1
1
openmpi/ompi/runtime/ompi_mpi_init.c

1153 строки
46 KiB
C
Исходник Обычный вид История

/*
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2005 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006-2007 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2006-2007 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2006 University of Houston. All rights reserved.
*
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#ifdef HAVE_SYS_TIME_H
#include <sys/time.h>
#endif /* HAVE_SYS_TIME_H */
#include "mpi.h"
#include "opal/class/opal_list.h"
#include "opal/mca/base/base.h"
#include "opal/mca/paffinity/base/base.h"
#include "opal/mca/maffinity/base/base.h"
#include "opal/runtime/opal_progress.h"
#include "opal/threads/threads.h"
#include "opal/util/argv.h"
#include "opal/util/show_help.h"
#include "opal/util/stacktrace.h"
#include "opal/util/num_procs.h"
#include "opal/runtime/opal.h"
#include "opal/event/event.h"
#include "orte/util/sys_info.h"
#include "orte/util/proc_info.h"
#include "orte/util/session_dir.h"
#include "orte/util/name_fns.h"
#include "orte/runtime/runtime.h"
#include "orte/mca/rml/rml.h"
#include "orte/mca/errmgr/errmgr.h"
#include "orte/mca/grpcomm/grpcomm.h"
#include "orte/runtime/orte_globals.h"
#include "ompi/constants.h"
#include "ompi/mpi/f77/constants.h"
#include "ompi/runtime/mpiruntime.h"
#include "ompi/runtime/params.h"
#include "ompi/communicator/communicator.h"
#include "ompi/group/group.h"
#include "ompi/info/info.h"
#include "ompi/errhandler/errcode.h"
#include "ompi/request/request.h"
#include "ompi/op/op.h"
#include "ompi/file/file.h"
#include "ompi/attribute/attribute.h"
#include "ompi/mca/allocator/base/base.h"
#include "ompi/mca/allocator/allocator.h"
#include "ompi/mca/rcache/base/base.h"
#include "ompi/mca/rcache/rcache.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/mpool/mpool.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/runtime/ompi_module_exchange.h"
#include "ompi/mca/pml/base/base.h"
#include "ompi/mca/osc/base/base.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/mca/coll/base/base.h"
#include "ompi/mca/io/io.h"
#include "ompi/mca/io/base/base.h"
#include "ompi/debuggers/debuggers.h"
#include "ompi/proc/proc.h"
#include "ompi/mca/pml/base/pml_base_bsend.h"
#include "ompi/mca/dpm/base/base.h"
#include "ompi/mca/pubsub/base/base.h"
#if OPAL_ENABLE_FT == 1
#include "ompi/mca/crcp/crcp.h"
#include "ompi/mca/crcp/base/base.h"
#endif
#include "ompi/runtime/ompi_cr.h"
static int slot_list_to_cpu_set(char *slot_str);
#include "orte/runtime/orte_globals.h"
/*
* Global variables and symbols for the MPI layer
*/
bool ompi_mpi_initialized = false;
bool ompi_mpi_finalized = false;
bool ompi_mpi_thread_multiple = false;
int ompi_mpi_thread_requested = MPI_THREAD_SINGLE;
int ompi_mpi_thread_provided = MPI_THREAD_SINGLE;
opal_thread_t *ompi_mpi_main_thread = NULL;
bool ompi_mpi_maffinity_setup = false;
/*
* These variables are here, rather than under ompi/mpi/c/foo.c
* because it is not sufficient to have a .c file that only contains
* variables -- you must have a function that is invoked from
* elsewhere in the code to guarantee that all linkers will pull in
* the .o file from the library. Hence, although these are MPI
* constants, we might as well just define them here (i.e., in a file
* that already has a function that is guaranteed to be linked in,
* rather than make a new .c file with the constants and a
* corresponding dummy function that is invoked from this function).
*
Clean up the Fortran MPI sentinel values per problem reported on the users mailing list: http://www.open-mpi.org/community/lists/users/2006/07/1680.php Warning: this log message is not for the weak. Read at your own risk. The problem was that we had several variables in Fortran common blocks of various types, but their C counterparts were all of a type equivalent to a fortran double complex. This didn't seem to matter for the compilers that we tested, but we never tested static builds (which is where this problem seems to occur, at least with the Intel compiler: the linker compilains that the variable in the common block in the user's .o file was of one size/alignment but the one in the C library was a different size/alignment). So this patch fixes the sizes/types of the Fortran common block variables and their corresponding C instantiations to be of the same sizes/types. But wait, there's more. We recently introduced a fix for the OSX linker where some C versions of the fortran common block variables (e.g., _ompi_fortran_status_ignore) were not being found when linking ompi_info (!). Further research shows that the code path for ompi_info to require ompi_fortran_status_ignore is, unfortunately, necessary (a quirk of how various components pull in different portions of the code base -- nothing in ompi_info itself requires fortran or MPI knowledge, of course). Hence, the real problem was that there was no code path from ompi_info to the portion of the code base where the C globals corresponding to the Fortran common block variables were instantiated. This is because the OSX linker does not automatically pull in .o files that only contain unintialized global variables; the OSX linker typically only pulls in a .o file from a library if it either has a function that is used or have a global variable that is initialized (that's the short version; lots of details and corner cases omitted). Hence, we changed the global C variables corresponding to the fortran common blocks to be initialized, thereby causing the OSX linker to pull them in automatically -- problem solved. At the same time, we moved the constants to another .c file with a function, just for good measure. However, this didn't really solve the problem: 1. The function in the file with the C versions of the fortran common block variables (ompi/mpi/f77/test_constants_f.c) did not have a code path that was reachable from ompi_info, so the only reason that the constants were found (on OSX) was because they were initialized in the global scope (i.e., causing the OSX compiler to pull in that .o file). 2. Initializing these variable in the global scope causes problems for some linkers where -- once all the size/type problems mentioned above were fixed -- the alignments of fortran common blocks and C global variables do not match (even though the types of the Fortran and C variables match -- wow!). Hence, initializing the C variables would not necessarily match the alignment of what Fortran expected, and the linker would issue a warning (i.e., the alignment warnings referenced in the original post). The solution is two-fold: 1. Move the Fortran variables from test_constants_f.c to ompi/mpi/runtime/ompi_mpi_init.c where there are other global constants that *are* initialized (that had nothing to do with fortran, so the alignment issues described above are not a factor), and therefore all linkers (including the OSX linker) will pull in this .o file and find all the symbols that it needs. 2. Do not initialize the C variables corresponding to the Fortran common blocks in the global scope. Indeed, never initialize them at all (because we never need their *values* - we only check for their *locations*). Since nothing is ever written to these variables (particularly in the global scope), the linker does not see any alignment differences during initialization, but does make both the C and Fortran variables have the same addresses (this method has been working in LAM/MPI for over a decade). There were some comments here in the OMPI code base and in the LAM code base that stated/implied that C variables corresponding to Fortran common blocks had to have the same alignment as the Fortran common blocks (i.e., 16). There were attempts in both code bases to ensure that this was true. However, the attempts were wrong (in both code bases), and I have now read enough Fortran compiler documentation to convince myself that matching alignments is not required (indeed, it's beyond our control). As long as C variables corresponding to Fortran common blocks are not initialized in the global scope, the linker will "figure it out" and adjust the alignment to whatever is required (i.e., the greater of the alignments). Specifically (to counter comments that no longer exist in the OMPI code base but still exist in the LAM code base): - there is no need to make attempts to specially align C variables corresponding to Fortran common blocks - the types and sizes of C variables corresponding to Fortran common blocks should match, but do not need to be on any particular alignment Finally, as a side effect of this effort, I found a bunch of inconsistencies with the intent of status/array_of_statuses parameters. For all the functions that I modified they should be "out" (not inout). This commit was SVN r11057.
2006-07-31 19:07:09 +04:00
* Additionally, there can be/are strange linking paths such that
* ompi_info needs symbols such as ompi_fortran_status_ignore,
Clean up the Fortran MPI sentinel values per problem reported on the users mailing list: http://www.open-mpi.org/community/lists/users/2006/07/1680.php Warning: this log message is not for the weak. Read at your own risk. The problem was that we had several variables in Fortran common blocks of various types, but their C counterparts were all of a type equivalent to a fortran double complex. This didn't seem to matter for the compilers that we tested, but we never tested static builds (which is where this problem seems to occur, at least with the Intel compiler: the linker compilains that the variable in the common block in the user's .o file was of one size/alignment but the one in the C library was a different size/alignment). So this patch fixes the sizes/types of the Fortran common block variables and their corresponding C instantiations to be of the same sizes/types. But wait, there's more. We recently introduced a fix for the OSX linker where some C versions of the fortran common block variables (e.g., _ompi_fortran_status_ignore) were not being found when linking ompi_info (!). Further research shows that the code path for ompi_info to require ompi_fortran_status_ignore is, unfortunately, necessary (a quirk of how various components pull in different portions of the code base -- nothing in ompi_info itself requires fortran or MPI knowledge, of course). Hence, the real problem was that there was no code path from ompi_info to the portion of the code base where the C globals corresponding to the Fortran common block variables were instantiated. This is because the OSX linker does not automatically pull in .o files that only contain unintialized global variables; the OSX linker typically only pulls in a .o file from a library if it either has a function that is used or have a global variable that is initialized (that's the short version; lots of details and corner cases omitted). Hence, we changed the global C variables corresponding to the fortran common blocks to be initialized, thereby causing the OSX linker to pull them in automatically -- problem solved. At the same time, we moved the constants to another .c file with a function, just for good measure. However, this didn't really solve the problem: 1. The function in the file with the C versions of the fortran common block variables (ompi/mpi/f77/test_constants_f.c) did not have a code path that was reachable from ompi_info, so the only reason that the constants were found (on OSX) was because they were initialized in the global scope (i.e., causing the OSX compiler to pull in that .o file). 2. Initializing these variable in the global scope causes problems for some linkers where -- once all the size/type problems mentioned above were fixed -- the alignments of fortran common blocks and C global variables do not match (even though the types of the Fortran and C variables match -- wow!). Hence, initializing the C variables would not necessarily match the alignment of what Fortran expected, and the linker would issue a warning (i.e., the alignment warnings referenced in the original post). The solution is two-fold: 1. Move the Fortran variables from test_constants_f.c to ompi/mpi/runtime/ompi_mpi_init.c where there are other global constants that *are* initialized (that had nothing to do with fortran, so the alignment issues described above are not a factor), and therefore all linkers (including the OSX linker) will pull in this .o file and find all the symbols that it needs. 2. Do not initialize the C variables corresponding to the Fortran common blocks in the global scope. Indeed, never initialize them at all (because we never need their *values* - we only check for their *locations*). Since nothing is ever written to these variables (particularly in the global scope), the linker does not see any alignment differences during initialization, but does make both the C and Fortran variables have the same addresses (this method has been working in LAM/MPI for over a decade). There were some comments here in the OMPI code base and in the LAM code base that stated/implied that C variables corresponding to Fortran common blocks had to have the same alignment as the Fortran common blocks (i.e., 16). There were attempts in both code bases to ensure that this was true. However, the attempts were wrong (in both code bases), and I have now read enough Fortran compiler documentation to convince myself that matching alignments is not required (indeed, it's beyond our control). As long as C variables corresponding to Fortran common blocks are not initialized in the global scope, the linker will "figure it out" and adjust the alignment to whatever is required (i.e., the greater of the alignments). Specifically (to counter comments that no longer exist in the OMPI code base but still exist in the LAM code base): - there is no need to make attempts to specially align C variables corresponding to Fortran common blocks - the types and sizes of C variables corresponding to Fortran common blocks should match, but do not need to be on any particular alignment Finally, as a side effect of this effort, I found a bunch of inconsistencies with the intent of status/array_of_statuses parameters. For all the functions that I modified they should be "out" (not inout). This commit was SVN r11057.
2006-07-31 19:07:09 +04:00
* which, if they weren't here with a collection of other global
* symbols that are initialized (which seems to force this .o file to
* be pulled into the resolution process, because ompi_info certainly
* does not call ompi_mpi_init()), would not be able to be found by
* the OSX linker.
*
* NOTE: See the big comment in ompi/mpi/f77/constants.h about why we
* have four symbols for each of the common blocks (e.g., the Fortran
* equivalent(s) of MPI_STATUS_IGNORE). Here, we can only have *one*
* value (not four). So the only thing we can do is make it equal to
* the fortran compiler convention that was selected at configure
* time. Note that this is also true for the value of .TRUE. from the
* Fortran compiler, so even though Open MPI supports all four Fortran
* symbol conventions, it can only support one convention for the two
* C constants (MPI_FORTRAN_STATUS[ES]_IGNORE) and only support one
* compiler for the value of .TRUE. Ugh!!
*
* Note that the casts here are ok -- we're *only* comparing pointer
* values (i.e., they'll never be de-referenced). The global symbols
* are actually of type (ompi_fortran_common_t) (for alignment
* issues), but MPI says that MPI_F_STATUS[ES]_IGNORE must be of type
* (MPI_Fint*). Hence, we have to cast to make compilers not
* complain.
*/
#if OMPI_WANT_F77_BINDINGS
# if OMPI_F77_CAPS
MPI_Fint *MPI_F_STATUS_IGNORE = (MPI_Fint*) &MPI_FORTRAN_STATUS_IGNORE;
MPI_Fint *MPI_F_STATUSES_IGNORE = (MPI_Fint*) &MPI_FORTRAN_STATUSES_IGNORE;
# elif OMPI_F77_PLAIN
MPI_Fint *MPI_F_STATUS_IGNORE = (MPI_Fint*) &mpi_fortran_status_ignore;
MPI_Fint *MPI_F_STATUSES_IGNORE = (MPI_Fint*) &mpi_fortran_statuses_ignore;
# elif OMPI_F77_SINGLE_UNDERSCORE
MPI_Fint *MPI_F_STATUS_IGNORE = (MPI_Fint*) &mpi_fortran_status_ignore_;
MPI_Fint *MPI_F_STATUSES_IGNORE = (MPI_Fint*) &mpi_fortran_statuses_ignore_;
# elif OMPI_F77_DOUBLE_UNDERSCORE
MPI_Fint *MPI_F_STATUS_IGNORE = (MPI_Fint*) &mpi_fortran_status_ignore__;
MPI_Fint *MPI_F_STATUSES_IGNORE = (MPI_Fint*) &mpi_fortran_statuses_ignore__;
# else
# error Unrecognized Fortran 77 name mangling scheme
# endif
#else
MPI_Fint *MPI_F_STATUS_IGNORE = NULL;
MPI_Fint *MPI_F_STATUSES_IGNORE = NULL;
#endif /* OMPI_WANT_F77_BINDINGS */
Clean up the Fortran MPI sentinel values per problem reported on the users mailing list: http://www.open-mpi.org/community/lists/users/2006/07/1680.php Warning: this log message is not for the weak. Read at your own risk. The problem was that we had several variables in Fortran common blocks of various types, but their C counterparts were all of a type equivalent to a fortran double complex. This didn't seem to matter for the compilers that we tested, but we never tested static builds (which is where this problem seems to occur, at least with the Intel compiler: the linker compilains that the variable in the common block in the user's .o file was of one size/alignment but the one in the C library was a different size/alignment). So this patch fixes the sizes/types of the Fortran common block variables and their corresponding C instantiations to be of the same sizes/types. But wait, there's more. We recently introduced a fix for the OSX linker where some C versions of the fortran common block variables (e.g., _ompi_fortran_status_ignore) were not being found when linking ompi_info (!). Further research shows that the code path for ompi_info to require ompi_fortran_status_ignore is, unfortunately, necessary (a quirk of how various components pull in different portions of the code base -- nothing in ompi_info itself requires fortran or MPI knowledge, of course). Hence, the real problem was that there was no code path from ompi_info to the portion of the code base where the C globals corresponding to the Fortran common block variables were instantiated. This is because the OSX linker does not automatically pull in .o files that only contain unintialized global variables; the OSX linker typically only pulls in a .o file from a library if it either has a function that is used or have a global variable that is initialized (that's the short version; lots of details and corner cases omitted). Hence, we changed the global C variables corresponding to the fortran common blocks to be initialized, thereby causing the OSX linker to pull them in automatically -- problem solved. At the same time, we moved the constants to another .c file with a function, just for good measure. However, this didn't really solve the problem: 1. The function in the file with the C versions of the fortran common block variables (ompi/mpi/f77/test_constants_f.c) did not have a code path that was reachable from ompi_info, so the only reason that the constants were found (on OSX) was because they were initialized in the global scope (i.e., causing the OSX compiler to pull in that .o file). 2. Initializing these variable in the global scope causes problems for some linkers where -- once all the size/type problems mentioned above were fixed -- the alignments of fortran common blocks and C global variables do not match (even though the types of the Fortran and C variables match -- wow!). Hence, initializing the C variables would not necessarily match the alignment of what Fortran expected, and the linker would issue a warning (i.e., the alignment warnings referenced in the original post). The solution is two-fold: 1. Move the Fortran variables from test_constants_f.c to ompi/mpi/runtime/ompi_mpi_init.c where there are other global constants that *are* initialized (that had nothing to do with fortran, so the alignment issues described above are not a factor), and therefore all linkers (including the OSX linker) will pull in this .o file and find all the symbols that it needs. 2. Do not initialize the C variables corresponding to the Fortran common blocks in the global scope. Indeed, never initialize them at all (because we never need their *values* - we only check for their *locations*). Since nothing is ever written to these variables (particularly in the global scope), the linker does not see any alignment differences during initialization, but does make both the C and Fortran variables have the same addresses (this method has been working in LAM/MPI for over a decade). There were some comments here in the OMPI code base and in the LAM code base that stated/implied that C variables corresponding to Fortran common blocks had to have the same alignment as the Fortran common blocks (i.e., 16). There were attempts in both code bases to ensure that this was true. However, the attempts were wrong (in both code bases), and I have now read enough Fortran compiler documentation to convince myself that matching alignments is not required (indeed, it's beyond our control). As long as C variables corresponding to Fortran common blocks are not initialized in the global scope, the linker will "figure it out" and adjust the alignment to whatever is required (i.e., the greater of the alignments). Specifically (to counter comments that no longer exist in the OMPI code base but still exist in the LAM code base): - there is no need to make attempts to specially align C variables corresponding to Fortran common blocks - the types and sizes of C variables corresponding to Fortran common blocks should match, but do not need to be on any particular alignment Finally, as a side effect of this effort, I found a bunch of inconsistencies with the intent of status/array_of_statuses parameters. For all the functions that I modified they should be "out" (not inout). This commit was SVN r11057.
2006-07-31 19:07:09 +04:00
/* Constants for the Fortran layer. These values are referred to via
common blocks in the Fortran equivalents. See
ompi/mpi/f77/constants.h for a more detailed explanation.
The values are *NOT* initialized. We do not use the values of
these constants; only their addresses (because they're always
passed by reference by Fortran).
Initializing upon instantiation these can reveal size and/or
alignment differences between Fortran and C (!) which can cause
warnings or errors upon linking (e.g., making static libraries with
the intel 9.0 compilers on 64 bit platforms shows alignment
differences between libmpi.a and the user's application, resulting
in a linker warning). FWIW, if you initialize these variables in
functions (i.e., not at the instantiation in the global scope), the
linker somehow "figures it all out" (w.r.t. different alignments
between fortan common blocks and the corresponding C variables) and
no linker warnings occur.
Note that the rationale for the types of each of these variables is
discussed in ompi/include/mpif-common.h. Do not change the types
without also modifying ompi/mpi/f77/constants.h and
ompi/include/mpif-common.h.
*/
#define INST(type, upper_case, lower_case, single_u, double_u) \
type lower_case; \
type upper_case; \
type single_u; \
type double_u
INST(int, MPI_FORTRAN_BOTTOM, mpi_fortran_bottom,
mpi_fortran_bottom_, mpi_fortran_bottom__);
INST(int, MPI_FORTRAN_IN_PLACE, mpi_fortran_in_place,
mpi_fortran_in_place_, mpi_fortran_in_place__);
INST(char *, MPI_FORTRAN_ARGV_NULL, mpi_fortran_argv_null,
mpi_fortran_argv_null_, mpi_fortran_argv_null__);
INST(double, MPI_FORTRAN_ARGVS_NULL, mpi_fortran_argvs_null,
mpi_fortran_argvs_null_, mpi_fortran_argvs_null__);
INST(int *, MPI_FORTRAN_ERRCODES_IGNORE, mpi_fortran_errcodes_ignore,
mpi_fortran_errcodes_ignore_, mpi_fortran_errcodes_ignore__);
INST(int *, MPI_FORTRAN_STATUS_IGNORE, mpi_fortran_status_ignore,
mpi_fortran_status_ignore_, mpi_fortran_status_ignore__);
INST (double, MPI_FORTRAN_STATUSES_IGNORE, mpi_fortran_statuses_ignore,
mpi_fortran_statuses_ignore_, mpi_fortran_statuses_ignore__);
/*
* Per MPI-2:9.5.3, MPI_REGISTER_DATAREP is a memory leak. There is
* no way to *de*register datareps once they've been registered. So
* we have to track all registrations here so that they can be
* de-registered during MPI_FINALIZE so that memory-tracking debuggers
* don't show Open MPI as leaking memory.
*/
opal_list_t ompi_registered_datareps;
Clean up the Fortran MPI sentinel values per problem reported on the users mailing list: http://www.open-mpi.org/community/lists/users/2006/07/1680.php Warning: this log message is not for the weak. Read at your own risk. The problem was that we had several variables in Fortran common blocks of various types, but their C counterparts were all of a type equivalent to a fortran double complex. This didn't seem to matter for the compilers that we tested, but we never tested static builds (which is where this problem seems to occur, at least with the Intel compiler: the linker compilains that the variable in the common block in the user's .o file was of one size/alignment but the one in the C library was a different size/alignment). So this patch fixes the sizes/types of the Fortran common block variables and their corresponding C instantiations to be of the same sizes/types. But wait, there's more. We recently introduced a fix for the OSX linker where some C versions of the fortran common block variables (e.g., _ompi_fortran_status_ignore) were not being found when linking ompi_info (!). Further research shows that the code path for ompi_info to require ompi_fortran_status_ignore is, unfortunately, necessary (a quirk of how various components pull in different portions of the code base -- nothing in ompi_info itself requires fortran or MPI knowledge, of course). Hence, the real problem was that there was no code path from ompi_info to the portion of the code base where the C globals corresponding to the Fortran common block variables were instantiated. This is because the OSX linker does not automatically pull in .o files that only contain unintialized global variables; the OSX linker typically only pulls in a .o file from a library if it either has a function that is used or have a global variable that is initialized (that's the short version; lots of details and corner cases omitted). Hence, we changed the global C variables corresponding to the fortran common blocks to be initialized, thereby causing the OSX linker to pull them in automatically -- problem solved. At the same time, we moved the constants to another .c file with a function, just for good measure. However, this didn't really solve the problem: 1. The function in the file with the C versions of the fortran common block variables (ompi/mpi/f77/test_constants_f.c) did not have a code path that was reachable from ompi_info, so the only reason that the constants were found (on OSX) was because they were initialized in the global scope (i.e., causing the OSX compiler to pull in that .o file). 2. Initializing these variable in the global scope causes problems for some linkers where -- once all the size/type problems mentioned above were fixed -- the alignments of fortran common blocks and C global variables do not match (even though the types of the Fortran and C variables match -- wow!). Hence, initializing the C variables would not necessarily match the alignment of what Fortran expected, and the linker would issue a warning (i.e., the alignment warnings referenced in the original post). The solution is two-fold: 1. Move the Fortran variables from test_constants_f.c to ompi/mpi/runtime/ompi_mpi_init.c where there are other global constants that *are* initialized (that had nothing to do with fortran, so the alignment issues described above are not a factor), and therefore all linkers (including the OSX linker) will pull in this .o file and find all the symbols that it needs. 2. Do not initialize the C variables corresponding to the Fortran common blocks in the global scope. Indeed, never initialize them at all (because we never need their *values* - we only check for their *locations*). Since nothing is ever written to these variables (particularly in the global scope), the linker does not see any alignment differences during initialization, but does make both the C and Fortran variables have the same addresses (this method has been working in LAM/MPI for over a decade). There were some comments here in the OMPI code base and in the LAM code base that stated/implied that C variables corresponding to Fortran common blocks had to have the same alignment as the Fortran common blocks (i.e., 16). There were attempts in both code bases to ensure that this was true. However, the attempts were wrong (in both code bases), and I have now read enough Fortran compiler documentation to convince myself that matching alignments is not required (indeed, it's beyond our control). As long as C variables corresponding to Fortran common blocks are not initialized in the global scope, the linker will "figure it out" and adjust the alignment to whatever is required (i.e., the greater of the alignments). Specifically (to counter comments that no longer exist in the OMPI code base but still exist in the LAM code base): - there is no need to make attempts to specially align C variables corresponding to Fortran common blocks - the types and sizes of C variables corresponding to Fortran common blocks should match, but do not need to be on any particular alignment Finally, as a side effect of this effort, I found a bunch of inconsistencies with the intent of status/array_of_statuses parameters. For all the functions that I modified they should be "out" (not inout). This commit was SVN r11057.
2006-07-31 19:07:09 +04:00
int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
{
int ret;
ompi_proc_t** procs;
size_t nprocs;
char *error = NULL;
bool timing = false;
int param, value;
struct timeval ompistart, ompistop;
char *slot_list = NULL;
#if 0
/* see comment below about sched_yield */
int num_processors;
#endif
/* Join the run-time environment - do the things that don't hit
the registry */
if (ORTE_SUCCESS != (ret = opal_init())) {
error = "ompi_mpi_init: opal_init failed";
goto error;
}
/* check to see if we want timing information */
param = mca_base_param_reg_int_name("ompi", "timing",
"Request that critical timing loops be measured",
false, false, 0, &value);
if (value != 0) {
timing = true;
gettimeofday(&ompistart, NULL);
}
/* Setup ORTE stage 1, note that we are not infrastructre */
if (ORTE_SUCCESS != (ret = orte_init(ORTE_NON_TOOL))) {
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
error = "ompi_mpi_init: orte_init failed";
goto error;
}
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
/* check for timing request - get stop time and report elapsed time if so */
if (timing) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init [%ld]: time from start to completion of orte_init %ld usec",
(long)ORTE_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
/* Once we've joined the RTE, see if any MCA parameters were
passed to the MPI level */
if (OMPI_SUCCESS != (ret = ompi_mpi_register_params())) {
error = "mca_mpi_register_params() failed";
goto error;
}
/* Setup process affinity */
if ( NULL != ( slot_list = getenv("slot_list"))) {
if (ORTE_SUCCESS != (ret = slot_list_to_cpu_set(slot_list))){
error = "ompi_mpi_init: error slot_list assigning";
goto error;
}
}
if (ompi_mpi_paffinity_alone) {
bool set = false;
param = mca_base_param_find("mpi", NULL, "paffinity_processor");
if (param >= 0) {
if (OMPI_SUCCESS == mca_base_param_lookup_int(param, &value)) {
if (value >= 0) {
opal_paffinity_base_cpu_set_t mpi_cpumask;
OPAL_PAFFINITY_CPU_ZERO(mpi_cpumask);
OPAL_PAFFINITY_CPU_SET(value,mpi_cpumask);
if (OPAL_SUCCESS == opal_paffinity_base_set(mpi_cpumask)) {
set = true;
}
}
}
if (!set) {
char *vpid;
orte_util_convert_vpid_to_string(&vpid, ORTE_PROC_MY_NAME->vpid);
opal_show_help("help-mpi-runtime",
"mpi_init:startup:paffinity-unavailable",
true, vpid);
free(vpid);
}
/* If we were able to set processor affinity, try setting
up memory affinity */
else {
if (OPAL_SUCCESS == opal_maffinity_base_open() &&
OPAL_SUCCESS == opal_maffinity_base_select()) {
ompi_mpi_maffinity_setup = true;
}
}
}
}
/* initialize datatypes. This step should be done early as it will
* create the local convertor and local arch used in the proc
* init.
*/
if (OMPI_SUCCESS != (ret = ompi_ddt_init())) {
error = "ompi_ddt_init() failed";
goto error;
}
/* Initialize OMPI procs */
if (OMPI_SUCCESS != (ret = ompi_proc_init())) {
error = "mca_proc_init() failed";
goto error;
}
/* initialize ops. This has to be done *after* ddt_init, but
befor mca_coll_base_open, since come collective modules
(e.g. the hierarchical) need them in the query function
*/
if (OMPI_SUCCESS != (ret = ompi_op_init())) {
error = "ompi_op_init() failed";
goto error;
}
/* Open up MPI-related MCA components */
if (OMPI_SUCCESS != (ret = mca_allocator_base_open())) {
error = "mca_allocator_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_rcache_base_open())) {
error = "mca_rcache_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_mpool_base_open())) {
error = "mca_mpool_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_pml_base_open())) {
error = "mca_pml_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_coll_base_open())) {
error = "mca_coll_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = ompi_osc_base_open())) {
error = "ompi_osc_base_open() failed";
goto error;
}
#if OPAL_ENABLE_FT == 1
if (OMPI_SUCCESS != (ret = ompi_crcp_base_open())) {
error = "ompi_crcp_base_open() failed";
goto error;
}
#endif
/* In order to reduce the common case for MPI apps (where they
don't use MPI-2 IO or MPI-1 topology functions), the io and
topo frameworks are initialized lazily, at the first use of
relevant functions (e.g., MPI_FILE_*, MPI_CART_*, MPI_GRAPH_*),
so they are not opened here. */
/* Select which MPI components to use */
if (OMPI_SUCCESS !=
(ret = mca_mpool_base_init(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_MPI_THREADS))) {
error = "mca_mpool_base_init() failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = mca_pml_base_select(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_MPI_THREADS))) {
error = "mca_pml_base_select() failed";
goto error;
}
/* select buffered send allocator component to be used */
ret=mca_pml_base_bsend_init(OMPI_ENABLE_MPI_THREADS);
if( OMPI_SUCCESS != ret ) {
error = "mca_pml_base_bsend_init() failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = mca_coll_base_find_available(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_MPI_THREADS))) {
error = "mca_coll_base_find_available() failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = ompi_osc_base_find_available(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_MPI_THREADS))) {
error = "ompi_osc_base_find_available() failed";
goto error;
}
#if OPAL_ENABLE_FT == 1
if (OMPI_SUCCESS != (ret = ompi_crcp_base_select() ) ) {
error = "ompi_crcp_base_select() failed";
goto error;
}
#endif
/* io and topo components are not selected here -- see comment
above about the io and topo frameworks being loaded lazily */
/* Initialize each MPI handle subsystem */
/* initialize requests */
if (OMPI_SUCCESS != (ret = ompi_request_init())) {
error = "ompi_request_init() failed";
goto error;
}
/* initialize info */
if (OMPI_SUCCESS != (ret = ompi_info_init())) {
error = "ompi_info_init() failed";
goto error;
}
/* initialize error handlers */
if (OMPI_SUCCESS != (ret = ompi_errhandler_init())) {
error = "ompi_errhandler_init() failed";
goto error;
}
/* initialize error codes */
if (OMPI_SUCCESS != (ret = ompi_mpi_errcode_init())) {
error = "ompi_mpi_errcode_init() failed";
goto error;
}
/* initialize internal error codes */
if (OMPI_SUCCESS != (ret = ompi_errcode_intern_init())) {
error = "ompi_errcode_intern_init() failed";
goto error;
}
/* initialize groups */
if (OMPI_SUCCESS != (ret = ompi_group_init())) {
error = "ompi_group_init() failed";
goto error;
}
/* initialize communicators */
if (OMPI_SUCCESS != (ret = ompi_comm_init())) {
error = "ompi_comm_init() failed";
goto error;
}
/* initialize file handles */
if (OMPI_SUCCESS != (ret = ompi_file_init())) {
error = "ompi_file_init() failed";
goto error;
}
/* initialize windows */
if (OMPI_SUCCESS != (ret = ompi_win_init())) {
error = "ompi_win_init() failed";
goto error;
}
/* initialize attribute meta-data structure for comm/win/dtype */
if (OMPI_SUCCESS != (ret = ompi_attr_init())) {
error = "ompi_attr_init() failed";
goto error;
}
/* check for timing request - get stop time and report elapsed time if so */
if (timing) {
gettimeofday(&ompistop, NULL);
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
opal_output(0, "ompi_mpi_init[%ld]: time from completion of orte_init to modex %ld usec",
(long)ORTE_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
/* exchange connection info - this function also acts as a barrier
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
* as it will not return until the exchange is complete
*/
if (OMPI_SUCCESS != (ret = orte_grpcomm.modex(NULL))) {
error = "orte_grpcomm_modex failed";
goto error;
}
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
if (timing) {
gettimeofday(&ompistop, NULL);
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
opal_output(0, "ompi_mpi_init[%ld]: time to execute modex %ld usec",
(long)ORTE_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
/* Fill in remote proc information */
if (OMPI_SUCCESS != (ret = ompi_proc_get_info())) {
ORTE_ERROR_LOG(ret);
error = "ompi_mpi_init: ompi_proc_get_info failed";
goto error;
}
/* Figure out the final MPI thread levels. If we were not
compiled for support for MPI threads, then don't allow
MPI_THREAD_MULTIPLE. */
ompi_mpi_thread_requested = requested;
if (OMPI_HAVE_THREAD_SUPPORT == 0) {
ompi_mpi_thread_provided = *provided = MPI_THREAD_SINGLE;
ompi_mpi_main_thread = NULL;
} else if (OMPI_ENABLE_MPI_THREADS == 1) {
ompi_mpi_thread_provided = *provided = requested;
ompi_mpi_main_thread = opal_thread_get_self();
} else {
if (MPI_THREAD_MULTIPLE == requested) {
ompi_mpi_thread_provided = *provided = MPI_THREAD_SERIALIZED;
} else {
ompi_mpi_thread_provided = *provided = requested;
}
ompi_mpi_main_thread = opal_thread_get_self();
}
ompi_mpi_thread_multiple = (ompi_mpi_thread_provided ==
MPI_THREAD_MULTIPLE);
if ((OMPI_ENABLE_PROGRESS_THREADS == 1) ||
(*provided != MPI_THREAD_SINGLE)) {
opal_set_using_threads(true);
}
/* start PML/BTL's */
ret = MCA_PML_CALL(enable(true));
if( OMPI_SUCCESS != ret ) {
error = "PML control failed";
goto error;
}
/* add all ompi_proc_t's to PML */
if (NULL == (procs = ompi_proc_world(&nprocs))) {
error = "ompi_proc_world() failed";
goto error;
}
ret = MCA_PML_CALL(add_procs(procs, nprocs));
free(procs);
if( OMPI_SUCCESS != ret ) {
error = "PML add procs failed";
goto error;
}
MCA_PML_CALL(add_comm(&ompi_mpi_comm_world));
MCA_PML_CALL(add_comm(&ompi_mpi_comm_self));
/* Init coll for the comms */
if (OMPI_SUCCESS !=
(ret = mca_coll_base_comm_select(MPI_COMM_WORLD))) {
error = "mca_coll_base_comm_select(MPI_COMM_WORLD) failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = mca_coll_base_comm_select(MPI_COMM_SELF))) {
error = "mca_coll_base_comm_select(MPI_COMM_SELF) failed";
goto error;
}
/*
* Dump all MCA parameters if requested
*/
if (ompi_mpi_show_mca_params) {
ompi_show_all_mca_params(ompi_mpi_comm_world.c_my_rank,
nprocs,
orte_system_info.nodename);
}
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component. This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done: As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in. In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in. The incoming changes revamp these procedures in three ways: 1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step. The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic. Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure. 2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed. The size of this data has been reduced in three ways: (a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes. To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose. (b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction. (c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using. While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly. 3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup. It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging. Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future. There are a few minor additional changes in the commit that I'll just note in passing: * propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details. * requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details. * cleanup of some stale header files This commit was SVN r16364.
2007-10-05 23:48:23 +04:00
/* wait for everyone to reach this point */
if (OMPI_SUCCESS != (ret = orte_grpcomm.barrier())) {
error = "orte_grpcomm_barrier failed";
goto error;
}
/* wire up the oob interface, if requested. Do this here because
it will go much faster before the event library is switched
into non-blocking mode */
if (OMPI_SUCCESS != (ret = ompi_init_preconnect_oob())) {
error = "ompi_mpi_do_preconnect_oob() failed";
goto error;
}
/* check for timing request - get stop time and report elapsed
time if so, then start the clock again */
if (timing) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init[%ld]: time from stage 2 cast to complete oob wireup %ld usec",
(long)ORTE_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
#if OMPI_ENABLE_PROGRESS_THREADS == 0
/* Start setting up the event engine for MPI operations. Don't
block in the event library, so that communications don't take
forever between procs in the dynamic code. This will increase
CPU utilization for the remainder of MPI_INIT when we are
blocking on ORTE-level events, but may greatly reduce non-TCP
latency. */
opal_progress_set_event_flag(OPAL_EVLOOP_NONBLOCK);
#endif
/* wire up the mpi interface, if requested. Do this after the
non-block switch for non-TCP performance. Do before the
polling change as anyone with a complex wire-up is going to be
using the oob. */
if (OMPI_SUCCESS != (ret = ompi_init_preconnect_mpi())) {
error = "ompi_mpi_do_preconnect_all() failed";
goto error;
}
/* Setup the publish/subscribe (PUBSUB) framework */
if (OMPI_SUCCESS != (ret = ompi_pubsub_base_open())) {
error = "ompi_pubsub_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = ompi_pubsub_base_select())) {
error = "ompi_pubsub_base_select() failed";
goto error;
}
/* Setup the dynamic process management (DPM) framework */
if (OMPI_SUCCESS != (ret = ompi_dpm_base_open())) {
error = "ompi_dpm_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = ompi_dpm_base_select())) {
error = "ompi_dpm_base_select() failed";
goto error;
}
/* Check whether we have been spawned or not. We introduce that
at the very end, since we need collectives, datatypes, ptls
etc. up and running here.... */
if (OMPI_SUCCESS != (ret = ompi_dpm.dyn_init())) {
error = "ompi_comm_dyn_init() failed";
goto error;
}
/*
* Startup the Checkpoint/Restart Mech.
* Note: Always do this so tools don't hang when
* in a non-checkpointable build
*/
if (OMPI_SUCCESS != (ret = ompi_cr_init())) {
error = "ompi_cr_init";
goto error;
}
/* Undo ORTE calling opal_progress_event_users_increment() during
MPI lifetime, to get better latency when not using TCP. Do
this *after* dyn_init, as dyn init uses lots of ORTE
communication and we don't want to hinder the performance of
that code. */
opal_progress_event_users_decrement();
/* see if the user specified yield_when_idle - if so, use it */
param = mca_base_param_find("mpi", NULL, "yield_when_idle");
mca_base_param_lookup_int(param, &value);
if (value < 0) {
/* TEMPORARY FIX - RIGHT NOW, WE DO NOT HAVE ACCESS TO
* INFO ON THE NUMBER OF LOCAL PROCS. THE ORTED IS SETTING
* THE MCA PARAM (OR THE PLS WILL, DEPENDING ON SYSTEM) SO
* THE FOLLOWING CODE WILL **NEVER** BE EXECUTED *EXCEPT*
* POSSIBLY BY SINGLETONS IN THE ABSENCE OF AN ENVIRO MCA PARAM
*/
#if 0
/* nope - so let's figure out what we can/should do...
* first, get the number of processors - if we can't then
* we can't do anything but set conservative values
*/
if (OPAL_SUCCESS == opal_get_num_processors(&num_processors)) {
/* got the num_processors - compare that to the number of
* local procs in this job to decide if we are oversubscribed
*/
if (ompi_proc_local_proc->num_local_procs > num_processors) {
/* oversubscribed - better yield */
opal_progress_set_yield_when_idle(true);
} else {
/* not oversubscribed - go ahead and be a hog! */
opal_progress_set_yield_when_idle(false);
}
} else {
/* couldn't get num_processors - be conservative */
opal_progress_set_yield_when_idle(true);
}
#endif
/* always just default to conservative */
opal_progress_set_yield_when_idle(true);
} else {
/* yep, they specified it - so set idle accordingly */
opal_progress_set_yield_when_idle(value == 0 ? false : true);
}
param = mca_base_param_find("mpi", NULL, "event_tick_rate");
mca_base_param_lookup_int(param, &value);
/* negative value means use default - just don't do anything */
if (value >= 0) {
opal_progress_set_event_poll_rate(value);
}
/* At this point, we are fully configured and in MPI mode. Any
communication calls here will work exactly like they would in
the user's code. Setup the connections between procs and warm
them up with simple sends, if requested */
error:
if (ret != OMPI_SUCCESS) {
const char *err_msg = opal_strerror(ret);
opal_show_help("help-mpi-runtime",
"mpi_init:startup:internal-failure", true,
"MPI_INIT", "MPI_INIT", error, err_msg, ret);
return ret;
}
/* Initialize the registered datarep list to be empty */
OBJ_CONSTRUCT(&ompi_registered_datareps, opal_list_t);
/* All done. Wasn't that simple? */
ompi_mpi_initialized = true;
/* Do we need to wait for a debugger? */
ompi_wait_for_debugger();
/* check for timing request - get stop time and report elapsed time if so */
if (timing) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init[%ld]: time from oob wireup to complete mpi_init %ld usec",
(long)ORTE_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
}
return MPI_SUCCESS;
}
/**
* This function receives a slot string ant translate it to
* cpu_set (long bitmap) using the PLPA module.
*/
static int socket_to_cpu_set(char **socket_list, int socket_cnt)
{
int i;
char **range;
int range_cnt;
int lower_range, upper_range;
int processor_id, num_processors;
int max_processor_id;
int rc;
opal_paffinity_base_cpu_set_t cpumask;
if (OPAL_SUCCESS != (rc = opal_paffinity_base_get_processor_info(&num_processors, &max_processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_ZERO(cpumask);
for (i=0; i<socket_cnt; i++) {
if (0 == strcmp("*", socket_list[i])) {
for ( processor_id=0; processor_id<=max_processor_id; processor_id++) {
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != ( rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runon cpu #%d (any socket)",(long)ORTE_PROC_MY_NAME->vpid, processor_id);
}
}
continue;
}
range = opal_argv_split(socket_list[i],'-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
processor_id = atoi(range[0]);
if (max_processor_id < processor_id) {
opal_output(0, "ERROR !!! max_processor_id (%d) < processor_id(%d), modify rankfile and run again\n",max_processor_id, processor_id);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != ( rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on cpu #%d", (long)ORTE_PROC_MY_NAME->vpid, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if (max_processor_id < upper_range || lower_range >= upper_range ) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_cpu) < %d , modify rankfile and run again\n",lower_range, max_processor_id, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for (processor_id=lower_range; processor_id<=upper_range; processor_id++) {
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on cpu #%d (%d-%d)",
(long)ORTE_PROC_MY_NAME->vpid, processor_id, lower_range, upper_range);
}
}
break;
default:
opal_argv_free(range);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
}
return ORTE_SUCCESS;
}
static int socket_core_to_cpu_set(char **socket_core_list, int socket_core_list_cnt)
{
int rc, i;
char **socket_core;
int socket_core_cnt;
char **range;
int range_cnt;
int lower_range, upper_range;
int socket, core, processor_id ;
int max_socket_num, max_core_num;
int num_sockets, num_cores;
opal_paffinity_base_cpu_set_t cpumask;
socket_core = opal_argv_split (socket_core_list[0], ':');
socket_core_cnt = opal_argv_count(socket_core);
OPAL_PAFFINITY_CPU_ZERO(cpumask);
socket = atoi(socket_core[0]);
if ( OPAL_SUCCESS != ( rc = opal_paffinity_base_get_socket_info(&num_sockets, &max_socket_num))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if ( max_socket_num < socket) {
opal_output(0,"ERROR !!! socket(%d) > max_socket_num(%d), modify rankfile and run again", socket, max_socket_num);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != ( rc = opal_paffinity_base_get_core_info(socket, &num_cores, &max_core_num))) {
opal_output(0,"Error !!! Invalid socket number (%d) in rankfile, modify rankfile and run again\n", socket);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if (0 == strcmp("*",socket_core[1])) {
for (core = 0; core <= max_core_num; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
} else {
range = opal_argv_split(socket_core[1], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
if ( max_core_num < core ) {
opal_output(0,"Error !!! core(%d) > max_core (%d) on socket %d, modify rankfile and run again\n",
core, max_core_num, socket);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if ( 0 > lower_range || max_core_num < upper_range || lower_range >= upper_range ) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_core) < %d ,modify rankfile and run again\n",
lower_range, max_core_num, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for (core=lower_range; core<=upper_range; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
opal_argv_free(socket_core);
}
for (i=1; i<socket_core_list_cnt; i++) {
socket_core = opal_argv_split (socket_core_list[i], ':');
socket_core_cnt = opal_argv_count(socket_core);
switch (socket_core_cnt) {
case 1:
range = opal_argv_split(socket_core[0], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
/* use PLPA to construct the child->cpu_set */
if ( max_core_num < core ) {
opal_output(0,"Error !!! max_core(%d) < core(%d), modify rankfile and run again\n",max_core_num, core);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
opal_output(0,"Error !!! Invalid socket : core pair ( #%d : %d), modify rankfile and run again\n",socket, core);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if ( 0 > lower_range || max_core_num < upper_range || lower_range >= upper_range) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_core) < %d, modify rankfile and run again\n",
lower_range, max_core_num, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for (core=lower_range; core<=upper_range; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
break;
case 2:
socket = atoi(socket_core[0]);
if (0 == strcmp("*",socket_core[1])) {
for (core=0; core<=max_core_num; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id ( socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
} else {
range = opal_argv_split(socket_core[1], '-');
range_cnt = opal_argv_count(range);
socket = atoi(socket_core[0]);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
if ( max_core_num < core ) {
opal_output(0,"Error !!! max_core(%d) < core(%d), modify rankfile and run again\n", max_core_num, core);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if ( 0 > lower_range || max_core_num < upper_range || lower_range > upper_range) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_core) < %d, modify rankfile and run again\n",
lower_range, max_core_num, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for ( core = lower_range; core <= upper_range; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %ld runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
}
break;
default:
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(socket_core);
}
return ORTE_SUCCESS;
}
static int slot_list_to_cpu_set(char *slot_str)
{
char **item;
char **socket_core;
orte_std_cntr_t item_cnt, socket_core_cnt;
int rc;
item = opal_argv_split (slot_str, ',');
item_cnt = opal_argv_count (item);
socket_core = opal_argv_split (item[0], ':');
socket_core_cnt = opal_argv_count(socket_core);
opal_argv_free(socket_core);
switch (socket_core_cnt) {
case 1:
if (ORTE_SUCCESS != (rc = socket_to_cpu_set(item, item_cnt))) {
opal_argv_free(item);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
break;
case 2:
if (ORTE_SUCCESS != (rc = socket_core_to_cpu_set(item, item_cnt))) {
opal_argv_free(item);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
break;
default:
opal_argv_free(item);
return ORTE_ERROR;
}
opal_argv_free(item);
return ORTE_SUCCESS;
}