random floating point errors in ice_fct.F90 in subroutine ice_fem_fct

[suvarchal]

floating point error occurs around:

fesom2/src/ice_fct.F90

Line 718 in 231c416

icepplus(n)=min(1.0_WP,tmax(n)/flux)

presumably a OMP racing condition? that leads to flux=0.0 its initialized value.

[suvarchal]

my fear is either locks or ordered are enough for ensuring non-racing condition here:

fesom2/src/ice_fct.F90

Lines 682 to 692 in 231c416

	if (flux>0) then
	#if !defined(DISABLE_OPENACC_ATOMICS)
	!$ACC ATOMIC UPDATE
	#endif
	icepplus(n)=icepplus(n)+flux
	else
	#if !defined(DISABLE_OPENACC_ATOMICS)
	!$ACC ATOMIC UPDATE
	#endif
	icepminus(n)=icepminus(n)+flux
	end if

we may have to use !$OMP ATOMIC UPDATE there?

[trackow]

Hey Suvi, just wanted to add that we now see this issue also on Levante (Rohit, not sure which branch of FESOM), and on Atos ECMWF (with various branches, e.g. the old fesom2.5-fix-cycle3 and ifs-support-bundle by Sebastian).

This makes me think this could have to do with some updates to Levante or Atos themselves, because we did not see those in the past with those branches and could run 5 years without issues like this on Levante (May) and Atos ECMWF around summer. Or could we have been this lucky?