[saga-rg] Fwd (andre at merzky.net): Fwd (s.jha at ucl.ac.uk): Re: OGSA is aligned...
Andre Merzky
andre at merzky.net
Tue Feb 14 08:39:20 CST 2006
Hi SAGA group,
We finally managed to sit together for an hour and discuss
the relation between SAGA and OGSA in some detail. Thanks
to the OGSA folks for making the time!
This email tries to capture the points and results of that
discussion. It is rather detailed for two reasons:
- we want the OGSA folks to check, and 'officially' agree
to these points (they did by now)
- we want to use these points as input to our middleware
alignment document
Observations:
- There are a number of paradigms which span OGSA and are
potentially apparent at the application level in some
use cases (TODO: recheck OGSA use case document)
Amongst these paradigms are, most notably,
- notification / events
- asynchroneous / decoupled operations
- state management
- life-time management
-- notification / events:
SAGA thinks we are fine here, as we have
A) a generic event and monitoring model, and
B) most events can be captured and reacted on
on implmentation level.
-- asynchroneous / decoupled operations:
SAGA thinks we are fine here as well, as
A) the API is asynchroneous as well, and
B) asynchroneous middleware calls can be made
synchroneous with a blocking wait (that has drawbacks
though, but we can live with those in our use cases).
-- state management:
SAGA thinks we are fine here as well, as
A) essential state information (file pointer, job state,
stream state) are exposed on API level already
B) other state information can be kept in the API
implementation on client side, or on intermediate
stateful services, depending on the system the API is
implemented on
-- life-time management:
This was the most interesting point I think. The use
cases are:
- an application runs for 3 months, and expects a file
pointer to be available for that time
- a remote file op implies a log file, but a network
failure prevents the lock from beeing removed. W/o
life-time management support, that lock lingers
(can't distinguish between a network drop and
application being busy).
Our discussion came to the conclusion, that:
- sensible life-time defaults in the implementation,
or out of bound, can probably catch most problems
- a generic application life-time attribute to a
session handle should catch most of the more
difficult cases (long running applications etc).
- per 80:20 rule, that should be sufficient for SAGA
until more specific use cases are known.
- more detailed use cases might motivate life-time
attributes (contexts) for SAGA object.
Summary:
At the end, we agreed with Steven's remark that OGSA and
other middleware properties should ideally NOT pop up on
application level at all. Having said that, there might
be use cases where an exposure of life-time, state,
discovery etc. might be beneficial for application.
Also it might not be obvious what is a middleware
property ("one middleware's property might be another
applications")
( Comment from Steven:
If stuff has to pop up it should do so in an infrastructure
neutral way
- i.e. it is a lifetime request (not a setting of a WSRF
resource property).
Also if requesting a specific lifetime there should be an
option for the layer to throw a 'NotSupportedExcepetion"
)
SAGA should wait for more detailed use cases before
pursuing that topic any further, for now we see NO point
of conflict between SAGA and OGSA based middleware.
Life-time management might be a simple and local addition
to session properties, and should be considered for
inclusion into the API, after cross checking with the
available OGSA use cases.
Cheers, Andre et.al.
--
"So much time, so little to do..." -- Garfield
More information about the saga-rg
mailing list