> On Jun 20, 2016, at 12:20 PM, marco restelli <[email protected]> wrote: > > 2016-06-20 19:01 GMT+0200, Barry Smith <[email protected]>: >> >>> On Jun 20, 2016, at 10:13 AM, marco restelli <[email protected]> wrote: >>> >>> Dear all, >>> while assembling a matrix in PETsc we have the following pattern: >>> >>> >>> do i=1,number_of_chunks >>> >>> ! generate a chunk of matrix entries >>> call compute_ith_chunk( i_idx , j_idx , values ) >> >> This just generates a "random" bunch of nonzero entries with no >> structure? What discretization method are you using? > > The matrix comes from the discretization of an integral equation and > the structure depends on the convolution kernel. Sure it is not > random, but it has a nontrivial pattern which I would prefer to deal > with in another part of the code. So, for the matrix assembly my > present ansatz is treating the entries as a collection of known but > arbitrary values - I understand that I am probably giving up on some > optimization, but this might be worth if it helps modularizing the > logic of the code.
If this is coming from a integral equation where normally the matrix would be dense (?) I assume you have some mechanism to "throw away" small values. Likely even then you have many nonzeros per row? Hundreds, thousands? When you have many nonzeros per row and and insert the individual entries "randomly" the performance will be poor because MatSetValues() has to do a binary search on each entry based on the column index to find the correct location to put the value in. Preallocation is crucial but even with preallocation the time for setting all the values if you don't take advantage of some structure will be really slow. Barry > >>> ! insert those entries >>> do j=1,size_of_the_ith_chunk >>> call MatSetValue( mat , i_idx(j),j_idx(j),value(j) , add_values ) >>> enddo >>> >>> enddo >>> >>> >>> The problem is that inserting the elements with MatSetValue seems to >>> have a significant overhead due to memory allocations and >>> deallocations. >> >> >>> >>> Is there a way to speed-up this process, preallocating memory? >>> >>> Notice that we know the number of elements that we have to insert for >>> each chunk, but we don't know to which extent they overlap, i.e. we do >>> not know how many nonzero entries will result in the final matrix. >> >> Take a look at MatPreallocateInitialize() and its other routines it is >> designed to make it easy to get a good preallocation. > > Ah, this looks promising! I will experiment with this. > > Thank you, > Marco
