Re: MPI Calculation Strategy with Lanczos Algorithm

[Jim Sims <jim.sims@nist.gov>]

Charles Bouldin wrote:

> Very nice. As for distributing the work across processors....this
> takes us into the general topic of load balancing, which we have
> avoided thus far by using homogeneous symmetric clusters and an
> algorithm that has the same computatonal time independent of energy.
> I think a simple way to do this is to work out, even if only roughly
> and empirically, how the number of iterations  scales with energy and
> then give the processors doing the higher energy part of the spectrum
> more energy points to calculate. If we can get a rough idea of the
> scaling, this slight adjustment of how we parcel out the work will be
> easy to implement. If we get fancier, the "worker" machines can each
> do a single energy and then come back and say "thank you sir, may I
> have another" to the master machine, which then assigns another
> energy point to that node.

I agree with all of this. Sounds like it is time to get fanicer and introduce
load balancing to the program as outlined by Chuck.

Jim