whole JSDL set?
 o the existing systems do not support all entries.
   The current efforts are focused on a simple, but feasible use case.


Q. Ali: why are the JSDL attributes at all important for scheduling?
  The matching of job and resource attributes could be generalized and
  independent from the specific attributes (just matching).
  Just the scheduler specific attributes are essential?
  o for the practical work on a feasibility study we need to agree also
    on these issues. In the long-run, the modeling of jobs and resources
    could be more extensive and generalized in the scheduler matching
    task except the Scheduling Description Attributes, see next item.


- Scheduler Description Language (SDL)
--------------------------------------
- presentation see slides:

Q. Ali: Does it make a difference to decide on a new scheduling language or
  some "JSDL extension" or something else? Only the attributes matter
 - no real difference, except process who finally produces a recommendation.
 - decision: the research group is not foreseen to produce recommendations,
   therefore we work on informational documents to outline the proposed steps
   for scheduler interoperability. The final adoption and transition to a WG
   can be discussed later.
 - thus, suggestion to publish the doc as INFO, and see whether people
   use and adopt it.

Q. what kind of dependencies are considered (the language has some
simple starting points for job dependencies)?
 - workflow is currently not in scope of the group
 - discussion showed that workflow is required by several groups
 - not clear how to cover this within this group
 - it is expected that the initial scheduler which is responsible for a
   job might consider dependencies and workflows during its scheduling.

Q. (Synapse) Is rescheduling covered?
 - currently not considered with the short term goals

Q. (Synapse) Standardisation of resource tags might be needed, is this covered?
 - this is an issue for a broader information model. This is out of scope for
   this group.

Q. How to interact with resource discovery in this process?
 - OGSA-RSS provides the interfaces for CSG and EPS.
   The resource discovery would be called through CSG.
   The scheduler would call this in its scheduling process.
   The scheduler itself would be accessed as an EPS

Q. is cost considered?
  - needs to be considered in the long run.
  - for now we consider flexible scheduling objectives which can include
    cost etc. as an optimization goal. Inclusion of cost would be simple, however
    the supporting services (accounting, billing, transformation of currency etc).
    could be complicating and are out of scope.

Q. Is load balancing considered? Is predictive scheduling considered?
  - This is part of the scheduling algorithm. The algorithm is a black box
    and can be adapted as needed. That is the reason why the interaction of
    schedulers is done through negotiation. This supports different independent
    scheduling strategies in the different schedulers.


- Negotiation Protocol
------------------------
- Presentation by Oliver, see slides



Session #2
29.01.07 - 11am-12:15pm

Lead: Ramin Yahyapour
notes by Philipp Wieder

- Execution Management
---------------------------
- Presentation by Ramin, see slides

After Agreement committment
- prior to job execution
 * Is the JSDL job description enough to start the job (short-term)?
  -> MSS: Could live the JSDL stage-in/sage-out.
  -> Usage of stage-in/stage-out might be a general option.
  -> OGSA BES? - Existing systems do not have a BES backend. Possible solution for later.
  -> Decision: use JSDL for stage-in/stage-out.
 * Is it ok that the job execution is automatically executed by the agreement provider?
  -> Yes.

- during job execution
 * Monitor the status of the agreement.
- after job execution
 -> stage-out via JDSL.
 -> evaluate the agreement after job execution

Authentication
- VOMS - UNICORE this summer
- VOMS - Globus unknown
-> Decision could not be made during the session, postponed (ask OMII-Europe guys for more info).

-> short term stay with JSDL, long-term check together with the "big picture" BES/VOMS/...

Discussion
----------
- Bring in the negotiation into WS-Agreement? -> Wolfgang: yes.
- Any aspects not covered? - Looks like it.
- No common information model yet, OGSA information model work still in preliminary phase.
- Proposal (Jay Unger): Use XML description of resource capabilities and XQuery to to the matching and XSLT to do the transformation.
 -> Talk to Michel Drescher about his approach.


Start discussion on "Big picture", architectural inclusion of other standards
-----------
- Whiteboard discussion  (more on paper)
- (Opwn) issues (among others)
 * Clarification wrt to OGSA-EMS is needed -> talk to these guys
 * Can/should the OGSA-BES interface be part of the scheduling interface?
  -> Oliver: Yes, is possible.


GSA telecon
------------
Will be set-up again bi-weekly, date and time will be circulated via mailing list



Action Items:
A1: Prepare INFO doc with architectural picture that shows the
   flow between the architectural blocks and outline the interaction
 = Ramin
A2: Prepare INFO doc with scheduler attributes, potential links to JSDL
 = Philipp (Ali volunteers to review the work)
A3: need to better understand the links to OGSA-BES and OGSA-HPCP
 = n.n.
A4: Describe the architectural blocks and its relation to other spec work like
   OGSA-RSS, OGSA-BES etc.
   -> See A1.
A5: Contact OMII-Europe wrt integration of VOMS into Globus, UNICORE, gLite
 = Philipp
A6: Talk to Michel Drescher about XML/XQuery/XPath approach to resource modeling
 = Philipp [DONE]


Next steps:
Picking up regular telecons (bi-weekly) e.g. end of February