Unify signal handling infrastructure (#1108)

* Unify signal handling infrastructure Fix bugs in parallel signal handling Add real-time checkpoint capability Add real-time heartbeat capability Fix bug when checkpointing a restarted run
sstsimulator · Jul 29, 2024 · 95aed59 · 95aed59
1 parent a4dbc4a
commit 95aed59
Show file tree

Hide file tree

Showing 43 changed files with 3,247 additions and 460 deletions.
diff --git a/src/sst/core/Makefile.am b/src/sst/core/Makefile.am
@@ -61,6 +61,8 @@ nobase_dist_sst_HEADERS = \
 	profile/syncProfileTool.h \
 	profile/componentProfileTool.h \
 	rankInfo.h \
+	realtime.h \
+	realtimeAction.h \
 	sparseVectorMap.h \
 	sst_types.h \
 	sstpart.h \
@@ -207,6 +209,7 @@ sst_core_sources = \
 	uninitializedQueue.cc \
 	unitAlgebra.cc \
 	module.cc \
+	realtime.cc \
 	ssthandler.cc \
 	sstpart.cc \
 	timeVortex.cc \

diff --git a/src/sst/core/cfgoutput/jsonConfigOutput.cc b/src/sst/core/cfgoutput/jsonConfigOutput.cc
@@ -149,18 +149,20 @@ JSONConfigGraphOutput::generate(const Config* cfg, ConfigGraph* graph)
     json::ordered_json outputJson;
 
     // Put in the program options
-    outputJson["program_options"]["verbose"]            = std::to_string(cfg->verbose());
-    outputJson["program_options"]["stop-at"]            = cfg->stop_at();
-    outputJson["program_options"]["print-timing-info"]  = cfg->print_timing() ? "true" : "false";
+    outputJson["program_options"]["verbose"]                = std::to_string(cfg->verbose());
+    outputJson["program_options"]["stop-at"]                = cfg->stop_at();
+    outputJson["program_options"]["print-timing-info"]      = cfg->print_timing() ? "true" : "false";
     // Ignore stopAfter for now
     // outputJson["program_options"]["stopAfter"] = cfg->stopAfterSec();
-    outputJson["program_options"]["heartbeat-period"]   = cfg->heartbeat_period();
-    outputJson["program_options"]["timebase"]           = cfg->timeBase();
-    outputJson["program_options"]["partitioner"]        = cfg->partitioner();
-    outputJson["program_options"]["timeVortex"]         = cfg->timeVortex();
-    outputJson["program_options"]["interthread-links"]  = cfg->interthread_links() ? "true" : "false";
-    outputJson["program_options"]["output-prefix-core"] = cfg->output_core_prefix();
-    outputJson["program_options"]["checkpoint-period"]  = cfg->checkpoint_period();
+    outputJson["program_options"]["heartbeat-sim-period"]   = cfg->heartbeat_sim_period();
+    outputJson["program_options"]["heartbeat-wall-period"]  = std::to_string(cfg->heartbeat_wall_period());
+    outputJson["program_options"]["timebase"]               = cfg->timeBase();
+    outputJson["program_options"]["partitioner"]            = cfg->partitioner();
+    outputJson["program_options"]["timeVortex"]             = cfg->timeVortex();
+    outputJson["program_options"]["interthread-links"]      = cfg->interthread_links() ? "true" : "false";
+    outputJson["program_options"]["output-prefix-core"]     = cfg->output_core_prefix();
+    outputJson["program_options"]["checkpoint-sim-period"]  = cfg->checkpoint_sim_period();
+    outputJson["program_options"]["checkpoint-wall-period"] = std::to_string(cfg->checkpoint_wall_period());
 
 
     // Put in the global param sets

diff --git a/src/sst/core/cfgoutput/pythonConfigOutput.cc b/src/sst/core/cfgoutput/pythonConfigOutput.cc
@@ -229,7 +229,10 @@ PythonConfigGraphOutput::generate(const Config* cfg, ConfigGraph* graph)
         outputFile, "sst.setProgramOption(\"print-timing-info\", \"%s\")\n", cfg->print_timing() ? "true" : "false");
     // Ignore stopAfter for now
     // fprintf(outputFile, "sst.setProgramOption(\"stopAfter\", \"%" PRIu32 "\")\n", cfg->stopAfterSec);
-    fprintf(outputFile, "sst.setProgramOption(\"heartbeat-period\", \"%s\")\n", cfg->heartbeat_period().c_str());
+    fprintf(
+        outputFile, "sst.setProgramOption(\"heartbeat-sim-period\", \"%s\")\n", cfg->heartbeat_sim_period().c_str());
+    fprintf(
+        outputFile, "sst.setProgramOption(\"heartbeat-wall-period\", \"%" PRIu32 "\")\n", cfg->heartbeat_wall_period());
     fprintf(outputFile, "sst.setProgramOption(\"timebase\", \"%s\")\n", cfg->timeBase().c_str());
     fprintf(outputFile, "sst.setProgramOption(\"partitioner\", \"%s\")\n", cfg->partitioner().c_str());
     fprintf(outputFile, "sst.setProgramOption(\"timeVortex\", \"%s\")\n", cfg->timeVortex().c_str());
@@ -238,7 +241,11 @@ PythonConfigGraphOutput::generate(const Config* cfg, ConfigGraph* graph)
         cfg->interthread_links() ? "true" : "false");
     fprintf(outputFile, "sst.setProgramOption(\"output-prefix-core\", \"%s\")\n", cfg->output_core_prefix().c_str());
 
-    fprintf(outputFile, "sst.setProgramOption(\"checkpoint-period\", \"%s\")\n", cfg->checkpoint_period().c_str());
+    fprintf(
+        outputFile, "sst.setProgramOption(\"checkpoint-sim-period\", \"%s\")\n", cfg->checkpoint_sim_period().c_str());
+    fprintf(
+        outputFile, "sst.setProgramOption(\"checkpoint-wall-period\", \"%" PRIu32 "\")\n",
+        cfg->checkpoint_wall_period());
 
     // Output the global params
     fprintf(outputFile, "# Define the global parameter sets:\n");

diff --git a/src/sst/core/checkpointAction.cc b/src/sst/core/checkpointAction.cc
@@ -28,13 +28,23 @@ REENABLE_WARNING
 
 namespace SST {
 
-CheckpointAction::CheckpointAction(Config* UNUSED(cfg), int this_rank, Simulation_impl* sim, TimeConverter* period) :
+CheckpointAction::CheckpointAction(
+    Config* UNUSED(cfg), RankInfo this_rank, Simulation_impl* sim, TimeConverter* period) :
     Action(),
-    rank(this_rank),
-    m_period(period)
+    rank_(this_rank),
+    period_(period),
+    generate_(false)
 {
-    sim->insertActivity(period->getFactor(), this);
-    if ( (0 == this_rank) ) { lastTime = sst_get_cpu_time(); }
+    next_sim_time_ = 0;
+    last_cpu_time_ = 0;
+
+    if ( period_ ) {
+        next_sim_time_ =
+            (period_->getFactor() * (sim->getCurrentSimCycle() / period_->getFactor())) + period_->getFactor();
+        sim->insertActivity(next_sim_time_, this);
+    }
+
+    if ( (0 == this_rank.rank) ) { last_cpu_time_ = sst_get_cpu_time(); }
 }
 
 CheckpointAction::~CheckpointAction() {}
@@ -43,20 +53,47 @@ void
 CheckpointAction::execute(void)
 {
     Simulation_impl* sim = Simulation_impl::getSimulation();
-    const double     now = sst_get_cpu_time();
+    createCheckpoint(sim);
+
+    next_sim_time_ += period_->getFactor();
+    sim->insertActivity(next_sim_time_, this);
+}
 
-    if ( 0 == rank ) {
+void
+CheckpointAction::createCheckpoint(Simulation_impl* sim)
+{
+
+    if ( 0 == rank_.rank ) {
+        const double now = sst_get_cpu_time();
         sim->getSimulationOutput().output(
             "# Simulation Checkpoint: Simulated Time %s (Real CPU time since last checkpoint %.5f seconds)\n",
-            sim->getElapsedSimTime().toStringBestSI().c_str(), (now - lastTime));
+            sim->getElapsedSimTime().toStringBestSI().c_str(), (now - last_cpu_time_));
 
-        lastTime = now;
+        last_cpu_time_ = now;
     }
 
     sim->checkpoint();
+}
 
-    SimTime_t next = sim->getCurrentSimCycle() + m_period->getFactor();
-    sim->insertActivity(next, this);
+void
+CheckpointAction::check()
+{
+    // TODO add logic for simulation-interval checkpoints in parallel
+    Simulation_impl* sim = Simulation_impl::getSimulation();
+    if ( generate_ ) { createCheckpoint(sim); }
+    generate_ = false;
+}
+
+void
+CheckpointAction::setCheckpoint()
+{
+    generate_ = true;
+}
+
+SimTime_t
+CheckpointAction::getNextCheckpointSimTime()
+{
+    return next_sim_time_;
 }
 
 } // namespace SST
diff --git a/src/sst/core/checkpointAction.h b/src/sst/core/checkpointAction.h
@@ -16,6 +16,7 @@
 #include "sst/core/config.h"
 #include "sst/core/cputimer.h"
 #include "sst/core/output.h"
+#include "sst/core/rankInfo.h"
 #include "sst/core/sst_types.h"
 
 #include <set>
@@ -35,22 +36,35 @@ class CheckpointAction : public Action
     /**
     Create a new checkpoint object for the simulation core to initiate checkpoints
     */
-    CheckpointAction(Config* cfg, int this_rank, Simulation_impl* sim, TimeConverter* period);
+    CheckpointAction(Config* cfg, RankInfo this_rank, Simulation_impl* sim, TimeConverter* period);
     ~CheckpointAction();
 
-    NotSerializable(SST::CheckpointAction) // Going to have to fix this
+    /** Generate a checkpoint next time check() is called */
+    void setCheckpoint();
 
-        private :
+    /** Called by TimeVortex to trigger checkpoint on simulation clock interval - not used in parallel simulation */
+    void execute(void) override;
 
-        CheckpointAction()
-    {}
+    /** Called by SyncManager to check whether a checkpoint should be generated */
+    void check();
+
+    /** Return next checkpoint time */
+    SimTime_t getNextCheckpointSimTime();
+
+    NotSerializable(SST::CheckpointAction);
+
+private:
+    CheckpointAction() {}
     CheckpointAction(const CheckpointAction&);
 
-    void           operator=(CheckpointAction const&);
-    void           execute(void) override;
-    int            rank;
-    TimeConverter* m_period;
-    double         lastTime;
+    void operator=(CheckpointAction const&);
+    void createCheckpoint(Simulation_impl* sim); // The actual checkpoint operation
+
+    RankInfo       rank_;          // RankInfo for this thread/rank
+    TimeConverter* period_;        // Simulation time interval for scheduling or nullptr if not set
+    double         last_cpu_time_; // Last time a checkpoint was triggered
+    bool           generate_;      // Whether a checkpoint should be done next time check() is called
+    SimTime_t      next_sim_time_; // Next simulationt ime a checkpoint should trigger at or 0 if not applicable
 };
 
 } // namespace SST