rpms/openmpi
7
0
Fork 0
Code Issues Pull Requests Releases Activity

Update to 4.1.1

Browse Source 
Enable psm2 support
fbtl-posix: link to common_ompio
Revert upstream commit c36d7459b6331c4d

Resolve: rhbz#1869443

Signed-off-by: Honggang Li <honli@redhat.com>
This commit is contained in:
Honggang Li 2021-07-15 21:38:30 -04:00
parent 4511aa6ee3
commit 5dda5cc9df
7 changed files with 418 additions and 1718 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
.gitignore vendored
View File

@ -54,3 +54,4 @@ openmpi-1.4.1-RH.tar.bz2
/openmpi-4.0.4.tar.bz2
/openmpi-4.0.5.tar.bz2
/openmpi-4.1.0.tar.bz2
/openmpi-4.1.1.tar.bz2

367
0001-Revert-ucx-check-supported-transports-and-devices-fo.patch Normal file
View File

@ -0,0 +1,367 @@
From 63c80c7692e55f634cbca6f67cc5c9cdef3a04d2 Mon Sep 17 00:00:00 2001
From: Honggang Li <honli@redhat.com>
Date: Mon, 28 Jun 2021 21:38:13 +0800
Subject: [PATCH] Revert "ucx: check supported transports and devices for
setting priority"
This reverts commit c36d7459b6331c4da825cad5a64326e7c1a272aa.
---
contrib/platform/mellanox/optimized.conf | 2 -
ompi/mca/pml/ucx/pml_ucx_component.c | 15 +-
opal/mca/common/ucx/common_ucx.c | 202 +----------------------
opal/mca/common/ucx/common_ucx.h | 15 --
opal/mca/common/ucx/configure.m4 | 2 -
5 files changed, 2 insertions(+), 234 deletions(-)
diff --git a/contrib/platform/mellanox/optimized.conf b/contrib/platform/mellanox/optimized.conf
index 543fd8d1e224..b86b37c9e2fa 100644
--- a/contrib/platform/mellanox/optimized.conf
+++ b/contrib/platform/mellanox/optimized.conf
@@ -61,8 +61,6 @@
coll = ^ml
hwloc_base_binding_policy = core
btl = self
-pml_ucx_tls = any
-pml_ucx_devices = any
# Basic behavior to smooth startup
mca_base_component_show_load_errors = 0
orte_abort_timeout = 10
diff --git a/ompi/mca/pml/ucx/pml_ucx_component.c b/ompi/mca/pml/ucx/pml_ucx_component.c
index 6aed6c41d11d..ed9cc6573e8e 100644
--- a/ompi/mca/pml/ucx/pml_ucx_component.c
+++ b/ompi/mca/pml/ucx/pml_ucx_component.c
@@ -107,26 +107,13 @@ static mca_pml_base_module_t*
mca_pml_ucx_component_init(int* priority, bool enable_progress_threads,
bool enable_mpi_threads)
{
- opal_common_ucx_support_level_t support_level;
int ret;
- support_level = opal_common_ucx_support_level(ompi_pml_ucx.ucp_context);
- if (support_level == OPAL_COMMON_UCX_SUPPORT_NONE) {
- return NULL;
- }
-
if ( (ret = mca_pml_ucx_init(enable_mpi_threads)) != 0) {
return NULL;
}
- /*
- * If found supported devices - set to the configured (high) priority.
- * Otherwise - Found only supported transports (which could be exposed by
- * unsupported devices), so set a priority lower than ob1.
- */
- *priority = (support_level == OPAL_COMMON_UCX_SUPPORT_DEVICE) ?
- ompi_pml_ucx.priority : 19;
- PML_UCX_VERBOSE(2, "returning priority %d", *priority);
+ *priority = ompi_pml_ucx.priority;
return &ompi_pml_ucx.super;
}
diff --git a/opal/mca/common/ucx/common_ucx.c b/opal/mca/common/ucx/common_ucx.c
index ac7a17d799a5..ae8e66877ab6 100644
--- a/opal/mca/common/ucx/common_ucx.c
+++ b/opal/mca/common/ucx/common_ucx.c
@@ -14,11 +14,8 @@
#include "opal/mca/base/mca_base_framework.h"
#include "opal/mca/pmix/pmix.h"
#include "opal/memoryhooks/memory.h"
-#include "opal/util/argv.h"
#include <ucm/api/ucm.h>
-#include <fnmatch.h>
-#include <stdio.h>
/***********************************************************************/
@@ -28,8 +25,7 @@ opal_common_ucx_module_t opal_common_ucx = {
.verbose = 0,
.progress_iterations = 100,
.registered = 0,
- .opal_mem_hooks = 0,
- .tls = NULL
+ .opal_mem_hooks = 0
};
static void opal_common_ucx_mem_release_cb(void *buf, size_t length,
@@ -40,15 +36,10 @@ static void opal_common_ucx_mem_release_cb(void *buf, size_t length,
OPAL_DECLSPEC void opal_common_ucx_mca_var_register(const mca_base_component_t *component)
{
- static const char *default_tls = "rc_verbs,ud_verbs,rc_mlx5,dc_mlx5,cuda_ipc,rocm_ipc";
- static const char *default_devices = "mlx*";
static int registered = 0;
static int hook_index;
static int verbose_index;
static int progress_index;
- static int tls_index;
- static int devices_index;
-
if (!registered) {
verbose_index = mca_base_var_register("opal", "opal_common", "ucx", "verbose",
"Verbose level of the UCX components",
@@ -69,29 +60,6 @@ OPAL_DECLSPEC void opal_common_ucx_mca_var_register(const mca_base_component_t *
OPAL_INFO_LVL_3,
MCA_BASE_VAR_SCOPE_LOCAL,
&opal_common_ucx.opal_mem_hooks);
-
- opal_common_ucx.tls = malloc(sizeof(*opal_common_ucx.tls));
- *opal_common_ucx.tls = strdup(default_tls);
- tls_index = mca_base_var_register("opal", "opal_common", "ucx", "tls",
- "List of UCX transports which should be supported on the system, to enable "
- "selecting the UCX component. Special values: any (any available). "
- "A '^' prefix negates the list. "
- "For example, in order to exclude on shared memory and TCP transports, "
- "please set to '^posix,sysv,self,tcp,cma,knem,xpmem'.",
- MCA_BASE_VAR_TYPE_STRING, NULL, 0, 0,
- OPAL_INFO_LVL_3,
- MCA_BASE_VAR_SCOPE_LOCAL,
- opal_common_ucx.tls);
-
- opal_common_ucx.devices = malloc(sizeof(*opal_common_ucx.devices));
- *opal_common_ucx.devices = strdup(default_devices);
- devices_index = mca_base_var_register("opal", "opal_common", "ucx", "devices",
- "List of device driver pattern names, which, if supported by UCX, will "
- "bump its priority above ob1. Special values: any (any available)",
- MCA_BASE_VAR_TYPE_STRING, NULL, 0, 0,
- OPAL_INFO_LVL_3,
- MCA_BASE_VAR_SCOPE_LOCAL,
- opal_common_ucx.devices);
registered = 1;
}
if (component) {
@@ -107,14 +75,6 @@ OPAL_DECLSPEC void opal_common_ucx_mca_var_register(const mca_base_component_t *
component->mca_type_name,
component->mca_component_name,
"opal_mem_hooks", 0);
- mca_base_var_register_synonym(tls_index, component->mca_project_name,
- component->mca_type_name,
- component->mca_component_name,
- "tls", 0);
- mca_base_var_register_synonym(devices_index, component->mca_project_name,
- component->mca_type_name,
- component->mca_component_name,
- "devices", 0);
}
}
@@ -163,166 +123,6 @@ OPAL_DECLSPEC void opal_common_ucx_mca_deregister(void)
opal_output_close(opal_common_ucx.output);
}
-#if HAVE_DECL_OPEN_MEMSTREAM
-static bool opal_common_ucx_check_device(const char *device_name, char **device_list)
-{
- char sysfs_driver_link[PATH_MAX];
- char driver_path[PATH_MAX];
- char *ib_device_name;
- char *driver_name;
- char **list_item;
- ssize_t ret;
-
- /* mlx5_0:1 */
- ret = sscanf(device_name, "%m[^:]%*d", &ib_device_name);
- if (ret != 1) {
- return false;
- }
-
- sysfs_driver_link[sizeof(sysfs_driver_link) - 1] = '\0';
- snprintf(sysfs_driver_link, sizeof(sysfs_driver_link) - 1,
- "/sys/class/infiniband/%s/device/driver", ib_device_name);
- free(ib_device_name);
-
- driver_path[sizeof(driver_path) - 1] = '\0';
- ret = readlink(sysfs_driver_link, driver_path, sizeof(driver_path) - 1);
- if (ret < 0) {
- MCA_COMMON_UCX_VERBOSE(2, "readlink(%s) failed: %s", sysfs_driver_link,
- strerror(errno));
- return false;
- }
-
- driver_name = basename(driver_path);
- for (list_item = device_list; *list_item != NULL; ++list_item) {
- if (!fnmatch(*list_item, driver_name, 0)) {
- MCA_COMMON_UCX_VERBOSE(2, "driver '%s' matched by '%s'",
- driver_path, *list_item);
- return true;
- }
- }
-
- return false;
-}
-#endif
-
-OPAL_DECLSPEC opal_common_ucx_support_level_t
-opal_common_ucx_support_level(ucp_context_h context)
-{
- opal_common_ucx_support_level_t support_level = OPAL_COMMON_UCX_SUPPORT_NONE;
- static const char *support_level_names[] = {
- [OPAL_COMMON_UCX_SUPPORT_NONE] = "none",
- [OPAL_COMMON_UCX_SUPPORT_TRANSPORT] = "transports only",
- [OPAL_COMMON_UCX_SUPPORT_DEVICE] = "transports and devices"
- };
-#if HAVE_DECL_OPEN_MEMSTREAM
- char *rsc_tl_name, *rsc_device_name;
- char **tl_list, **device_list, **list_item;
- bool is_any_tl, is_any_device;
- bool found_tl, negate;
- char line[128];
- FILE *stream;
- char *buffer;
- size_t size;
- int ret;
-#endif
-
- is_any_tl = !strcmp(*opal_common_ucx.tls, "any");
- is_any_device = !strcmp(*opal_common_ucx.devices, "any");
-
- /* Check for special value "any" */
- if (is_any_tl && is_any_device) {
- MCA_COMMON_UCX_VERBOSE(1, "ucx is enabled on any transport or device",
- *opal_common_ucx.tls);
- support_level = OPAL_COMMON_UCX_SUPPORT_DEVICE;
- goto out;
- }
-
-#if HAVE_DECL_OPEN_MEMSTREAM
- /* Split transports list */
- negate = ('^' == (*opal_common_ucx.tls)[0]);
- tl_list = opal_argv_split(*opal_common_ucx.tls + (negate ? 1 : 0), ',');
- if (tl_list == NULL) {
- MCA_COMMON_UCX_VERBOSE(1, "failed to split tl list '%s', ucx is disabled",
- *opal_common_ucx.tls);
- goto out;
- }
-
- /* Split devices list */
- device_list = opal_argv_split(*opal_common_ucx.devices, ',');
- if (device_list == NULL) {
- MCA_COMMON_UCX_VERBOSE(1, "failed to split devices list '%s', ucx is disabled",
- *opal_common_ucx.devices);
- goto out_free_tl_list;
- }
-
- /* Open memory stream to dump UCX information to */
- stream = open_memstream(&buffer, &size);
- if (stream == NULL) {
- MCA_COMMON_UCX_VERBOSE(1, "failed to open memory stream for ucx info (%s), "
- "ucx is disabled", strerror(errno));
- goto out_free_device_list;
- }
-
- /* Print ucx transports information to the memory stream */
- ucp_context_print_info(context, stream);
-
- /* Rewind and read transports/devices list from the stream */
- fseek(stream, 0, SEEK_SET);
- while ((support_level != OPAL_COMMON_UCX_SUPPORT_DEVICE) &&
- (fgets(line, sizeof(line), stream) != NULL)) {
- rsc_tl_name = NULL;
- ret = sscanf(line,
- /* "# resource 6 : md 5 dev 4 flags -- rc_verbs/mlx5_0:1" */
- "# resource %*d : md %*d dev %*d flags -- %m[^/ \n\r]/%m[^/ \n\r]",
- &rsc_tl_name, &rsc_device_name);
- if (ret != 2) {
- free(rsc_tl_name);
- continue;
- }
-
- /* Check if 'rsc_tl_name' is found provided list */
- found_tl = is_any_tl;
- for (list_item = tl_list; !found_tl && (*list_item != NULL); ++list_item) {
- found_tl = !strcmp(*list_item, rsc_tl_name);
- }
-
- /* Check if the transport has a match (either positive or negative) */
- assert(!(is_any_tl && negate));
- if (found_tl != negate) {
- if (is_any_device ||
- opal_common_ucx_check_device(rsc_device_name, device_list)) {
- MCA_COMMON_UCX_VERBOSE(2, "%s/%s: matched both transport and device list",
- rsc_tl_name, rsc_device_name);
- support_level = OPAL_COMMON_UCX_SUPPORT_DEVICE;
- } else {
- MCA_COMMON_UCX_VERBOSE(2, "%s/%s: matched transport list but not device list",
- rsc_tl_name, rsc_device_name);
- support_level = OPAL_COMMON_UCX_SUPPORT_TRANSPORT;
- }
- } else {
- MCA_COMMON_UCX_VERBOSE(2, "%s/%s: did not match transport list",
- rsc_tl_name, rsc_device_name);
- }
-
- free(rsc_device_name);
- free(rsc_tl_name);
- }
-
- MCA_COMMON_UCX_VERBOSE(2, "support level is %s", support_level_names[support_level]);
- fclose(stream);
- free(buffer);
-
-out_free_device_list:
- opal_argv_free(device_list);
-out_free_tl_list:
- opal_argv_free(tl_list);
-out:
-#else
- MCA_COMMON_UCX_VERBOSE(2, "open_memstream() was not found, ucx is disabled");
-#endif
- return support_level;
-}
-
void opal_common_ucx_empty_complete_cb(void *request, ucs_status_t status)
{
}
diff --git a/opal/mca/common/ucx/common_ucx.h b/opal/mca/common/ucx/common_ucx.h
index 92cdd738ef98..202131ac8907 100644
--- a/opal/mca/common/ucx/common_ucx.h
+++ b/opal/mca/common/ucx/common_ucx.h
@@ -88,8 +88,6 @@ typedef struct opal_common_ucx_module {
int progress_iterations;
int registered;
bool opal_mem_hooks;
- char **tls;
- char **devices;
} opal_common_ucx_module_t;
typedef struct opal_common_ucx_del_proc {
@@ -97,23 +95,10 @@ typedef struct opal_common_ucx_del_proc {
size_t vpid;
} opal_common_ucx_del_proc_t;
-typedef enum {
- /* No supported transports found (according to configured list of supported
- transports) */
- OPAL_COMMON_UCX_SUPPORT_NONE,
-
- /* Have supported transports but not supported devices */
- OPAL_COMMON_UCX_SUPPORT_TRANSPORT,
-
- /* Have both supported transports and supported devices */
- OPAL_COMMON_UCX_SUPPORT_DEVICE,
-} opal_common_ucx_support_level_t;
-
extern opal_common_ucx_module_t opal_common_ucx;
OPAL_DECLSPEC void opal_common_ucx_mca_register(void);
OPAL_DECLSPEC void opal_common_ucx_mca_deregister(void);
-OPAL_DECLSPEC opal_common_ucx_support_level_t opal_common_ucx_support_level(ucp_context_h context);
OPAL_DECLSPEC void opal_common_ucx_mca_proc_added(void);
OPAL_DECLSPEC void opal_common_ucx_empty_complete_cb(void *request, ucs_status_t status);
OPAL_DECLSPEC int opal_common_ucx_mca_pmix_fence(ucp_worker_h worker);
diff --git a/opal/mca/common/ucx/configure.m4 b/opal/mca/common/ucx/configure.m4
index af8628a889c6..27e07c2005b2 100644
--- a/opal/mca/common/ucx/configure.m4
+++ b/opal/mca/common/ucx/configure.m4
@@ -18,8 +18,6 @@ AC_DEFUN([MCA_opal_common_ucx_CONFIG],[
[common_ucx_happy="yes"],
[common_ucx_happy="no"])
- AC_CHECK_DECLS([open_memstream], [], [], [[#include <stdio.h>]])
-
AS_IF([test "$common_ucx_happy" = "yes"],
[$1],
[$2])
--
2.31.1

33
266189935aef4fce825d0db831b4b53accc62c32.patch Normal file
View File

@ -0,0 +1,33 @@
From 266189935aef4fce825d0db831b4b53accc62c32 Mon Sep 17 00:00:00 2001
From: Jeff Squyres <jsquyres@cisco.com>
Date: Tue, 22 Jun 2021 22:28:37 -0400
Subject: [PATCH] fbtl-posix: link to common_ompio
The posix fbtl calls mca_common_ompio_progress(), which resides in
common/ompio (i.e., libmca_common_ompio.la). So add that into
mca_fbtl_posix_la_LIBADD (like we do in a few other OMPIO-based
components). Failure to do this *can* lead to the posix fbtl
component failing to load (depending on whether other OMPIO-based
components that pull in libmca_common_ompio were loaded first).
Thanks to Honggang Li for raising the issue.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
---
ompi/mca/fbtl/posix/Makefile.am | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/ompi/mca/fbtl/posix/Makefile.am b/ompi/mca/fbtl/posix/Makefile.am
index a7b0624d3ec..1ce19cb09b7 100644
--- a/ompi/mca/fbtl/posix/Makefile.am
+++ b/ompi/mca/fbtl/posix/Makefile.am
@@ -34,7 +34,8 @@ mcacomponentdir = $(ompilibdir)
mcacomponent_LTLIBRARIES = $(component_install)
mca_fbtl_posix_la_SOURCES = $(sources)
mca_fbtl_posix_la_LDFLAGS = -module -avoid-version
-mca_fbtl_posix_la_LIBADD = $(top_builddir)/ompi/lib@OMPI_LIBMPI_NAME@.la
+mca_fbtl_posix_la_LIBADD = $(top_builddir)/ompi/lib@OMPI_LIBMPI_NAME@.la \
+ $(OMPI_TOP_BUILDDIR)/ompi/mca/common/ompio/libmca_common_ompio.la
noinst_LTLIBRARIES = $(component_noinst)
libmca_fbtl_posix_la_SOURCES = $(sources)

1682
8322.patch
View File

File diff suppressed because it is too large Load Diff

25
8348.patch
View File

@ -1,25 +0,0 @@
From 838568da9fce85b4555b0e0cbd899c8e8ef75696 Mon Sep 17 00:00:00 2001
From: George Bosilca <bosilca@icl.utk.edu>
Date: Wed, 6 Jan 2021 13:30:40 -0500
Subject: [PATCH] A started generalized request should be marked as pending.
Fixes #8340
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
(cherry picked from commit 434a2515f8aab11f505b2fca0b3d8cc41e24cef2)
---
ompi/request/grequest.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/ompi/request/grequest.c b/ompi/request/grequest.c
index c895b4232b6..02affd642aa 100644
--- a/ompi/request/grequest.c
+++ b/ompi/request/grequest.c
@@ -163,6 +163,7 @@ int ompi_grequest_start(
greq->greq_free.c_free = gfree_fn;
greq->greq_cancel.c_cancel = gcancel_fn;
greq->greq_base.req_status = ompi_status_empty;
+ greq->greq_base.req_complete = REQUEST_PENDING;
*request = &greq->greq_base;
return OMPI_SUCCESS;

26
openmpi.spec
View File

@ -26,12 +26,11 @@
%endif
# Run autogen - needed for some patches
# For Patch0
%bcond_without autogen
Name: openmpi%{?_cc_name_suffix}
Version: 4.1.0
Release: 7%{?dist}
Version: 4.1.1
Release: 2%{?dist}
Summary: Open Message Passing Interface
License: BSD and MIT and Romio
URL: http://www.open-mpi.org/
@ -42,11 +41,8 @@ Source1: openmpi.module.in
Source2: openmpi.pth.py2
Source3: openmpi.pth.py3
Source4: macros.openmpi
# Fix AVX library linkage
Patch0: https://patch-diff.githubusercontent.com/raw/open-mpi/ompi/pull/8322.patch
# Fix generalized requests (mpi4py test failure)
Patch1: https://patch-diff.githubusercontent.com/raw/open-mpi/ompi/pull/8348.patch
Patch1: 266189935aef4fce825d0db831b4b53accc62c32.patch
Patch2: 0001-Revert-ucx-check-supported-transports-and-devices-fo.patch
BuildRequires: gcc-c++
BuildRequires: gcc-gfortran
@ -82,6 +78,9 @@ BuildRequires: perl-interpreter
BuildRequires: perl(Getopt::Long)
BuildRequires: pmix-devel
BuildRequires: python%{python3_pkgversion}-devel
%ifarch x86_64
BuildRequires: libpsm2-devel
%endif
%if %{with ucx}
BuildRequires: ucx-devel
%endif
@ -91,7 +90,7 @@ BuildRequires: rpm-mpi-hooks
%endif
Provides: mpi
%if 0%{?rhel}
%if 0%{?rhel} == 7
# Need this for /etc/profile.d/modules.sh
Requires: environment-modules
%endif
@ -169,7 +168,7 @@ OpenMPI support for Python 3.
%prep
%autosetup -p1
%autosetup -p1 -n %{name}-%{version}
%if %{with autogen}
./autogen.pl --force
%endif
@ -358,6 +357,13 @@ make check
%changelog
* Thu Jul 15 2021 Honggang Li <honli@redhat.com> - 4.1.1-2
- Update to 4.1.1
- Enable psm2 support
- fbtl-posix: link to common_ompio
- Revert upstream commit c36d7459b6331c4d
- Resolve: rhbz#1869443
* Fri Apr 16 2021 Mohan Boddu <mboddu@redhat.com> - 4.1.0-7
- Rebuilt for RHEL 9 BETA on Apr 15th 2021. Related: rhbz#1947937

2
sources
View File

@ -1 +1 @@
SHA512 (openmpi-4.1.0.tar.bz2) = eaf086ab4929ce5a9a3e867c8315bf802ff4dc75d3f05d740e22dfd97803a4559212dacbe06920d42ac6644f46057eb6980cccf5a8b0a7df9c5bdf5bffc0b3a6
SHA512 (openmpi-4.1.1.tar.bz2) = 0d85ba45a40c0879f266e5286615e2cf94eb3570f0a705194525821d5c85d460cefc3a2da8207e6e84c479d3d0da656e2342cc2d6f88c4b4577ca22bbeacc89d