[petsc-users] KSP_SpeedUP

[Jed Brown]

On Wed, 14 Apr 2010 19:07:10 -0300, Pedro Torres <torres.pedrozpk at gmail.com> wrote:
> This may be a dummy question but in the second bullet says "its own memory
> bandwith of roughly 2 or more gigabytes", this means gigabytes/seconds

Yes, but this is a bit dated and very architecture dependent.  For
example, relative to current AMD/Intel offerings, BlueGene has much
slower cores and slightly slower memory, leading to (usually) better
scalability.  In addition to the factors of in-socket scalability, it
can be faster (if your network and MPI support it) for the network
hardware to perform the copies in send/recieve operations on the same
node (i.e. even when you have shared memory, the copies are often better
done by network hardware than by the kernel).  Arguably the only
meaningful scalability study is by choosing how to utilize each node and
then increasing the number of nodes.

Jed