Merge pull request #8037 from devreal/opal-thread-yield

devreal · web-flow · commit 46a4bc3de5ae · 2021-02-23T17:40:14.000+01:00
Move yield capability to opal thread component
diff --git a/ompi/runtime/ompi_mpi_params.c b/ompi/runtime/ompi_mpi_params.c
@@ -44,6 +44,8 @@
 #include "opal/util/show_help.h"
 #include "opal/runtime/opal.h"
 #include "opal/runtime/opal_params.h"
+#include "opal/mca/threads/threads.h"
+
 /*
  * Global variables
  *
@@ -63,7 +65,8 @@ bool ompi_mpi_keep_fqdn_hostnames = false;
 bool ompi_have_sparse_group_storage = OPAL_INT_TO_BOOL(OMPI_GROUP_SPARSE);
 bool ompi_use_sparse_group_storage = OPAL_INT_TO_BOOL(OMPI_GROUP_SPARSE);
 
-bool ompi_mpi_yield_when_idle = false;
+/* if the threads module requires yielding we use that as default but allow it to be overridden */
+bool ompi_mpi_yield_when_idle = OPAL_THREAD_YIELD_WHEN_IDLE_DEFAULT;
 int ompi_mpi_event_tick_rate = -1;
 char *ompi_mpi_show_mca_params_string = NULL;
 bool ompi_mpi_have_sparse_group_storage = !!(OMPI_GROUP_SPARSE);
@@ -152,7 +155,9 @@ int ompi_mpi_register_params(void)
                                  OPAL_INFO_LVL_9,
                                  MCA_BASE_VAR_SCOPE_READONLY,
                                  &ompi_mpi_oversubscribe);
-    ompi_mpi_yield_when_idle = ompi_mpi_oversubscribe;
+
+    /* yield if the node is oversubscribed and allow users to override */
+    ompi_mpi_yield_when_idle |= ompi_mpi_oversubscribe;
     (void) mca_base_var_register("ompi", "mpi", NULL, "yield_when_idle",
                                  "Yield the processor when waiting for MPI communication (for MPI processes, will default to 1 when oversubscribing nodes)",
                                  MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
diff --git a/opal/mca/threads/argobots/threads_argobots_threads.h b/opal/mca/threads/argobots/threads_argobots_threads.h
@@ -6,7 +6,7 @@
  * 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,
+ * Copyright (c) 2004-2020 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.
@@ -37,4 +37,14 @@ struct opal_thread_t {
     void *t_ret;
 };
 
+
+/* Argobots are cooperatively scheduled so yield when idle */
+#define OPAL_THREAD_YIELD_WHEN_IDLE_DEFAULT true
+
+static inline
+void opal_thread_yield(void)
+{
+    ABT_thread_yield();
+}
+
 #endif /* OPAL_MCA_THREADS_ARGOBOTS_THREADS_ARGOBOTS_THREADS_H */
diff --git a/opal/mca/threads/pthreads/Makefile.am b/opal/mca/threads/pthreads/Makefile.am
@@ -29,4 +29,6 @@ libmca_threads_pthreads_la_SOURCES = \
     threads_pthreads_threads.h \
     threads_pthreads_tsd.h \
     threads_pthreads_wait_sync.c \
-    threads_pthreads_wait_sync.h
+    threads_pthreads_wait_sync.h \
+    threads_pthreads_yield.c \
+    threads_pthreads.h
diff --git a/opal/mca/threads/pthreads/threads_pthreads.h b/opal/mca/threads/pthreads/threads_pthreads.h
@@ -0,0 +1,27 @@
+/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
+/*
+ * Copyright (c) 2020      High Performance Computing Center Stuttgart,
+ *                         University of Stuttgart.  All rights reserved.
+ *
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+
+#ifndef OPAL_MCA_THREADS_PTHREADS_THREADS_PTHREADS_H
+#define OPAL_MCA_THREADS_PTHREADS_THREADS_PTHREADS_H
+
+#include "opal_config.h"
+#include <stdint.h>
+#include <time.h>
+
+typedef void (opal_threads_pthreads_yield_fn_t)(void);
+
+OPAL_DECLSPEC int opal_threads_pthreads_yield_init(const mca_base_component_t *component);
+
+OPAL_DECLSPEC extern opal_threads_pthreads_yield_fn_t *opal_threads_pthreads_yield_fn;
+
+#endif /* OPAL_MCA_THREADS_PTHREADS_THREADS_PTHREADS_H */
diff --git a/opal/mca/threads/pthreads/threads_pthreads_component.c b/opal/mca/threads/pthreads/threads_pthreads_component.c
@@ -6,7 +6,7 @@
  * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
- * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
+ * Copyright (c) 2004-2020 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.
@@ -26,8 +26,11 @@
 #include "opal/mca/threads/thread.h"
 #include "opal/mca/threads/threads.h"
 #include "opal/constants.h"
+#include "opal/mca/threads/pthreads/threads_pthreads.h"
+
 
 static int opal_threads_pthreads_open(void);
+static int opal_threads_pthreads_register(void);
 
 const opal_threads_base_component_1_0_0_t mca_threads_pthreads_component = {
     /* First, the mca_component_t struct containing meta information
@@ -41,13 +44,19 @@ const opal_threads_base_component_1_0_0_t mca_threads_pthreads_component = {
                               OPAL_RELEASE_VERSION),
 
         .mca_open_component = opal_threads_pthreads_open,
+        .mca_register_component_params = opal_threads_pthreads_register
     },
     .threadsc_data = {
         /* The component is checkpoint ready */
         MCA_BASE_METADATA_PARAM_CHECKPOINT
     },
 };
 
+int opal_threads_pthreads_register(void)
+{
+    return opal_threads_pthreads_yield_init(&mca_threads_pthreads_component.threadsc_version);
+}
+
 int opal_threads_pthreads_open(void)
 {
     return OPAL_SUCCESS;
diff --git a/opal/mca/threads/pthreads/threads_pthreads_threads.h b/opal/mca/threads/pthreads/threads_pthreads_threads.h
@@ -6,7 +6,7 @@
  * Copyright (c) 2004-2006 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
- * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
+ * Copyright (c) 2004-2020 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.
@@ -30,11 +30,23 @@
 #include <pthread.h>
 #include <signal.h>
 
+#include "opal/mca/threads/threads.h"
+#include "opal/mca/threads/pthreads/threads_pthreads.h"
+
 struct opal_thread_t {
     opal_object_t super;
     opal_thread_fn_t t_run;
     void *t_arg;
     pthread_t t_handle;
 };
 
+/* Pthreads do not need to yield when idle */
+#define OPAL_THREAD_YIELD_WHEN_IDLE_DEFAULT false
+
+static inline
+void opal_thread_yield(void)
+{
+    opal_threads_pthreads_yield_fn();
+}
+
 #endif /* OPAL_MCA_THREADS_PTHREADS_THREADS_PTHREADS_THREADS_H */
diff --git a/opal/mca/threads/pthreads/threads_pthreads_yield.c b/opal/mca/threads/pthreads/threads_pthreads_yield.c
@@ -0,0 +1,89 @@
+/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
+/*
+ * Copyright (c) 2020      High Performance Computing Center Stuttgart,
+ *                         University of Stuttgart.  All rights reserved.
+ *
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+#include "opal_config.h"
+#include <time.h>
+#ifdef HAVE_SCHED_H
+#include <sched.h>
+#endif
+
+#include "opal/constants.h"
+#include "opal/mca/threads/thread.h"
+#include "opal/mca/threads/pthreads/threads_pthreads.h"
+
+static void opal_thread_pthreads_yield_sched_yield(void);
+static void opal_thread_pthreads_yield_nanosleep(void);
+
+typedef enum {
+    OPAL_PTHREADS_YIELD_SCHED_YIELD = 0,
+    OPAL_PTHREADS_YIELD_NANOSLEEP
+} opal_threads_pthreads_yield_strategy_t;
+
+static mca_base_var_enum_value_t yield_strategy_values[] = {
+    {OPAL_PTHREADS_YIELD_SCHED_YIELD, "sched_yield"},
+    {OPAL_PTHREADS_YIELD_NANOSLEEP,   "nanosleep"},
+    {0, NULL}};
+
+
+
+/* Number of nanoseconds to nanosleep, if enabled */
+static uint64_t yield_nsleep_nanosecs;
+/* The time to nanosleep, if enabled */
+static struct timespec yield_nsleep_time = {.tv_sec = 0, .tv_nsec = 1};
+static opal_threads_pthreads_yield_strategy_t yield_strategy = OPAL_PTHREADS_YIELD_SCHED_YIELD;
+
+opal_threads_pthreads_yield_fn_t *opal_threads_pthreads_yield_fn = &opal_thread_pthreads_yield_sched_yield;
+
+int opal_threads_pthreads_yield_init(const mca_base_component_t *component)
+{
+    mca_base_var_enum_t *yield_strategy_enumerator;
+    mca_base_var_enum_create("pthread_yield_strategies", yield_strategy_values, &yield_strategy_enumerator);
+
+    (void) mca_base_component_var_register(component, "yield_strategy",
+                                           "Pthread yield strategy to use",
+                                           MCA_BASE_VAR_TYPE_INT, yield_strategy_enumerator, 0, 0, OPAL_INFO_LVL_3,
+                                           MCA_BASE_VAR_SCOPE_LOCAL, &yield_strategy);
+    switch(yield_strategy) {
+        case OPAL_PTHREADS_YIELD_NANOSLEEP:
+            opal_threads_pthreads_yield_fn = &opal_thread_pthreads_yield_nanosleep;
+            break;
+        default:
+            /* use initial value */
+            break;
+    }
+
+    OBJ_RELEASE(yield_strategy_enumerator);
+
+    yield_nsleep_nanosecs = (yield_nsleep_time.tv_sec * 1E9) + yield_nsleep_time.tv_nsec;
+    (void) mca_base_component_var_register(component, "nanosleep_time",
+                                           "Number of nanoseconds to sleep when using nanosleep as the pthread yield strategy",
+                                           MCA_BASE_VAR_TYPE_UINT64_T, NULL, 0, 0, OPAL_INFO_LVL_3,
+                                           MCA_BASE_VAR_SCOPE_LOCAL, &yield_nsleep_nanosecs);
+    yield_nsleep_time.tv_sec  = yield_nsleep_nanosecs / 1E9;
+    yield_nsleep_time.tv_nsec = yield_nsleep_nanosecs - (uint64_t)(yield_nsleep_time.tv_sec * 1E9);
+
+    return OPAL_SUCCESS;
+
+}
+
+void opal_thread_pthreads_yield_sched_yield(void)
+{
+#ifdef HAVE_SCHED_H
+    sched_yield();
+#endif
+}
+
+void opal_thread_pthreads_yield_nanosleep(void)
+{
+    nanosleep(&yield_nsleep_time, NULL);
+}
+
diff --git a/opal/mca/threads/qthreads/threads_qthreads_threads.h b/opal/mca/threads/qthreads/threads_qthreads_threads.h
@@ -6,7 +6,7 @@
  * 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,
+ * Copyright (c) 2004-2020 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.
@@ -38,4 +38,13 @@ struct opal_thread_t {
     aligned_t *t_thread_ret_ptr;
 };
 
+/* Qthreads are cooperatively scheduled so yield when idle */
+#define OPAL_THREAD_YIELD_WHEN_IDLE_DEFAULT true
+
+static inline
+void opal_thread_yield(void)
+{
+    qthread_yield();
+}
+
 #endif /* OPAL_MCA_THREADS_QTHREADS_THREADS_QTHREADS_THREADS_H */
diff --git a/opal/mca/threads/threads.h b/opal/mca/threads/threads.h
@@ -133,6 +133,8 @@ OPAL_DECLSPEC opal_thread_t *opal_thread_get_self(void);
 OPAL_DECLSPEC void opal_thread_kill(opal_thread_t *, int sig);
 OPAL_DECLSPEC void opal_thread_set_main(void);
 
+static inline void opal_thread_yield(void);
+
 END_C_DECLS
 
 #endif /* OPAL_MCA_THREADS_THREADS_H */
diff --git a/opal/runtime/opal_progress.c b/opal/runtime/opal_progress.c
@@ -6,7 +6,7 @@
  * 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,
+ * Copyright (c) 2004-2020 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.
@@ -27,10 +27,6 @@
 
 #include "opal_config.h"
 
-#ifdef HAVE_SCHED_H
-#include <sched.h>
-#endif
-
 #include "opal/runtime/opal_progress.h"
 #include "opal/util/event.h"
 #include "opal/mca/base/mca_base_var.h"
@@ -39,6 +35,7 @@
 #include "opal/util/output.h"
 #include "opal/runtime/opal_params.h"
 #include "opal/runtime/opal.h"
+#include "opal/mca/threads/threads.h"
 
 #define OPAL_PROGRESS_USE_TIMERS (OPAL_TIMER_CYCLE_SUPPORTED || OPAL_TIMER_USEC_SUPPORTED)
 #define OPAL_PROGRESS_ONLY_USEC_NATIVE (OPAL_TIMER_USEC_NATIVE && !OPAL_TIMER_CYCLE_NATIVE)
@@ -68,7 +65,7 @@ static volatile opal_progress_callback_t *callbacks_lp = NULL;
 static size_t callbacks_lp_len = 0;
 static size_t callbacks_lp_size = 0;
 
-/* do we want to call sched_yield() if nothing happened */
+/* do we want to yield() if nothing happened */
 bool opal_progress_yield_when_idle = false;
 
 #if OPAL_PROGRESS_USE_TIMERS
@@ -212,7 +209,7 @@ static int opal_progress_events(void)
  * be called.  We don't propogate errors from the progress functions,
  * so no action is taken if they return failures.  The functions are
  * expected to return the number of events progressed, to determine
- * whether or not we should call sched_yield() during MPI progress.
+ * whether or not we should yield the CPU during MPI progress.
  * This is only losely tracked, as an error return can cause the number
  * of progressed events to appear lower than it actually is.  We don't
  * care, as the cost of that happening is far outweighed by the cost
@@ -246,16 +243,16 @@ opal_progress(void)
         opal_progress_events();
     }
 
-#if OPAL_HAVE_SCHED_YIELD
     if (opal_progress_yield_when_idle && events <= 0) {
         /* If there is nothing to do - yield the processor - otherwise
          * we could consume the processor for the entire time slice. If
          * the processor is oversubscribed - this will result in a best-case
          * latency equivalent to the time-slice.
+         * With some thread implementations, yielding might be required
+         * to ensure correct scheduling of all communicating threads.
          */
-        sched_yield();
+        opal_thread_yield();
     }
-#endif  /* defined(HAVE_SCHED_YIELD) */
 }
 
 
