Enable auto chackplointing on SigTerm

Author: ferdonline

coreneuron/apps/main1.cpp

@@ -12,8 +12,9 @@
  * @brief File containing main driver routine for CoreNeuron
  */
 
-#include <cstring>
 #include <climits>
+#include <cstring>
+#include <csignal>
 #include <memory>
 #include <vector>
 
@@ -114,11 +115,24 @@ char* prepare_args(int& argc, char**& argv, int use_mpi, const char* arg) {
     // return actual data to be freed
     return first;
 }
-}
+
+}  // extern "C"
 
 namespace coreneuron {
+
 void call_prcellstate_for_prcellgid(int prcellgid, int compute_gpu, int is_init);
 
+
+static std::string check_restore() {
+    auto restore_path = corenrn_param.restorepath;
+    const auto auto_chkpt_path = corenrn_param.outpath + "/_corenrn_ckpt";
+    if (restore_path.empty() && fs_isdir(auto_chkpt_path.c_str())) {
+        restore_path = auto_chkpt_path;
+    }
+    return restore_path;
+}
+
+
 void nrn_init_and_load_data(int argc,
                             char* argv[],
                             bool is_mapping_needed = false,
@@ -168,7 +182,7 @@ void nrn_init_and_load_data(int argc,
     set_globals(corenrn_param.datpath.c_str(), (corenrn_param.seed >= 0), corenrn_param.seed);
 
     // set global variables for start time, timestep and temperature
-    std::string restore_path = corenrn_param.restorepath;
+    std::string restore_path = check_restore();
     t = restore_time(restore_path.c_str());
 
     if (corenrn_param.dt != -1000.) {  // command line arg highest precedence
@@ -402,6 +416,20 @@ std::unique_ptr<ReportHandler> create_report_handler(ReportConfiguration& config
     return report_handler;
 }
 
+/**
+ * \brief Installs a SIGTERM handler so that we finish the current simulation without losing data
+ * \return True if a checkpoint was performed. False otherwise (not enough elapsed time)
+ */
+static void install_sigterm_handler() {
+    auto sigh = [](int) {
+        std::cerr << "SIGTERM caught! Halting sim and dumping checkpoint" << std::endl;
+        coreneuron::stoprun = true;
+    };
+    if (std::signal(SIGTERM, sigh) == SIG_ERR) {
+        std::cerr << "Could not install SIGTERM handler" << std::endl;
+    }
+}
+
 }  // namespace coreneuron
 
 /// The following high-level functions are marked as "extern C"
@@ -482,6 +510,9 @@ extern "C" int run_solve_core(int argc, char** argv) {
     if (nrnmpi_myid == 0) {
         mkdir_p(output_dir.c_str());
     }
+
+    install_sigterm_handler();
+
 #if NRNMPI
     nrnmpi_barrier();
 #endif

coreneuron/io/file_utils.cpp

@@ -51,3 +51,13 @@ int mkdir_p(const char* path) {
     delete[] dirpath;
     return 0;
 }
+
+bool fs_exists(const char* path) {
+    struct stat buffer;
+    return (stat(path, &buffer) == 0);
+}
+
+bool fs_isdir(const char* path) {
+    struct stat buffer;
+    return (stat(path, &buffer) == 0 && S_ISDIR(buffer.st_mode));
+}

coreneuron/io/file_utils.hpp

@@ -21,4 +21,12 @@
  */
 int mkdir_p(const char* path);
 
+/** @brief Checks an arbitrary path exists
+ */
+bool fs_exists(const char* path);
+
+/** @brief Checks an arbitrary path is an existing directory
+ */
+bool fs_isdir(const char* path);
+
 #endif /* ifndef NRN_FILE_UTILS */

coreneuron/nrnconf.h

@@ -37,7 +37,7 @@ extern double pi;
 extern double t, dt;
 extern int rev_dt;
 extern int secondorder;
-extern bool stoprun;
+extern bool volatile stoprun;
 extern const char* bbcore_write_version;
 #define tstopbit   (1 << 15)
 #define tstopset   stoprun |= tstopbit

coreneuron/sim/fadvance_core.cpp

@@ -6,6 +6,7 @@
 # =============================================================================.
 */
 
+#include <cstdlib>
 #include <functional>
 
 #include "coreneuron/coreneuron.hpp"
@@ -23,12 +24,19 @@
 #include "coreneuron/utils/progressbar/progressbar.h"
 #include "coreneuron/utils/profile/profiler_interface.h"
 #include "coreneuron/io/nrn2core_direct.h"
+#include "coreneuron/io/nrn_checkpoint.hpp"
+
+// Do an auto checkpoint only if execution lasted longer than this var (secs)
+#define CHECKPOINT_MIN_RUNTIME (4 * 3600)  // 4h
 
 namespace coreneuron {
 
 extern corenrn_parameters corenrn_param;
 static void* nrn_fixed_step_thread(NrnThread*);
 static void* nrn_fixed_step_group_thread(NrnThread*, int, int, int&);
+static bool nrn_auto_checkpoint();
+static time_t sim_start_time;
+
 
 void dt2thread(double adt) { /* copied from nrnoc/fadvance.c */
     if (adt != nrn_threads[0]._dt) {
@@ -109,6 +117,7 @@ void nrn_fixed_single_steps_minimal(int total_sim_steps, double tstop) {
 #endif
         nrn_fixed_step_minimal();
         if (stoprun) {
+            nrn_auto_checkpoint();
             break;
         }
         current_steps++;
@@ -141,6 +150,7 @@ void nrn_fixed_step_group_minimal(int total_sim_steps) {
         nrn_flush_reports(nrn_threads[0]._t);
 #endif
         if (stoprun) {
+            nrn_auto_checkpoint();
             break;
         }
         current_steps++;
@@ -377,4 +387,26 @@ void* nrn_fixed_step_lastpart(NrnThread* nth) {
 
     return nullptr;
 }
+
+/**
+ * \brief Does a checkpoint of the simulation in enough time has passed
+ * \return True if a checkpoint was performed. False otherwise (not enough elapsed time)
+ */
+static bool nrn_auto_checkpoint() {
+    time_t cur_time = time(NULL);
+    int elapsed_secs = difftime(sim_start_time, cur_time);
+    if (elapsed_secs < CHECKPOINT_MIN_RUNTIME) {
+        return false;
+    }
+    // Write to tmp location first because allocated time may not be enough to complete
+    const auto ckpt_tmp = corenrn_param.outpath + "/_corenrn_ckpt_dirty",
+               ckpt_dir = corenrn_param.outpath + "/_corenrn_ckpt";
+    Instrumentor::phase p("Checkpointing");
+    write_checkpoint(nrn_threads, nrn_nthread, ckpt_tmp.c_str());
+    system(("/bin/rm -rf '" + ckpt_dir + "'; " + "/bin/mv '" + ckpt_tmp + "' '" + ckpt_dir + "'")
+               .c_str());
+    return true;
+}
+
+
 }  // namespace coreneuron

coreneuron/utils/nrnoc_aux.cpp

@@ -15,7 +15,7 @@
 #include "coreneuron/utils/nrnoc_aux.hpp"
 
 namespace coreneuron {
-bool stoprun;
+bool volatile stoprun;
 int v_structure_change;
 int diam_changed;
 #define MAXERRCOUNT 5