Changes between Version 67 and Version 68 of doc/app/iofiles

Timestamp:

Sep 15, 2011 12:20:59 PM (13 years ago)

Author:

raasch

Comment:

--

doc/app/iofiles

@@ -4 +4 @@
 On parallel computers many of the files are read and/or written by the central processing element 0 (PE0) only. These files have processor-independent content (and therefore they can be read or written by other PEs just as well). However, certain files have processor-dependent content. For the binary output of data for restart runs (local file [#BINOUT BINOUT]), for example, each PE outputs only the data of its subdomain. So each processing element writes into its own file with its own file name. These files lie in a subdirectory of the temporary working directory. Their names are build off from the underline ("_") and the four digit processor ID. The data written for restart runs would be e.g. on the files BINOUT/_0000 (PE0), BINOUT/_0001 (PE1), BINOUT/_0002 (PE2) etc. Such files, which have processor-dependent content on parallel computers, are marked in the following list by the fact, that to the file name a line ("/") is attached. If appropriate output files are to be copied through '''mrun''' to permanent files, and/or files with processor-dependent content are supposed to be copied as input files into the temporary working directory of the model, you have to indicate a special file attribute in the appropriate file connection statement (see [wiki:arpe arpe], [wiki:flpe flpe] in the '''mrun''' description). Then the permanent file name will be interpreted as a directory name, in which the input files are expected and/or to which output files are written. The file names in these directories are always named _0000, _0001, _0002 etc. \\\\
 For internal use, the model may open a set of further files, which are not defined by the user but contain no usable information and therefore are not included in this list. \\\\
+Depending on the configuration of the underlying file-system (e.g. Lustre) and the capacity of the I/O-hardware, simultaneous output to a larger number of files, i.e. if a larger number of PEs is used (>1000), may lead to severe problems as job aborts or even to a complete crash of the whole system. In order to circumvent this problem, the maximum number of parallel I/O streams (i.e. the number of files which are simultaneously written), can be limited by using the '''mrun'''-option {{{-w <max # of streams>}}}, where {{{<max # of streams>}}} should be smaller or equal 256 (on the SGI-ICE of HLRN-II). The (parallel) output is then done in a sequential order for blocks of 256 PEs. \\\\
 In case of [../examples/coupled coupled] atmosphere-ocean runs, both the atmosphere and the [../inipar#ocean ocean] executable use the same temporary working directory. However, each executable requires its own, unique set of files for I/O. In order to distinguish between atmosphere and ocean files, coupled atmosphere-ocean runs use the following filename convention. The atmosphere executable uses the set of normal filenames given in the table below. The ocean executable uses a set of modified filenames that have the string '_O' added to their normal name. The coupled ocean filenames are given in brackets in the table below where applicable. The string '_O' is simply appended to most filenames; exceptions are highlighted in bold face. ('''Note''': uncoupled ocean runs use the normal set of filenames without '_O'.) \\