Hi all, I first sought about a solution like Hartmut, but SAGA is targeting Grid newbies, so JSDL might be out of the 80/20 rule scope. I am quit happy with the current set the JD provides; now would I use a Job Submission Framework (JSF) for NAREGI, the current SAGA won't work perfectly. NAREGI has extended the scope to support MPI jobs and I need to support that. I can always stick to my SAGA-style implementation for the moment. To allow power users to still make use of the SAGA adding a simple JSDL property that over rules all others and is passed directly through to the back-end would be fine. The API documentation should clearly tell that if the JSDL tag is used it supersedes all other settings. -- Best regards, Pascal Kleijer ---------------------------------------------------------------- HPC Marketing Promotion Division, NEC Corporation 1-10, Nisshin-cho, Fuchu, Tokyo, 183-8501, Japan. Tel: +81-(0)42/333.6389 Fax: +81-(0)42/333.6382 Andre Merzky wrote:
Ha, actually, that is what I expected this discussion would lead to :-P
You write:
- JSDL is an OGF Standard now and there is no way of ignoring it in SAGA
and Pascal:
The necessity to support JSDL 1.0.x is a necessity, since it is an OFG standard.
I think we need to step back a little. Yes, JSDL is an OGF standard - but so is OGSA, OGSA Security, OGSA-DAIS.
Yes, we should support JSDL, and in particular be implementable on top of JSDL. We also should support OGSA, and in particular be implementable on top of OGSA. Same with DAIS.
That does NOT mean that we need an API to write JSDL documents, nor an API to write OGSA Base NOtification documents, nor an API to write DAIS Collection Access XML docs.
SAGA is NOT a general purpose API for existing OGF standards. It is _designed_ to be incomplete.
Quoting [Hartmut Kaiser] (Feb 19 2007):
Hi all,
After thinking about this issue a bit more I would like to make a even more 'radical' suggestion.
Let's recap: - JSDL is an OGF Standard now and there is no way of ignoring it in SAGA - the current SAGA JD is more constrained in terms of scope if compared to JSDL - there will be a lot of middleware supporting JSDL in the future - there wll be a lot of tooling support to create JSDL in the future
A viable solution to this in SAGA could be (duck):
- redo the SAGA job_service and job interfaces so that these work solely based on JSDL strings
Further, the saga job description has a (small) number of keys which are not supported by JSDL, and will not be in the near future. Those are scheduling related: Queue, JobStartTime etc.
- supply a SAGA JD component providing a _simple_ interface for JSDL construction:
The JSDL schema has repeatable, hierarchical elements, with 0..n occurences of sections etc etc. It can include XML snippets from arbitrary name spaces. Just one example from many in the JSDL spec:
"7.1 Attribute Extension
Every JSDL element allows for additional attributes, as many as needed, provided these attributes have a namespace other than the normative namespaces defined by JSDL."
There is a good API to write XML documents (DOM), and there are many tools to do so. Why should SAGA re-do that? None of these APIs/tools is simple - they cannot be, by design.
Please check the JSDL spec: JSDL is simple for middleware level - but it is too complex (and I mean really) on the level SAGA is targeting at.
So, providing an API to create JSDL is, IMHO, out of scope.
job_description::construct(string jsdl); job_description::as_jsdl();
job_service::create_job(string jsdl); job::migrate(..., string jsdl, ...);
where we clearly could stick to JSDL V1 for this conversion and the allowed attribute set on our job descriptions.
I honestly don't see the reason why we should deviate from our design principles (simple API, orient on use cases, don't do more than users ask for) just because JSDL exists.
Its nice it exists, we support it (it can and should be used by SAGA _implementations_). But it has no more appeal on API level than has OGSA Base Notification, for example.
So, my proposal is:
- stick to the JD we have in SAGA (it works nicely with JSDL)
- support upcoming JSDL extensions in due time, either by new keys on the job description (JSDL scheduling), or programmatically (i.e. JSDL parameter sweeps)
- if there are use cases asking for it (only then), provide a jsdl_to_jd converter
Well, I realise I am repeating myself - not a good sign. We probably should give the topic anoter week of rants, and then do a straw poll on the mailing list, or continue to fight on a phone call... :)
Cheers, Andre.
Regards Hartmut
-----Original Message----- From: saga-rg-bounces@ogf.org [mailto:saga-rg-bounces@ogf.org] On Behalf Of Pascal Kleijer Sent: Sunday, February 18, 2007 8:51 PM To: SAGA RG Subject: Re: [SAGA-RG] Fwd (andre@merzky.net): Re: JSDL - SAGA
Dear all,
I have given me the week-end to get some reflection on the problem. As mentioned it is not a trivial issue. If things goes out of hand, we might as well rename the project CAGA of not go gaga :P
The necessity to support JSDL 1.0.x is a necessity, since it is an OFG standard. I do not think JSDL 1.0 is overly complex right now. The Job Description should be made such that passing from one form to another is automatic (JD -> JSDL). The support of a property key for JSDL isn't bad; if used, all other properties can be potentially ignored. The JD would not have to parse it. The current specification is simple enough to not violate the "S" rule of SAGA.
The problem we have is with all the up coming extensions. In my case, I clearly violated the SAGA API rules in order to make the system run with NAREGI. I tweaked the problem to solve my case. In the 80/20 rule, it probably means that the SAGA job submission won't be usable. I will have to write code that does not deal with SAGA and goes directly to the underwear. Now I am speaking as a Grid "expert" and not as a Grid newbie (which is the SAGA user base). If you want to use complex Grid programming and do complex things, SAGA might not be the right solution.
The proposal made by Dan is not elegant in my opinion because it will force developer to know the underlying technology they want to use; it will also clutter the JD. The proposal from Thilo might work for OO languages because you can hide a lot behind patterns, but it won't mean a specific implementation will support such scheme.
The passing through solution seems to be mildly seen by the some GAT people here. I would say no, if the error handling is explicit enough to return a message telling that this JSDL cannot be interpreted by the underlying system. I will not see this as a bad solution either.
--
Best regards, Pascal Kleijer
---------------------------------------------------------------- HPC Marketing Promotion Division, NEC Corporation 1-10, Nisshin-cho, Fuchu, Tokyo, 183-8501, Japan. Tel: +81-(0)42/333.6389 Fax: +81-(0)42/333.6382
Another related idea is for SAGA to have one class for job descriptions with a quite loose ability to add "hints", such as in MPI I/O. The hints would not be required and might not be used, but in some cases would be used, depending on what SAGA is sitting on top of. Dan
Thilo Kielmann wrote (on 2/16/07 5:54 AM):
Folks,
I think we are having conflicting goals here. (Technical goals, not personal ones ;-)
On one hand, we have the "S for simplicity" in SAGA, and we must keep it. On the other hand, we have the necessity to support JSDL, future JSDL extensions, or any other types of job decscriptions that
(And still might to be invented.)
Assume, JSDL++ (whatever it will look like in near future) will become a widely adopted standard. (Or anything else, doesn't matter in the following.) Then, SAGA implementations will have to use JSDL++, and to form JSDL++ from SAGA job descriptions. Simultaneously, users are likely to use JSDL++ themselves, and may wnat to use JSDL++ to express their resource needs. At this point in time, SAGA will sit in the middle, and it may be very clumsy to first translate from JSDL++ to a SAGA job description, and then back to JSDL++ somewhere "down under" in the implementation.
For the very purpose of SAGA as a universal and simple grid API, it has to: - be independent of job description standards (mostly simpler than them) - support job description standards
My suggestion:
SAGA should have one class of job descriptions, and the
to create subclasses for more specific job descriptions (like JSDL). Such subclasses could be defined as separate extension documents (just like gridcpr). (Or was it "resource descriptions"???)
With this approach, users could still write programs that are independent of the underlying job submission machinery, having a simplified view on jobs and resource attributes etc. etc. At the same time, subclassed job/resource descriptions could be "passed through" transparently from the API to the implementeation, without being converted back and forth, a process that is very likely to loose some important details.
Is this a route to go?
Thilo
On Tue, Feb 13, 2007 at 06:28:58PM +0100, 'Andre Merzky' wrote:
From: 'Andre Merzky' <andre@merzky.net> <mailto:andre@merzky.net> To: SAGA RG <saga-rg@ogf.org> <mailto:saga-rg@ogf.org> Subject: [SAGA-RG] Fwd (andre@merzky.net <mailto:andre@merzky.net>): Re: JSDL - SAGA
Hi group(s),
a couple of us had a recent discussion (f2f and email) about JSDL and SAGA. The question is: should we support JSDL fully on API level? E.g., should we allow the application programmer to specify/use JSDL documents for job creation?
The reasons for doing that are compelling: JSDL is one of
acknowledged standards in OGF, and the number of backends supporting JSDL is rapidly increasing it seems. Supporting JSDL directly would allow to interface with other tools using JSDL, and would allow to reuse JSDL documents where
already available.
Well, I have however some problems with that approach, which are outlined in the cited email below.
Do you guys have any other thoughts, and what solution would you prefer?
Cheers, Andre.
----- Forwarded message from 'Andre Merzky' <andre@merzky.net> <mailto:andre@merzky.net> -----
> From: 'Andre Merzky' <andre@merzky.net> <mailto:andre@merzky.net> > To: Shantenu Jha <sjha@cct.lsu.edu> <mailto:sjha@cct.lsu.edu> > Cc: Hartmut Kaiser <hkaiser@cct.lsu.edu> <mailto:hkaiser@cct.lsu.edu>, > 'Thilo Kielmann' <kielmann@cs.vu.nl> <mailto:kielmann@cs.vu.nl>, > 'Andre Merzky' <andre@merzky.net> <mailto:andre@merzky.net> > Subject: Re: JSDL - SAGA > > Hartmut and I discussed that somewhat last week. So he knows I am > not wholehartedly for that option. SAGA is supposed to abstract > the low level details, not to expose them. > > JSDL is going to define a number of extensions now. Some of these > extensions are very useful for us, others not. > Mostly, they will be more complex than JSDL itself. > > Are we going to support the extensions? Which? All/some? > How to select? What error do we report on unsupported extensions? > Do we mandate that extensions are supported by the backends? > Which? > > Even w/o extensions: is the job description updated after an JSDL > attrib is set? What about those attribs which are not JSDL keys? > Assume an implementation which implements the existing SAGA job > description keys: MUST it support complete JSDL now? What error > whould it report? > > These are probably all solvable problems, and I do agree
Daniel S. Katz wrote: people want to use. possibility the most these are that there
> are advantages, i.e. the re-use of existing JSDL documents. > Anyway, IMHO we should be careful, consider if we really have > enough use cases etc. Also, a free function > jsdl_to_job_description may do the trick, w/o complicating the job > package itself. > > Cheers, Andre. > > Another point I'd like to raise is: if HPC profile bekomes a widely accepted OGF standard, do we support it directly, too? Or OGSA-Workflow? Where to stop, and where is the 'S' in that approach?
Andre.
> Quoting [Shantenu Jha] > >> What little I know, I think so to. >> >> >> Quoting [Hartmut Kaiser] >> >>> Agree 100% >>> The easiest way is probably just to add a attribute in the job >>> description taking the whole JSDL. >>> >>> >> Quoting [Thilo Kielmann] >> >>> Yes! >>> >>> >>> Quoting [Shantenu Jha] >>> >>>> Shouldn't we ensure that SAGA consumes JSDL w/o any >>>> problem/changes? >>>> >>>> -- "So much time, so little to do..." -- Garfield -- saga-rg mailing list saga-rg@ogf.org <mailto:saga-rg@ogf.org> http://www.ogf.org/mailman/listinfo/saga-rg
----------------------------------------------------------------------
--
-- saga-rg mailing list saga-rg@ogf.org http://www.ogf.org/mailman/listinfo/saga-rg