FTN_MPIO(5) FTN_MPIO(5)
FTN_MPIO - Fortran Multi-Threaded I/O
The 64-bit Fortran compiler and runtime libraries provide multi-threaded
I/O functionality for Fortran, in other words, this capability is
available only to programs compiled with -mips3 or -mips4 option and not
those compiled with -mips1 or -mips2. To do Fortran I/O in parallel, the
threaded subroutine must be compiled with the -mpio option. A few
things need to be noted:
1) All subroutines in source files compiled with -mpio option use a
different I/O interface which allows MP-safe Fortran I/O. However, these
MP interfaces do take extra runtime check to ensure the integrity of the
I/O operations and so would effect runtime performance a little (probably
less than 5%). For applications which wish to make the most out of
runtime performance, only those subroutines which get executed in
parallel or contain parallelized loops with I/O statements in them should
be compiled with -mpio option. The rest could and should be compiled
without -mpio option.
2) The I/O runtime library will ensure that only one thread gets a
lock on a particular logical unit to avoid several threads stepping on
each others. The lock is only released after this thread completes the
I/O operation. This implementation means three things: firstly, multiple
threads can write to multiple files connected to different Fortran
logical units at the same time; secondly, you cannot have multiple
threads writing to the same logical unit and expect it to be faster as,
most likely, it will be significantly slower due to all the lockings,
queryings, and unlockings on that same logical unit; and, thirdly, when
you open the same file as different logical units and then having
multiple threads writing to those logical units to the same file the
integrity of the file will be compromised.
3) The I/O runtime routines do not have control over which threads
get executed and will execute whichever thread that obtains the lock on a
particular logical unit first. Therefore, this ability to have multithreaded
I/O is most useful in direct access or keyed access I/O where
each thread performs I/O on a pre-determined record (or where each thread
performs I/O on a different file) and, therefore, the result is
independent on the execution order of the threads. For sequential files,
since the order of the records written out or read in depends on which
thread gets executed first, the result could be incorrect unless the
records are not sequentially related and can be written or read in any
order. For example:
DO 10, I=1,100
PRINT *, I
10 CONTINUE
If this loop is threaded, then the numbers printed out will no longer in
the sequential order from 1 to 100 and applications depending on the
numbers being in that strict order will get the wrong result.
4) Because of the problem with sequential files above, the compiler
(and/or pfa) will not parallelize loops containing I/O statements
Page 1
FTN_MPIO(5) FTN_MPIO(5)
automatically as it would compromise the correctness of the application.
Loops containing I/O statements must be parallelized manually by adding
the appropriate directives. Users who wish to use the ability of pfa,
for example, to analyze the loop for multi-threading can comment out the
I/O statements to get the .m file with all the parallelization directives
added to the appropriate loops, and then uncomment the I/O statements and
compile the .m file with the -mpio option. That is if they are sure that
the I/O statements won't be effected by the order in which they are
executed. Also because of the performance loss associated with multithreaded
Fortran I/O the -mpio option is never the default regardless of
other compilation options. The user must explicitly specify this option
in case he wants to have I/O statements inside multi-threaded
loops/subroutines.
5) The Fortran standard forbids the use of a function call in a list
of items to be written out in a WRITE/PRINT statement if that function
itself performs I/O. This nested I/O usage used to give random result or
coredump before the multi-threaded I/O implementation and still does now
if the application is not compiled with -mpio option. With the -mpio
option, this will give the correct result if the I/O operation inside the
function is not performed on the same logical unit as the one in the I/O
statement where the function call is used but will result in a deadlock
and the process will hang if the nested I/O operations are done on the
same logical unit.
6) Internal file I/O is treated as if it were an independent logical
unit. That means only one thread can do internal file I/O at a time.
7) Since the MP I/O runtime library sets its own internal lock on
the logical unit, it does not effect the normal I/O operation to the
files and does not impose additional constraint on the operability of the
I/O statements. In other words, I/O operations to NFS-mounted files,
tapes, sockets, etc. do not have any more limitations than those they
already have, if any.
8) The -mpio option effects the runtime library calls generated from
the I/O statements directly: a different set of MP-safe I/O interfaces
will be used instead of the default interfaces. That means this option,
unlike those compilation options which changes the behavior of the
runtime I/O library such as -vms_cc by setting a flag inside the Fortran
main program, is not effected by the fact whether you have the main
program written in C or Fortran. 9) The default maximum number of
Fortran logical units that can be opened at the same time is 100,
including 4 system files. In single-threaded mode this limit is
automatically increased if the number of opened exceeds it. However,
while I/O operations are being executed in multi-threaded mode, if
additional files are opened causing the number of opened files to exceed
this limit, the runtime library will return an error message asking the
user to set the environment variable FORTRAN_OPENED_UNITS to a bigger
number and then abort. In order to run the program to completion, the
user will need to set FORTRAN_OPENED_UNITS to a number big enough to
accommodate for all the files being opened at the same time. To avoid
this from occurring at all without having to set FORTRAN_OPENED_UNITS,
all OPEN statements can be done in single-threaded mode before doing
other I/O operations in parallel.
Page 2
FTN_MPIO(5) FTN_MPIO(5)
Calvin Vu
PPPPaaaaggggeeee 3333 [ Back ]
|
