11:39 a.m.
Hi,
I'm currently facing a problem running a function inside an MPI block,
when lists inside a bundle are used and accessed. (The context is the
johansensmall.gfn package, where the combination of exogenous terms and
running it under MPI fails.)
I'm not finished with the bughunting and want to try to produce a
minimal example, but I wanted to ask here already whether my suspicion
is plausible:
Because the MPI block doesn't inherit the dataset context of the calling
environment, for the bootstrap simulation I create a new dataset
environment via 'nulldata <...> --preserve'. In there I create the
simulated series. This works in principle. Am I right in assuming that
gretl is having problems when I use a previously packaged list like
mybundle.xlist and try to stuff it into the new dataset environment? I
have double-checked that $nobs matches the length of the variables in
the list, but is that enough?
thanks
sven
11:49 a.m.
On Wed, 12 Aug 2020, Sven Schreiber wrote:
Hi,
I'm currently facing a problem running a function inside an MPI block,
when lists inside a bundle are used and accessed. (The context is the
johansensmall.gfn package, where the combination of exogenous terms and
running it under MPI fails.)
I'm not finished with the bughunting and want to try to produce a
minimal example, but I wanted to ask here already whether my suspicion
is plausible:
Because the MPI block doesn't inherit the dataset context of the calling
environment, for the bootstrap simulation I create a new dataset
environment via 'nulldata <...> --preserve'. In there I create the
simulated series. This works in principle. Am I right in assuming that
gretl is having problems when I use a previously packaged list like
mybundle.xlist and try to stuff it into the new dataset environment? I
have double-checked that $nobs matches the length of the variables in
the list, but is that enough?
In passing a list via a bundle you're just passing a vector of
series ID numbers. This will be usable qua list in the target
function only if, in the dataset context in the target, there exist
series answering to those IDs.
Allin
5:37 p.m.
Am 12.08.2020 um 17:49 schrieb Allin Cottrell:
On Wed, 12 Aug 2020, Sven Schreiber wrote:
> I'm currently facing a problem running a function inside an
MPI block,
> when lists inside a bundle are used and accessed. (The context is the
> johansensmall.gfn package, where the combination of exogenous terms and
> running it under MPI fails.)
>
> I'm not finished with the bughunting and want to try to produce a
> minimal example, but I wanted to ask here already whether my suspicion
> is plausible:
>
> Because the MPI block doesn't inherit the dataset context of the calling
> environment, for the bootstrap simulation I create a new dataset
> environment via 'nulldata <...> --preserve'. In there I create the
> simulated series. This works in principle. Am I right in assuming that
> gretl is having problems when I use a previously packaged list like
> mybundle.xlist and try to stuff it into the new dataset environment? I
> have double-checked that $nobs matches the length of the variables in
> the list, but is that enough?
In passing a list via a bundle you're just passing a vector of series ID
numbers. This will be usable qua list in the target function only if, in
the dataset context in the target, there exist series answering to those
IDs.
Hm, right. I guess I need to use "--send-data", actually I don't
remember why I didn't do that right away.
But if I were to do only: "--send-data=mylist", are the ID numbers of
mylist in the original dataset somehow translated to the MPI context
where only the subset is present?
(What I mean is: suppose I have series x, y, z. And mylist=deflist(x,z),
so consisting of IDs 1 and 3. Inside the MPI block only x and z then
exist, presumably with local IDs 1 and 2. If I reference mylist in
there, will it look for IDs 1 and 3 or for 1 and 2? And of course
what's worse, what if mylist is wrapped in a bundle?)
thanks
sven
5:13 a.m.
Am 12.08.2020 um 23:37 schrieb Sven Schreiber:
Am 12.08.2020 um 17:49 schrieb Allin Cottrell:
> In passing a list via a bundle you're just passing a vector
of series ID
> numbers. This will be usable qua list in the target function only if, in
> the dataset context in the target, there exist series answering to those
> IDs.
Hm, right. I guess I need to use "--send-data", actually I don't
remember why I didn't do that right away.
But if I were to do only: "--send-data=mylist", are the ID numbers of
mylist in the original dataset somehow translated to the MPI context
where only the subset is present?
Answering myself after some testing: the ID numbers inside the MPI block
again start from 1 and are simply consecutive. Series need to be
referenced by name. OK.
The 'mylist' object itself does not exist inside the MPI block, so
syntax like mylist[1] is impossible. Fair enough.
Getting to the point of lists in bundles, the following provokes an
error "The list b.L does not exist" or sth like that:
<hansl>
nulldata 10
series x = 0
series y = normal()
list L = y # So ID number 2
bundle b = null
list b.L = L
mpi --send-data=b.L
if $mpirank == 0
varlist
endif
end mpi
</hansl>
Given that inside the mpi block b.L would not exist (just like mylist
above didn't exist), I think it would be OK to allow this kind of
construct. The list b.L would define some variables to be passed into
MPI just like the list mylist did. Everything is in the context of the
current dataset anyway.
thanks
sven
1557
days inactive
1558
days old
3 comments
2 participants
participants (2)
-
Allin Cottrell -
Sven Schreiber