Dear Federico, For your transiesta runs, you will probably quickly come to shortcomings. The parallelization scheme, for now, is not really scalable for huge systems, nor excessive number of processors. As transiesta is only parallelized across energy points you will immediately see your system stall for number of processors > NE. So dont do that. :)
Kind regards Nick 2013/3/8 Federico Iori <[email protected]> > Thanks Bartek. > > The point is exactly to know about a blugene for me, since there the code > should scale on thousand of cpus instead. We have access only to this kind > of machine basically so the idea is trying to move in this new direction. > > In naycase I will keep the informations you gave! > ciao! > > ff > > > On 03/08/2013 04:08 PM, Bartek Szyja wrote: > >> Hi Federico, >> >> I have been using SIESTA on huygens in Amsterdam, which is also IBM >> machine, but it's running linux (I don't know what OS is installed on >> Fermi) and has POWER6 CPUs. It depends on the model, but generally the >> calculations scale reasonably well up to 4 nodes (128 CPUs). There is no >> significant speedup from using SMT. I also haven't done any tests with >> k-points however nor order-N - the scaling might be better. >> >> Cheers, >> Bartek >> >> >> >> >> On Fri, 2013-03-08 at 15:14 +0100, Federico Iori wrote: >> >>> Hi Siesta people! >>> >>> I am wondering if some of you has ever did some scaling test of Siesta >>> and Transiesta on Fermi, the IBM-BlueGene /Q installed in Cineca in >>> Bologna. I was considering the idea to run on that machine but I would >>> like to know if the code is really able to scale on this kind of machine. >>> >>> Do you have any idea? >>> >>> Thanks a lot >>> >>> Federico Iori. >>> >> >> >> >
