Host memory usage increasing

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

I’m playing around with the PGI accelerators. I noticed that when I run the accelerator version of the code below my system memory increases linearly until the program finishes.

Could someone explain why this is and what appropriate data clauses I should use to fix this?

      PROGRAM TEST
      IMPLICIT DOUBLE PRECISION (A-H,O-Z)
C
      PARAMETER (NAT=27,NRADPT=96,NLEBPT=1202,NGRIDPT=96*1202)
      PARAMETER (ONE=1.0D+00)
C
      DIMENSION
     > GTEMPA(NAT,NGRIDPT),
     > GTEMPB(NAT,NAT),
     > GTEMPC(NAT)
C
      DO IRAD=1,NRADPT
        DO IANG=1,NLEBPT
        IPTME=(IRAD-1)*NLEBPT+IANG
!$acc data region copyout(gtempa(1:nat,iptme))
!$acc region local(gtempb(1:nat,1:nat),gtempc(1:nat))
          DO IATM=1,NAT
            GTEMPC(IATM)=ONE
            DO JATM=1,NAT
              GTEMPB(JATM,IATM)=ONE
              IF(IATM .EQ. JATM) CYCLE
              GTEMPB(JATM,IATM)=DBLE(JATM+IATM)
            ENDDO
            GTEMPC(IATM)=PRODUCT(GTEMPB(1:NAT,IATM),1)
          ENDDO
          GTEMPA(1:NAT,IPTME)=GTEMPC(1:NAT)
!$acc end region
!$acc end data region
        ENDDO
      ENDDO
      WRITE(999,*) GTEMPA
      END




     15, Generating copyout(gtempa(:,iptme))
     16, Generating local(gtempb(:,:))
         Generating local(gtempc(:))
     17, Loop is parallelizable
         Accelerator kernel generated
         17, !$acc do parallel, vector(27) ! blockidx%x threadidx%x
             Using register for 'gtempc'
     19, Loop is parallelizable
     24, product reduction inlined
         Loop is parallelizable
     26, Loop is parallelizable
         Accelerator kernel generated
         26, !$acc do parallel, vector(27) ! blockidx%x threadidx%x
2

Hi sslgamess,

What flags are you using? I’m able to recreate the issue but only when I use the “time” profiling sub-option (-ta=nvidia,time). It seems fine without the “time” sub-option.

I’ve send a report to our engineers (TPR#18371) to investigate if this is a memory leak with the profiling code, expected, or something else.

Thanks,
Mat

3

Hi Mat,

Yes, I am using the time flag.

Thanks for submitting the TPR.

4

We did indeed find a memory leak for programs using the -ta=nvidia,time option. It turned out the runtime was creating new cudaEvents, without recycling the old ones. That will be fixed in the 12.2 release.

Thanks.

5

Thanks for the update Michael

6

TPR 18371 was fixed in the 12.6 release.
Sorry about the delay.

dave