Dear Sylvain, Quoting [Sylvain Reynaud] (Jun 04 2009):
SAGA Job Management: =================== * CPUArchitecture and OperatingSystemType should be scalar (instead of vector) attributes for consistency with JSDL specification.
Good find! These are indeed errors in the spec, and should be scalar attributes. Will fix.
* page 166 says: "Attributes marked as 'not supported by JSDL' might disappear in future versions of the SAGA API". - Queue: this attribute makes the job description dependent on the targeted execution site, this information should be put in the URL instead.
Interesting point. The problem I see is that its hard to define a standard way on *how* to encode it in the URL, as each URL component (host, path, query, ...) may already be interpreted by the backend. For example, a globus job manager URL may well look like https://some.remote.host:9443/wsrf/services/ManagedExecutableJobService?65e5... Where would you put the queue?
- JobStartTime and JobContact: IMHO, these attributes are not in the 80% of the 80/20 rule, they are not supported by most middlewares, and they can be implemented by the user application.
Those have been included because DRMAA has these, and the DRMAA folx found them dead useful. JobStartTime I expect to get defined in JSDL when they get started on scheduling (which is not in the near future AFAIK). In SAGA, it probably should be move to the resource management package, when that emerges. JobContact is a tricky one: it seems dead useful to be able to specify a custom name to a job, and to identify the job thus in an easy way, but (a) I have yet to see a backend which supports it, (b) SAGA does not provide a way to actually *use* this name, e.g. as jobid, and (c) such a mapping can trivially be implemented on application level. I am with you here. Anyway, I think it is kind of dangerous to remove stuff from the spec simply because we don't see a use at the moment. Not sure how representative we are...
- Interactive: although this attribute may not be in the 80% of the 80/20 rule, I think it is usefull and should be kept.
Many of our original use cases had a visualization background, where interactive jobs are more likely to be useful than in other fields...
* job are missing a state "QUEUED" in order to enable timeout on jobs queued for a long time, synchronization with job start, or just because many users want to know if their jobs are running or queued. This can not be done with the job.state_detail attribute because its content is not uniform. As it was said at OGF23, some job services don't have any queue, but I think this is not the usual case.
I don't see why job.state_detail can't help you here: your SAGA implementation should always be able to bap the respective backend queue state to something like 'jsaga:queued'. The spec says that job_detail SHOULD be formatted as 'model:state', but you seem to have a very valid use case for defining your own backend model for jsaga. Reason why I am hesitant to agree with you, despite of the validity of your use case, is: while defining the spec and the state model, we tried only to expose those state at the top level which are the result of a SAGA API call, or can be left with a SAGA API call. There is not saga.job.queue() or saga::job::unqueue, so moving Queued on the top state model would break that rule of thumb.
SAGA Name Spaces: ================ * add a flag to disable checking existence of entry in constructor and open methods, because the cost for this check is not negligible with some protocols (then subsequent method calls on this object may throw an IncorrectState exception if the entry does not exist).
Makes sense. We could also overload 'Exclusive', which, at the moment, is only evaluated if 'Create' is specified. It has the same semantic meaning so (inversed): if 'Exclusive' is not specified on 'Create', an existing file is ignored. Would it make sense to allow Exclusive to be evaluated on all c'tors and open calls?
SAGA Session: ============ * An exception should be thrown when several contexts of the session can be used for a specific method call. Else, some files may be created with unexpected owner, some jobs may fail at the end of execution because of unexpected permissions, some accounts may be locked because of too many failed connection attempts.
Uh, this would badly break our use cases! In our implementation, you may very well copy a file instance with http, but read it with ftp, and delete it with nfs! And we need the respective contexts for these backends on the same session... Yes, results are not completely predictable if multiple backends can be used, but the user has always the option to create a specific session with exactly one context attached. So, that's a tough one.
* AFAIK, existing SAGA implementations support many technologies and may be used with several grids. Default session is very convenient, but it is not usable when several grids can be accessed, because a single default session for all is not enough. Argument of session constructor could be an identifier (e.g. a grid identifier) instead of a boolean.
Again, you can create a context for a Grid type, and attache just that one context to a session. Isn't that what you need? I may be missing something here... Thanks, Andre. -- Nothing is ever easy.