[petsc-users] PetscLayout and GetRange

[Weston Lowrie]

Jed,  Thank you for the good advice!

On Wed, Jan 16, 2013 at 9:07 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote:

> You can't change the size of the vector after it has been created. You can
> reuse the memory if you manage it separately (by calling
> VecCreateMPIWithArray()). Don't try to "trick" the Vec.
>
> I'm pretty sure your performance anxiety is premature. Just call
> VecDestroy() and create a new Vec for the next iteration.
>
> Note: it's inappropriate to think of a Vec as a dynamic array that you're
> using to accumulate an unknown amount of data. The Vec is meant to be used
> for linear algebra (including collective semantics). If you're building
> something dynamically, build it with a dynamic data structure and then put
> into a Vec once the structure is built.
>
>
> On Wed, Jan 16, 2013 at 8:01 AM, Weston Lowrie <wlowrie at uw.edu> wrote:
>
>> On a related note:
>> When using VecGetArray(), it uses the info from VecSetSizes(), which is
>> set on Vec creation.  Can this be updated after the Vec has been used?
>>
>> My issue is that even if I know the proper local size that I want, when I
>> call VecGetArray() it's going to use the local size determined on Vec
>> creation.
>>
>> To avoid this problem I could use VecSetValues() for the range I am
>> interested in, rather then using the pointer generated from VecGetArray()?
>>
>> Wes
>>
>> On Tue, Jan 15, 2013 at 5:31 PM, Weston Lowrie <wlowrie at uw.edu> wrote:
>>
>>> I don't think there is a inefficiency here.  It will be just one extra
>>> empty vector since all the other "real" ones are identical size.  Sounds
>>> like a good strategy to me.
>>>
>>> I'm not quite at the stage where I can profile yet.  I will send if I
>>> can get it to a point where it might be significant.
>>>
>>> Thanks for the help,
>>> Wes
>>>
>>>
>>>
>>> On Tue, Jan 15, 2013 at 5:26 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
>>>
>>>> On Tue, Jan 15, 2013 at 4:21 PM, Weston Lowrie <
>>>> wlowrie at u.washington.edu> wrote:
>>>>
>>>>> That's interesting.  If I understand you correctly, I would create a
>>>>> vector of the size I want specifically for calculating the ownership range,
>>>>> then use that on the real vectors.  Sounds like that would work.
>>>>>
>>>>> In my case, with many vectors, it does not make sense to copy them to
>>>>> a resized vector every time I want them to grow leading to many creates and
>>>>> destroys.
>>>>>
>>>>
>>>> You can't dynamically resize vectors like that, and the global offsets
>>>> change when you resize.
>>>>
>>>> Please profile before jumping to the conclusion that there is some
>>>> terrible inefficiency here. Unless all your loop does is create Vecs of
>>>> different sizes, chances are that the VecCreate is insignificant. If you
>>>> have a profile in which it's a big deal, please send the profile and
>>>> explain what you are doing and why.
>>>>
>>>>
>>>>>
>>>>> Wes
>>>>>
>>>>>
>>>>> On Tue, Jan 15, 2013 at 5:15 PM, Jed Brown <jedbrown at mcs.anl.gov>wrote:
>>>>>
>>>>>> VecGetOwnershipRange()
>>>>>>
>>>>>> You can use VecCreateMPIWithArray() using your own array preallocated
>>>>>> to be as long as you want. If you profile, you'll probably find this is not
>>>>>> a meaningful optimization.
>>>>>>
>>>>>>
>>>>>> On Tue, Jan 15, 2013 at 4:12 PM, Weston Lowrie <wlowrie at uw.edu>wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> I have a problem where I want to grow the size of a Vec (or Mat)
>>>>>>> many times during program execution.  I think for efficiency purposes I
>>>>>>> would just want to allocate a maximum size, and then only use the portion
>>>>>>> that I need.  In the vector case, it is rather simple, just use the
>>>>>>> beginning of the vector, and add values to the end.
>>>>>>>
>>>>>>> This leads to me to the problem of processor ownership ranges.  From
>>>>>>> a previous email I noticed one could use the PetscLayout object and keep
>>>>>>> adjusting it as the useful part of the vector grows.  Does this sound like
>>>>>>> a good approach?
>>>>>>>
>>>>>>> I noticed the PetscLayout is not available in Fortran bindings.  Any
>>>>>>> workarounds for this?  I suppose I can just manually calculate the
>>>>>>> processor ranges?
>>>>>>>
>>>>>>> Thanks for the help,
>>>>>>> Wes
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20130116/4d60dfbc/attachment-0001.html>