Re: Alpha speed on disk devices

Thread View: vmsnet.alpha

2 messages

[T] Back to Threads [V] Tree View [G] Group Articles

2 total messages Started by Art Kreymer Fri, 22 Jan 1993 21:23

#34

Author: Art Kreymer
Date: Fri, 22 Jan 1993 21:23

54 lines
2370 bytes

    There was earlier discussion of I/O problems on some Alpha systems.
    Fortran unformatted I/O to disk seems to be somewhat slower than
    expected, and, surprisingly, is CPU limited. That is, the CPU is
    running at 100% utilization in USER mode while the I/O is in progress.

    The I/O problem is almost certainly related to the Fortran Compiler.
    Note the following timings, writing a 32,000 block (16 Mbyte) file
    composed of records of length 8000 words (32 Kbytes)

    Formatted writes and VESTed images use much less CPU than native
    unformatted writes. Implied I/O do loops suffer no or negative
    overhead.

    Elapsed    CPU    Method
      15         1.   UNFORMATTED  WRITE (1) DATA  (vested image from VAX)
      25.        1.   FORMATTED    WRITE (1,'(A32000)') CDATA
      20.       15.   UNFORMATTED  WRITE (1) DATA  (native image)
      15.       15.   UNFORMATTED  WRITE (1) (DATA(J),J=1,8000)

    Here's the test program:

          PROGRAM WRITE

          PARAMETER ( LEN = 8000 )
          INTEGER DATA ( LEN )
          CHARACTER*(4*LEN) CDATA
          EQUIVALENCE ( CDATA , DATA )

          DO I = 1 , LEN
             DATA(I) = I
          ENDDO

          OPEN ( 1 , FILE = 'TEST' ,
         +   STATUS = 'NEW' ,
         +   FORM = 'UNFORMATTED' ,    ! unformatted
    !     +   FORM = 'FORMATTED' ,     !   formatted
         +   RECL = 8000 ,             ! unformatted
    !     +   RECL = 32000 ,           ! formatted
         +   RECORDTYPE = 'VARIABLE' ,
         +   INITIALSIZE = 32 000 )

          DO I = 1 , 512
             WRITE (1)  DATA                 ! unformatted
    !        WRITE (1) (DATA(J),J=1,LEN)     ! unformatted implied do
    !         WRITE (1, '(A32000)' ) CDATA   !   formatted
          ENDDO

          WRITE (*,*) ' DONE '
          CLOSE (1)

          CALL EXIT
          STOP
          END

Re: Alpha speed on disk devices

#48

Author: Michele Michelotto
Date: Tue, 26 Jan 1993 08:11

47 lines
2899 bytes

In article <1jpojvINNin@fnnews.fnal.gov>, kreymer@fnsg01.fnal.gov (Art Kreymer) writes:
|>    There was earlier discussion of I/O problems on some Alpha systems.
|>    Fortran unformatted I/O to disk seems to be somewhat slower than
|>    expected, and, surprisingly, is CPU limited. That is, the CPU is
|>    running at 100% utilization in USER mode while the I/O is in progress.
|>
|>    The I/O problem is almost certainly related to the Fortran Compiler.
|>    Note the following timings, writing a 32,000 block (16 Mbyte) file
|>    composed of records of length 8000 words (32 Kbytes)
|>
|>    Formatted writes and VESTed images use much less CPU than native
|>    unformatted writes. Implied I/O do loops suffer no or negative
|>    overhead.
|>
|>    Elapsed    CPU    Method
|>      15         1.   UNFORMATTED  WRITE (1) DATA  (vested image from VAX)
|>      25.        1.   FORMATTED    WRITE (1,'(A32000)') CDATA
|>      20.       15.   UNFORMATTED  WRITE (1) DATA  (native image)
|>      15.       15.   UNFORMATTED  WRITE (1) (DATA(J),J=1,8000)
|>
|>    Here's the test program:
.. deleted...

There is another strange behaviour: On the VAX you obtain moreless the same
results (in KB/s) if you declare DATA as integer*1(character),integer*2,
integer*4, integer*8;if you do unformatted write. This is normal because
the write are buffered. This happens also with the VESTed images omn the
Alpha.

With native code on the Alpha you obtain very good result with i*8
(and low cpu/elapsed ratio) and very bad result with i*1 (with cpu/elaps
that is close to 1.).

OK. Now I can declare an equivalence of my character buffer with a REAL*16
buffer, I can do all my operation with the character buffer and then do the
I/O with the REAL*16 buffer. I can achieve very good performance but

WHY ON THE VAX I DON'T NEED TO DO THESE TRICKS?
--
						Ciao, Michele
******************************************************************************
* Michele Michelotto, INFN Padova, Delphi group.                             *
* Internet: michelot@dxcern.cern.ch     BITNET: MICHELOTTO@IPDINFN           *
* Tel (+39).49.844347                   HEPNET: VAXFPD::MICHELOTTO           *
*                     "Ho bevuto CHE COSA ?"   (Socrate)                     *
******************************************************************************

Thread Navigation

This is a paginated view of messages in the thread with full content displayed inline.

Messages are displayed in chronological order, with the original post highlighted in green.

Use pagination controls to navigate through all messages in large threads.

Back to All Threads