ok, so considering performance on MIC
can the library MAGMA be used as an alternate to Viennacl for PETSc or FEniCS? http://www.nics.tennessee.edu/files/pdf/hpcss/04_03_LinearAlgebraPar.pdf (from slide 37 onwards) MAGMA seems to have sparse version which i think is doing all that any sparse non linear solver can do.. MAGMA-sparse.. will this be helpful in using with MIC *with regards..* *M. Sai Rajeswar* *M-tech Computer Technology* *IIT Delhi----------------------------------Cogito Ergo Sum---------* On Wed, Jun 11, 2014 at 8:34 PM, Karl Rupp <[email protected]> wrote: > Hi, > > > Im a masters student from Indian Institute of Technology delhi. Im > >> working on PETSc.. for performance, which is my area of interest. Can >> you please help me in knowing 'How to run PETSc on MIC' . That would >> be of great help to me. >> > > my experience is that 'performance' and 'MIC' for bandwidth-limited > operations don't go together. Regardless, you can use ViennaCL by building > via > --download-viennacl > for using the MIC via OpenCL, but you are usually much better off with a > proper multi-socket CPU node. > > Feel free to have a look at my recent slides from the Intl. OpenCL > Workshop here: > http://iwocl.org/wp-content/uploads/iwocl-2014-tech- > presentation-Karl-Rupp.pdf > PDF page 32 shows that in the OpenCL case one achieves only up to 20% of > peak bandwidth for 1900 different kernel configurations even for simple > kernels such as vector copy, vector addition, dot products, or dense > matrix-vector products. With some tricks one can probably get 30%, but > that's it. > > PETSc does not provide any 'native' OpenMP execution on MIC for similar > reasons. > > Best regards, > Karli > >
