[petsc-users] Improve in speed/efficiency from partitioning in 1 direction to 3 directions

[Jed Brown]

The preconditioner also matters. If you are using a one-level
preconditioner, the 3D partition can cut the iteration count by a factor of
approximately the cube root of the number of processes relative to the 1D
partition.

There are many factors that can limit the performance of your code. For
performance questions, always send the output of -log_summary.

On Tue, Sep 4, 2012 at 11:55 PM, TAY wee-beng <zonexo at gmail.com> wrote:

> On 4/9/2012 1:54 PM, Aron Ahmadia wrote:
>
>> There are a lot of other factors at play here, including how much time
>> you are spending working on the code, how far you're trying to scale
>> the algorithm and how much communication dominates your problem for
>> the architectures and algorithms you are using.  Unless you are trying
>> to take this code to thousands of processors I probably wouldn't worry
>> about it.
>>
>> A
>>
>
> Ok thanks for the suggestions Aron. I'll look at others parts for
> optimization.
>
>
>> On Tue, Sep 4, 2012 at 11:17 AM, TAY wee-beng <zonexo at gmail.com> wrote:
>>
>>> On 4/9/2012 11:11 AM, Aron Ahmadia wrote:
>>>
>>>> This doesn't strike me as a particularly large problem.  I'm not sure
>>>> it's worth doing unless you are going to be looking at more unknowns
>>>> in the future.
>>>>
>>>> A
>>>>
>>> Hi Aron,
>>>
>>> It will increase to 500x500x1200 or more. In that case, do you think it's
>>> worth it?
>>>
>>> Thanks!
>>>
>>>  On Tue, Sep 4, 2012 at 10:08 AM, TAY wee-beng <zonexo at gmail.com> wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> My Fortran CFD code is currently partitioned in the z direction. Total
>>>>> grid
>>>>> size is around is 153x248x620. Hence depending on the no. of procs,
>>>>> the z
>>>>> direction 620 is partitioned. The grid size changes but the ratio is
>>>>> around
>>>>> there.
>>>>>
>>>>> Parititoning in 3 directions was initially too complex for me. However,
>>>>> it
>>>>> seems to be much simplified with the use of DM. However, there's still
>>>>> a
>>>>> lot
>>>>> of work to be done to make it working.
>>>>>
>>>>> I'm wondering how much improvement in speed/efficiency will I get, if I
>>>>> partition from 1 direction to 3 directions. Is it worth the effort?
>>>>>
>>>>> --
>>>>> Yours sincerely,
>>>>>
>>>>> TAY wee-beng
>>>>>
>>>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120905/4797b526/attachment.html>