[vtkusers] broadcasting an vtkUnstructuredGrid

Martin Baumann mailsgetlost at web.de
Wed Aug 2 11:35:09 EDT 2006


Kevin H. Hobbs wrote:
> On Wed, 2006-08-02 at 15:30 +0200, Martin Baumann wrote:
>   
>> Hi,
>>
>>     
>>> Can you preposition the whole file or pieces of the file onto local
>>> storage on each node?
>>>       
>> copying the files on every node is no solution in my case as the files 
>> can be very large.
>> And not all of the nodes have enough memory capacity to hold the files.
>>     
>
> You mean storage right?
>   
Yes, storage.
> OK, the clusters I use have local disks on each node mounted as /tmp
> just for this. It's really nice not to have to deal with file
> contention, NFS and so on.
>
>   
>>> Do you rally want the same data on all the nodes? 
>>>       
>> I am programming a parallel particle tracer. A particle could move to 
>> any place in the
>> grid in a calculation step. So each node should have access to all data.
>>     
>
> So the nodes have enough memory to hold the whole data set, but not
> enough storage?
>   
The data set is timedependent. That means for each timestep there is an 
extra file.
Every node can hold the data of one time step in memory. But as there 
can be a lot
of time steps, the sum of all files is too large to copy to each nodes 
local hard disc.
> They can't ask for the pieces they need? Are there so many particles
> that each process would have a particle in most pieces? 
There can be a lot of particles. And they can be anywhere within the grid.
I read that finding positions and neighbours in unstructured grids is slow.
So is it possible (and fast) to "order" a certain piece of the grid that 
contains a given
point and all points that are closer than lets say 10 to that point?
> Would it make sense to break the data up and pass particles instead of data?
>   
That sounds interesting. Hmm... I am thinking about that...


Regards, M.B.



More information about the vtkusers mailing list