ant elder wrote:
On Tue, Feb 8, 2011 at 11:17 AM, Simon Nash <n...@apache.org> wrote:
Mike Edwards wrote:
On 08/02/2011 10:08, ant elder wrote:
On Tue, Feb 8, 2011 at 9:56 AM, Simon Laws<simonsl...@googlemail.com>
wrote:
I can see that relaxing the current spec restrictiveness would be
difficult as things would break in some circumstances, but it does
seem like there will be situations when the assembler really knows it
would be fine but even so they can't get the PBR optimization because
the impls aren't coded with the @AllowsPassByReference annotation.
Isn't that a similar situation to where the remotable annotation was
added to the interface SCDL element, and there could be a similar
allowsPassByReference attribute added to the service and reference
SCDL elements?
...ant
Hi Ant.
Can you say a bit more about the "situations when the assembler really
knows it would be fine"? Reading previous posts to this thread doesn't
convince me that that is the case.
Simon
We'll have to wait for Raymond to see if thats anything like his
situation, but there are lots of ways the assembler could know, he may
well have full access to the source but still be limited in making any
updates due to production lifecycle or control issues. This seems like
a similar situation as Java EE appservers have and I believe many
appservers have mechanisims for overriding and enabling PBR support in
their ORB or EJB modules to gain performance benefits.
...ant
Folks,
One possibility to start with would be to add a Tuscany extension
attribute which could be used by an Assembler to mark a particular service
and/or reference as "AllowsPassByReference".
This would cope with the situation where the Java components are not
marked, but the assembler knows enough about them to know that the marking
is correct.
Yours, Mike.
I'm very dubious about the wisdom of this. I bear the scars from a
previous life of a situation where a flag was added to an enterprise
application server (which shall remain nameless) to override the PBV
semantics required by the architecture in the interests of performance
optimization when the user of the product "knew" that using PBR would
be safe. Needless to say, on a number of occasions this flag was set
incorrectly and this resulted in accusations of runtime bugs that were
eventually shown after laborious debugging to have been caused by the
PBR "optimization" being enabled incorrectly.
IMO it's the implementation's responsibility to make the correct
assertions about whether it's PBR-safe. Implementations that don't
assert @AllowsPassByReference will have worse performance than those
that do support it. If a component developer cares about performance,
they need to code their implementation to play by the APBR rules and
turn on the flag to say they have done that. If the flag isn't turned
on by the implementer, I can't imagine that an assembler will have a
detailed enough knowledge of the implementation to be able to safely
provide the APBR assertion that the component developer couldn't/
wouldn't/didn't provide.
(Deep breath) Now I feel better after getting that off my chest :-)
Simon
As an attempt to justify this: we don't have any samples that use
@AllowsPassByReference, and I don't recall ever seeing many in Tuscany
or elsewhere that do, but i think all those samples i've seen probably
would have worked fine with it as they don't modify anything
afterwards. I think it would be relatively uncommon to code something
that would be unsuitable for using @AllowsPassByReference. So Joe
developer writing his code probably wont use @AllowsPassByReference
either, may not even know it exists, and any performance issues likely
wont get found until much later when running in a production system by
which time it can be very hard and time consuming to get the code
modified, re-test, QA'd and redeploying into the live system.
OK, that's a good point and we should fix the Tuscany samples so that
people reading them get the message about using APBR.
We could also get the runtime to log a low-level info message when APBR
isn't used and components are wired locally: something like "no APBR for
wire, data will be copied". That would alert people to the performance
overhead earlier in the development cycle.
I don't know the circumstances around your example of it being hard to
track down but i'll hazard a guess that that was at least partially
due to firstly that the PBR override feature was poorly document and
almost a "secret" option so developers/debuggers didn't know of its
existence and secondly that the activation and logging of the feature
was done out side of the application making it hard to see it was
being used. The Tuscany extension approach would mitigate those by
having the feature activated right in the application composite so its
hard to miss and having clear logging messages when a composite is
started where the feature is used.
Actually in its first incarnation the "force PBR override" setting was a
global flag that applied to the whole runtime, was set on by default,
and was poorly documented. (Hard to believe, I know.) After some
lobbying from myself the default was changed to off and the documentation
was improved, but people still saw it as a "go faster" flag and turned it
on when they shouldn't have done. Eventually it was made more granular
instead of a global thing and that helped some more. But the lessons
I learned were that 1) users experiencing performance problems will reach
for any magic go-faster switch and ignore all health warnings, and 2)
when it doesn't work they'll report it as a runtime bug.
Simon
That said i've no time now so wont be implementing this but if someone
else wanted to that would be ok to me.
...ant
