On Wed, 13 Jun 2007, Tim Jenness wrote:
> On Wed, 13 Jun 2007, Brad Cavanagh wrote:
>
>> ==13509== Open file descriptor 29:
>> /home/bradc/data/oracdr/reduced/uist/20070608/adam_13501/GLOBAL.sdf
>
> We got this to come up because we disabled the hds exit handler (Which
> usually closes all these) on the basis that valgrind will never report them
> if the exit handler runs but what's relevant to oracdr is the state
> following an obey not the state following the A-task shutdown.
>
> It would seem that if, for performance reasons, the parameter files are not
> closed after an action then if two monoliths modify GLOBAL.sdf that the first
> monolith will start to get confued. The private parameter files don't cause a
> problem because no-one else is writing to them.
This behaviour has changed.
Using the 2005 release I can see that various parts of ICL CCDPACK never
leave the GLOBAL file open (if this was the case I'm sure we'd have come
across this problem many years ago). It is briefly opened to write out the
new GLOBAL parameters, but then immediately closed. The monolith parameter
file always remains open, as expected for efficient access (and we all
know these cannot be shared).
Now in hokulei, the GLOBAL file is seen to be open after accessing some
parameters so remains open after the write update on task run down.
(BTW Tim, reads like this operate just using a HDS path, like:
$ADAM_USER/adam/GLOBAL.parameter, some of which is coded into the IFC
file, so it's a bit invisible to grepping, this is because associations
don't have to be global),
I'll have a poke around and see if an obvious cause leaps out.
Peter.
|