TELEMAC-MASCARET Forum: bug in parallel preconditioning ARTEMIS (1/1)

Welcome, Guest

TOPIC: bug in parallel preconditioning ARTEMIS

bug in parallel preconditioning ARTEMIS 9 years 8 months ago #15918

maxim.sorockin
OFFLINE
Fresh Boarder
Posts: 13

I found one more bug in ARTEMIS. It affects parallel run.

The call of subroutine CNTPRE in subroutine BERKHO may change preconditioning flag SLVART%PRECON.
So in parallel different processors may have different value of preconditioning flag, because in CNTPRE no global checking.
Then in subroutine SOLVE some of processors may call and some may don't call PRECDT subroutine (see part of code from SOLVE)

      IF(2*(CFG%PRECON/2).EQ.CFG%PRECON.OR.
     &   3*(CFG%PRECON/3).EQ.CFG%PRECON.OR.
     &   5*(CFG%PRECON/5).EQ.CFG%PRECON) THEN
        CALL PRECDT(X,A,B,TBB%ADR(IT1)%P,MESH,
     &              CFG%PRECON,PREXSM,DIADON,S)
      ENDIF

But PRECDT through calls of PRECD4 -> PARCOM -> PARCOM2 -> PARACO - > P_IREAD, P_IWRIT, P_WAIT_PARACO call MPI functions. So some of processors call and some don't call MPI functions. That cause crash of ARTEMIS in parallel.

May be global checking of preconditioning flag in or after CNTPRE subroutine is needed.

The administrator has disabled public write access.

bug in parallel preconditioning ARTEMIS 9 years 8 months ago #15920

jmhervouet

Hello,

Thanks for pointing out this problem, you are perfectly right. I enclose a proposition of correction of subroutine cntpre.f, using function P_IMIN, so that all the processors end up with the minimum value of preconditioning that has been found (the action of cntpre is just dividing by 2, 3 or 5).

I just tested it in parallel on a case that has no problem of diagonal equal to 0, so this is without guaranty, and moreover I am not sure of what can happen when a diagonal is really 0, but well, let's see.

With best regards,

Jean-Michel Hervouet

Attachments:

cntpre.f (4KB)

The administrator has disabled public write access.

bug in parallel preconditioning ARTEMIS 9 years 8 months ago #15925

maxim.sorockin OFFLINE Fresh Boarder Posts: 13	Thank you Jean-Michel. Very appreciate for your answer! Did you saw my topic about ARTEMIS with MUMPS? Do you have any thoughts about described problem? I will be very grateful for the advice. Maxim Sorokin
	The administrator has disabled public write access.

bug in parallel preconditioning ARTEMIS 9 years 8 months ago #15926

jmhervouet

Hello maxim,

Yes, I was away for some time but looked yesterday at all your remarks. I agree with you for them all and arrive at the same conclusionsnd corrections. I remained stuck by the dimensioning problem of STOSEG4 and contacted the author. Though he has now moved and does something else I am confident he will give us a feed back soon on this matter.

With best regards,

Jean-Michel Hervouet

The administrator has disabled public write access.

bug in parallel preconditioning ARTEMIS 9 years 8 months ago #15961

maxim.sorockin OFFLINE Fresh Boarder Posts: 13	Very appreciate for your attention, Jean-Michel. Thank you! Maxim Sorokin
	The administrator has disabled public write access.