[petsc-users] Recommended processors according to problem size for good scaling

[TAY wee-beng]

Hi,

Is there a recommended no. of unknowns for each processor for good
scaling? Supposed I have 1 million unknowns (cells), how many processors
should I use?

I remember reading that 50k unknowns per procs is a good choice. Is that
true?

-- Thank you. Yours sincerely, TAY wee-beng

[Karl Rupp]

Hi,

I assume you have read
http://www.mcs.anl.gov/petsc/documentation/faq.html#computers

As a rule of thumb, 50k cells is a reasonable lower limit. It really
depends a lot on your interconnect, the dimensionality of your problem
(2d vs. 3d, ratio of 'surface' to 'volume' of the cell batch), etc.
Also, most likely your speed-up will level off after about 4 threads per
socket, as then the memory channels become saturated.

In either case, it's best to run your own tiny scaling study to get
actual numbers. -log_summary is really handy for that purpose :-)

Best regards,
Karli

Post by TAY wee-beng
Hi,
Is there a recommended no. of unknowns for each processor for good
scaling? Supposed I have 1 million unknowns (cells), how many processors
should I use?
I remember reading that 50k unknowns per procs is a good choice. Is that
true?
-- Thank you. Yours sincerely, TAY wee-beng