Johannes and all, Lot of thanks to Johannes for his notes. I have improved below the formulation of : - what I said, - what I understood other participants said. Best regards. Etienne URBAH Meeting - Tuesday, 8 Dec 2009, 16:00 (CET) Participants: Morris, Balazs, Etienne, Johannes, Aleksander Agenda: Discussion of the work of Etienne on the "OGF PGI - AGU Execution Service Strawman Rendering" Morris: Go through meeting notes from last meeting Would be good to have gLite participation Balazs: We need at least one more person from gLite Real work can only be done with a screensharing tool Just discuss the ideas behind Etienne's work Balazs (chairing the strawman rendering document discussion): Doc is the latest version of the doc (version 37) by Aleksander on the PGI GridForge page Page 4 Synchronize, merge Etiennes's changes with actual model Inside createActivity (second sentence) Etienne: The initial state of a job is the submitted state Balazs: Substates ? Etienne: Substates of 'Submitted' state are : Incoming, Waiting, Outgoing - Incoming = Job inside input queue, has not been processed yet - Waiting = JSDL is being processed - Outgoing = JobId or EPR has been given to the job, ready for next state Any change must be performed in a consistent way Balazs: Combining substates in a certain way -> clarify more Not combine submitted outgoing with response Etienne: It is useful that each state really represent something If you do not give JobId at the end of 'Submitted' state, you have to completely include the 'Submitted' state inside the 'Pre-processing' state. Balazs: Only the first level states, 2nd third are optional? Etienne: Important: at the end of submitted state - JobId associated Balazs: Why at the end ? Etienne: At latest at the end ! Could be returned earlier also It is useful for the execution service for checking if it is possible (JSDL checking) Balazs: Not validate JSDL here When does execution service return response ? Morris: Page 4, comments - In terms of UNICORE it is alway possible directly, you don't do brokering - Not exactly sure how this relates to the state model - Out of the scope of the session ? Balazs: Now, we are not discussing staging of data We are discussing in which state job should be I do not see this written down now Morris: It is in the job description document validation Balazs: Missing the state model How do you read the job description validation? Morris: Matchmaking If there is a must - there is a must In gLite, first from the broker, then from execution service Balazs: First mandatory validation, then response containing maybe validation errors You have to validate first, to check for errors Etienne: First mandatory validations: XML, Schema, Semantic, Some validation may be only possible later ? Balazs: All validation should be done before, then assign state You submit a job, execution service does couple of things like validation The first time the submitter can check the state is when he/she has the JobId Etienne: At the beginning, I did not write the second level substates of 'Submitted'. I added them after some request, and I am ready to remove them. Balazs: Hold point is important : Substate of 'Submitted' which is a 'Hold' ! Etienne: Concerning the 'Submitted:Hold' substate, I sent a mail on 13 August Balazs: changeState operation resume like operations andy kind of hold states Morris: We need a figure : - not only the state model figure - but a step by step base like figure of presentation Etienne: We have to agree on the scope of the 'Submitted' state : Inside the 'Submitted' state, as soon a job has its JobId, the job should go to the 'Pre-processing' state, where it can be put in a 'Hold' substate Balazs: Is 'Pre-processing' different from data staging ? Etienne: 'Pre-processing' includes data staging Any other actions can be performed That's why there are third level 'Hold' substates There is 'Pre-processing:Hold:Suspended' when user wants 'Hold' without relationship to staging Morris: i.e. preprocessiong stage, only one CPU before using the whole machine not related to data staging Balazs: 'Pre-processing' is very complex state Morris: During the execution vs. pre and post processing "before finished" Other requirement for putting it somewhere Etienne: There is also the 'Post-processing:Hold:Suspend' state Morris: Clearly identify it is 'Post-processing' : Job is finished Etienne: Inside second level 'Hold' substates, third level substates can occur in any order at any time you want Balazs: Pre, post processing very complex, dangerous I do not see a sequence how the job goes through I really would like to see a linear state model Etienne: I agree with Balazs, inside the preprocessing hold there is no clear indication in which order the third level substates are processed I implicitely suppose that they can be processed in any order and any number of times - If it is achieved only by the implementation, then it is OK - If the order must be well known from the beginning, we have to completely review the state model Balazs: When exactly this will happen? Put the job in 'Hold' state before any kind of data staging takes place Etienne: If we keep model exactly as it is, it is not very precise, and several middlewares can implement it more easily. If we create a too precise model, it will be only implemented by one middleware. Too many details will limit the chance to have different middlewares. Are you sure you want to do that ? Balazs: Just want to stop the Job before 'Pre-processing' Morris: Use case? You don't do data staging then? Balazs: Use case : For safety reasons, the Submitter just wants to be sure that the job does not perform anything. Morris: 'Pre' means before 'Pre-processing' is not data staging Balazs: Inside ARC it is a safety measure to be able to stop the job before running Morris: I do not really see the difference User gets the feedback -> 'Submitted' 'Submitted:Hold' sounds strange No distinction with 'Pre-processing' Balazs: 'Submitted:Hold' is the end of 'Submitted' Morris: I am seeking to understand the difference ... Consistency in the state model In the case of 'automatic resubmission', people may want to specify certain 'Hold' points Use case maybe accounting problems Etienne: Inside the submitted state, I propose to change the list of substates : Create one substate where the JobId is associated to the Job When the Job already has its JobId, the submitter is able to perform operations on the job If the JobId is associated to the Job in the middle of the 'Submitted' state, there can be a 'Submitted:Hold' substate where the JobId has already been given to the Job We have to synchronize the state model with the messages exchanged Aleksander: How can it be an activity without a JobId ? Morris: Validation steps createActivity operations should only be successful if validation is successful If validation is not successful, no Job is created, no 'Hold' is possible Etienne: Inside the 'Submitted' state, I propose to : - rename 'Submitted:Waiting' to 'Submitted:Validating' - add another substate, 'Submitted:Hold' JobId must be associated to the Job Corresponding messages have to be sent back to the Submitter Sequence : Validation, Creation of JobId, then send JobId to Submitter Morris: Still keep the 'Waiting' substate Etienne: No, rename 'Waiting' by 'Validating' Balazs: When everything is done, the job is ready to go to the next state Pre-processing, Delegated, Post-processing Etienne: In which order should the third level substates be processed? Balazs: 'Hold' blocks : blocking before going to the next state Etienne: Second level 'Hold' substate before outgoing substate ? Balazs: Yes Etienne: I understand the requirement In fact, inside the 3 first level states preproc, delegated, postproc, there already is a second level 'Hold' substate corresponding to what Balazs wishes Balazs: Different states ? Second level 'Hold', third level 'Hold' ? Morris: There might be middlewares without third level Balazs: needs preprocessing hold Morris: How is the third level flow ? Balazs: Third and second level are optional Morris: I disagree : Second level should be adopted by all submitted, preprocessing, delegating, postprocessing i.e. preprocessing hold failed recoverable not many middlewares can do this I agree with Balazs: aligning the third level of states Etienne: I had written that only the first level was mandatory, but I agree to write that the first two levels are mandatory Besides, inside 'Submitted', I will add a 'Hold' substate Morris: 'Submitted:Incoming' is missing when comparing to the state model Is 'Submitted:Incoming' a substate or is it not? Balazs: For a substate, the JobId must already exist Etienne: I agree with Balazs : The initial substate of the 'Submitted' state should have no name I will remove the 'Submitted:Incoming' label, because at this point there is no JobId yet - The Job cannot be handled by the Submitter, it is just an internal substate Morris: But we should take into account problems reported in logfiles Etienne: Such problems are implementation-specific. We should expose only states which can be used by the Submitter So I will remove the 'Submitted:Incoming' label, which is not relevant for the Submitter Is there now an agreement on what is going on in the submitted state ? Morris: 'Submitted:Validating' should be renamed as 'Submitted:Validated' because there is a JobId only if the JSDL has already been validated. Etienne. I agree, the JobId is given at 'Submitted:Validated' substate Now, I propose to return to the document Inside the 'Submitted' state, the Execution Service cannot always provide a storage location Balazs: So we are switching the subject to to the data staging issue Data staging on createActivity Pull approach Service can publish information Etienne: Please look at my mail, where I present 3 ways for message transfer Which way is the best ? Someone: We can not suppose that the Submitter implements subscription to notifications. Balazs: The Submitter needs the info about the data staging location Etienne: Without notifications, the consistent solution is that the Submitter has to repeatedly poll until the job comes to manual staging If the Execution Service returns the JobId only at manual staging when the Job is inside the 'Pre-processing:Hold' substate, I cannot guarantee the consistency This has to be discussed, because I a afraid that it is not so easy to link the 'CreateActivity' return message of to such a late state Morris. one figure not enough maybe an animation hopefully can work on this urgently needed here !!! Balazs: I strongly wish to avoid notifications Publishing the staging location is not be coupled to job state Execution Service needs time to figure out the staging location - It publishes it after some time Why should this be coupled to a job state ?? Etienne: In some cases, the Execution Service has knowledge about the storage location already inside the submitted state, in some other cases it has not this knowledge Balazs: The Job should not go to 'Pre-processing' before if the Execution Service does not know the staging location Morris: Compared to job description validation i.e. we need 2 GB, requirement How it could be still valid to have it in submitted? Etienne: A standard computing element really owns storage, so that at 'Submitted:Validated', an Execution Service may already have allocated storage to the Job But if the Execution Service is divided into sequential parts where the 'Submitted' part does not know about physical storage but only about capabilities of batch systems and storage systems, there is a problem Balazs: We should edit the text when everybody can see the document Aleksander: validation, allocation data staging after allocation id should be assigned as soon job is validated client receives information, polling or notification Morris: We will continue next time with desktop sharing tool Etienne: Important questions : - Are these 3 ways described in my email consistent ? - Which ways are preferred? Aleksander: Which mail ? Etienne: Sent on 26 October, resent on 4 December Morris: Continue next week ----------------------------------------------------- Etienne URBAH LAL, Univ Paris-Sud, IN2P3/CNRS Bat 200 91898 ORSAY France Tel: +33 1 64 46 84 87 Skype: etienne.urbah Mob: +33 6 22 30 53 27 mailto:urbah@lal.in2p3.fr ----------------------------------------------------- On Tue, 08 Dec 2009, Johannes Watzl wrote:
Dear all,
please see below the notes from today's meeting.
Best, Johannes
...