Picongpu: Mismatch -d and --gridDist

Created on 5 Feb 2019  路  7Comments  路  Source: ComputationalRadiationPhysics/picongpu

There appears to be no check whether the number of GPUs given via -d matches the number of GPUs given in --gridDist.

The simulation still runs "fine" but has some empty cells in the front.

Could this setup be checked during start up?

affects latest release bug user input

All 7 comments

Please give a example for a broken co figuration.

We currently check --gridDist with -g to match (in sum) and expand the 2nd/third dimension if skipped in --gridDist to be of equal spacing.

The example below caused probably the issue described in https://github.com/ComputationalRadiationPhysics/projects-electrons/issues/132:

-d 663
-g 768 2304 768
--gridDist '160{2}, 64{2}, 160{2}' '288{8}' '320{1}, 128{1}, 320{1}'

There are only 6 devices in y allocated via -g but the distribution assumes 8.

Hm, that should be checked indeed! What is the actual grid size it will now build? 2304 or 1728?

Quick notes:

  • there should be no spaces after the commas ,
  • there should be no spaces inside the { } (just here)
  • did you forget the trailing '? (fixed, just here)
  • you can skip the {1} as it's implied (optional)

Tested syntax (besides the 8-6 mismatch) is:
--gridDist '160{2},64{2},160{2}' '288{8}' '320,128,320'

Indeed, while copy-pasting I lost the trailing ' - I will add added it.

In the submit script used, there are spaces after the commas. Does this cause errors?

The spaces in the { } are caused by markdown formatting to include bold font. There are no spaces in the submit.start.

The target number of cells is: 2304.

Fixed in #2876 :)

Was this page helpful?
0 / 5 - 0 ratings

Related issues

psychocoderHPC picture psychocoderHPC  路  4Comments

HighIander picture HighIander  路  4Comments

saipavankalyan picture saipavankalyan  路  3Comments

berceanu picture berceanu  路  3Comments

ax3l picture ax3l  路  3Comments