attempt to reduce sync cost with openmp

CMakeLists.txt

@@ -157,10 +157,10 @@ endif ()
 #set(WCS_VERTEX_LIST_TYPE "::boost::listS"
 #  CACHE STRING "The BGL vertex list type to use in WCS")
 
-######################### For clang ###############################
 #set(WCS_OUT_EDGE_LIST_TYPE "::boost::vecS"
 #  CACHE STRING "The BGL out edge list type to use in WCS")
 
+######################### For clang ###############################
 #  Setup specific to using clang on Livermore Computing (LC) platforms
 #
 #  Currently, to use clang on Livermore Computing (LC) platforms with
@@ -563,6 +563,28 @@ if (WCS_HAS_METIS)
   list(APPEND WCS_EXEC_TARGETS ssa_partition-bin)
 endif (WCS_HAS_METIS)
 
+# add executable ssa_partition_inline
+if (WCS_HAS_METIS)
+  add_executable( ssa_partition_inline-bin src/ssa_partition_inline.cpp)
+  target_include_directories(ssa_partition_inline-bin PUBLIC
+    $<BUILD_INTERFACE:${CMAKE_BINARY_DIR}>
+    $<BUILD_INTERFACE:${CMAKE_BINARY_DIR}/src>
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_INCLUDEDIR}>)
+
+  target_include_directories(ssa_partition_inline-bin
+      PRIVATE $<TARGET_PROPERTY:wcs,INTERFACE_INCLUDE_DIRECTORIES>
+      ${Protobuf_DIR}/include ${Boost_INCLUDE_DIRS})
+
+  target_link_libraries(ssa_partition_inline-bin wcs ${LIB_FILESYSTEM}
+                        -L${Boost_LIBRARY_DIRS} boost_program_options)
+
+  set_target_properties(ssa_partition_inline-bin PROPERTIES OUTPUT_NAME ssa_partition_inline)
+  set_target_properties(ssa_partition_inline-bin PROPERTIES CMAKE_INSTALL_RPATH
+                        "${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}")
+  list(APPEND WCS_EXEC_TARGETS ssa_partition_inline-bin)
+endif (WCS_HAS_METIS)
+
 # add executable graphml2dot
 add_executable( graphml2dot-bin src/utils/graphml2dot.cpp )
 set_target_properties(graphml2dot-bin PROPERTIES OUTPUT_NAME graphml2dot)

src/sim_methods/ssa_nrm.cpp

@@ -456,6 +456,12 @@ bool SSA_NRM::advance_time_and_iter(const sim_time_t t_new)
   m_sim_time = t_new;
   return true;
 }
+
+void SSA_NRM::advance_step(const sim_time_t t_new)
+{
+  ++ m_sim_iter;
+  m_sim_time = t_new;
+}
 #endif // defined(_OPENMP) && defined(WCS_OMP_RUN_PARTITION)
 
 bool SSA_NRM::forward(const revent_t firing)

src/sim_methods/ssa_nrm.hpp

@@ -82,6 +82,7 @@ class SSA_NRM : public Sim_Method {
  #if defined(_OPENMP) && defined(WCS_OMP_RUN_PARTITION)
   /// Used in parallel mode with partitioned network
   bool advance_time_and_iter(const sim_time_t t_new);
+  void advance_step(const sim_time_t t_new);
   void update_reactions(const sim_time_t t_fired,
                         const Sim_Method::affected_reactions_t& affected,
                         reaction_times_t& affected_rtimes);

src/ssa_partition.cpp

@@ -224,6 +224,12 @@ void wcs_init(const wcs::SSA_Params& cfg, const partition_idx_t& parts)
 #endif
   omp_set_dynamic(0);
   omp_set_schedule(omp_sched_static, 0);
+
+  if (omp_get_max_threads() < shared_state.m_nparts) {
+    WCS_THROW("Set OMP_NUM_THREADS env variable larger than " \
+              + std::to_string(shared_state.m_nparts));
+    // The globl states have already been created.
+  }
   omp_set_num_threads(shared_state.m_nparts);
 
   shared_state.m_max_time = cfg.m_max_time;