[petsc-users] Petsc performance

[Barry Smith]

  Sang,

    The most important single number by far is total ACHIEVABLE memory bandwidth http://en.wikipedia.org/wiki/Memory_bandwidth. The streams benchmark is a good way of measuring (we provide it in src/benchmarks/MPIVersion.c) the total achievable memory bandwidth when you have access to the machine.  Of course when you are buying a machine you may not have access. The total PEAK memory bandwidth can be a large amount higher than the achievable but is a way to start looking at machines. Ideally the vendor would give you access to the machine in order to run streams before you purchase.

   Barry



> On Oct 18, 2014, at 6:58 AM, Sang pham van <pvsang002 at gmail.com> wrote:
> 
> Dear everyone,
> 
> I want to purchase a high performance workstation which should have from 32 to 64 nodes, price is not so important. The workstation will be used for cfd simulation purpose.
> Would any one help me to chose one? I would very much appreciate  your suggestions. 
> 
> Thank you.
> 
> Sang
> 
> On Thu, Oct 16, 2014 at 10:01 AM, Marzouki Sondos <sondes.marzouki at gmail.com> wrote:
> Hello,
> 
> I still  hesitate whether to use Petsc library or not. And to take my decision I have some questions. Please can you help me?
> 
> Let s suppose that I have two matrixes having the same dimension, A0 and A1 but A0 is sparser than A1. Which means that A0 have less number of non zeroes.
> 
> My question is: Is Petsc less efficient( time) for dense matrix??
>  
> 
> Any suggestions or clarifications are welcome
> Thanks a lot,
> -- 
> Sondos Marzouki, Science of Computer Engineer
>  
> Tel: +216 27 66 32 24
> Address:  8300, St-Patrick Street, Suite 300, Montreal, QC, H8N 2H1, CANADA
> Skype: sondes.marzouki
> sondos at emworks.com
> EMWorks Inc
> www.emworks.com
>  
>