diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6eab72be46..86bdc45019 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -986,14 +986,23 @@ if(${TARGET} STREQUAL "charm4py")
     target_link_libraries(charm cudart cudahybridapi)
   endif()
 
+  if(${TRACING})
+    if(CMAKE_SYSTEM_NAME STREQUAL "Darwin")
+      target_link_libraries(charm trace-summary trace-projections "-Llib/ -standalone -all_load -c++stl -shared")
+    else()
+      target_link_libraries(charm trace-summary trace-projections "-Llib/ -standalone -whole-archive -c++stl -shared")
+    endif()
+  endif()
+
   add_dependencies(charm hwloc)
-endif()
 
-# Check that we are able to build and link an executable
-add_executable(ckhello ${CMAKE_SOURCE_DIR}/tests/charm++/simplearrayhello/hello.C)
-add_dependencies(ckhello ck ldb-rand memory-default threads-default conv-static
-  converse ckmain ckqt
-  moduleNDMeshStreamer modulecompletion)
+else()
+  # Check that we are able to build and link an executable
+  add_executable(ckhello ${CMAKE_SOURCE_DIR}/tests/charm++/simplearrayhello/hello.C)
+  add_dependencies(ckhello ck ldb-none memory-default threads-default conv-static
+    converse ckmain ckqt
+    moduleNDMeshStreamer modulecompletion)
+endif()
 
 # Create conv-mach-opt.sh
 set(optfile_sh ${CMAKE_BINARY_DIR}/include/conv-mach-opt.sh)
diff --git a/src/ck-core/charm.h b/src/ck-core/charm.h
index cee65d539a..aee83ad923 100644
--- a/src/ck-core/charm.h
+++ b/src/ck-core/charm.h
@@ -148,6 +148,9 @@ extern void CkStartQDExt_ArrayCallback(int aid, int* idx, int ndims, int epIdx,
 extern void CkStartQDExt_SectionCallback(int sid_pe, int sid_cnt, int rootPE, int ep);
 extern void registerCreateCallbackMsgExtCallback(void (*cb)(void*, int, int, int, int *, char**, int*));
 extern void registerPyReductionExtCallback(int (*cb)(char**, int*, int, char**));
+extern int CkTraceRegisterUserEvent(char *EventDesc, int eventID);
+extern void CkTraceBeginUserBracketEvent(int eventID);
+extern void CkTraceEndUserBracketEvent(int eventID);
 
 #endif
 /*********************************************************/
@@ -226,11 +229,12 @@ extern void CkRegisterBase(int derivedIdx, int baseIdx);
 /** Sets the ifNotThere policy of an EP **/
 extern void CkRegisterIfNotThere(int epIdx, CkArray_IfNotThere policy);
 #if CMK_CHARM4PY
-extern void CkRegisterMainChareExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx);
-extern void CkRegisterGroupExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx);
-extern void CkRegisterSectionManagerExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx);
-extern void CkRegisterArrayMapExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx);
-extern void CkRegisterArrayExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx);
+void CkRegisterMainChareExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx);
+extern void CkRegisterGroupExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx);
+extern void CkRegisterGroupExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx);
+extern void CkRegisterSectionManagerExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx);
+extern void CkRegisterArrayMapExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx);
+extern void CkRegisterArrayExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx);
 extern void CkRegisterReadonlyExt(const char *name, const char *type, size_t msgSize, char *msg);
 #endif
 
diff --git a/src/ck-core/ck.C b/src/ck-core/ck.C
index 33bfc82bcf..8ab873db9b 100644
--- a/src/ck-core/ck.C
+++ b/src/ck-core/ck.C
@@ -2637,6 +2637,19 @@ void CkArrayExtSend_multi(int aid, int *idx, int ndims, int epIdx, int num_bufs,
   }
 }
 
+int CkTraceRegisterUserEvent(char *EventDesc, int eventNum)
+{
+  return traceRegisterUserEvent(EventDesc, eventNum);
+}
+
+void CkTraceBeginUserBracketEvent(int eventID)
+{
+  traceBeginUserBracketEvent(eventID);
+}
+void CkTraceEndUserBracketEvent(int eventID)
+{
+  traceEndUserBracketEvent(eventID);
+}
 #endif
 
 //------------------- Message Watcher (record/replay) ----------------
diff --git a/src/ck-core/init.C b/src/ck-core/init.C
index 728466dfc3..cd12163a40 100644
--- a/src/ck-core/init.C
+++ b/src/ck-core/init.C
@@ -90,6 +90,21 @@ extern void QdCreate(int);
 extern void QdProcess(int);
 #endif
 
+#if CMK_CHARM4PY && CMK_TRACE_ENABLED
+extern void _registerTraceProjections();
+extern void _registerTraceSummary();
+extern void _createTraceprojections(char **argv);
+extern void _createTracesummary(char **argv);
+
+void _createTraces(char **argv) { (void)argv;
+_createTraceprojections(argv);
+_createTracesummary(argv);
+}
+
+#elif CMK_CHARM4PY
+void _createTraces(char **argv){};
+#endif
+
 void CkRestartMain(const char* dirname, CkArgMsg* args);
 
 #define  DEBUGF(x)     //CmiPrintf x;
@@ -1524,14 +1539,21 @@ void _initCharm(int unused_argc, char **argv)
 		_registerCkMemCheckpoint();
 #endif
 #if CMK_CHARM4PY
-                /**
-                  Load balancers are currently registered in Charm++ through a C file that is generated and
-                  and compiled by charmc when making an executable. That file contains appropriate calls to
-                  register whatever load balancers are being linked in.
-                  Without an executable (charm4py just uses libcharm.so), the load balancers in libcharm.so
-                  have to somehow be registered during init.
-                */
-		_registerTreeLB();
+    /**
+       Load balancers are currently registered in Charm++ through a C file that is generated and
+       and compiled by charmc when making an executable. That file contains appropriate calls to
+       register whatever load balancers are being linked in.
+       Without an executable (charm4py just uses libcharm.so), the load balancers in libcharm.so
+       have to somehow be registered during init.
+    */
+    _registerTreeLB();
+
+#if CMK_TRACE_ENABLED
+      _registerTraceProjections();
+      _registerTraceSummary();
+#endif
+
+
 #endif
 
 		/**
diff --git a/src/ck-core/register.C b/src/ck-core/register.C
index 11604725f0..0274c30543 100644
--- a/src/ck-core/register.C
+++ b/src/ck-core/register.C
@@ -33,6 +33,7 @@ void _registerInit(void)
 int CkRegisterMsg(const char *name, CkPackFnPtr pack, CkUnpackFnPtr unpack,
                   CkDeallocFnPtr dealloc, size_t size)
 {
+  auto messageSize = _msgTable.size();
   return _msgTable.add(new MsgInfo(name, pack, unpack, dealloc, size));
 }
 
@@ -101,37 +102,41 @@ void CkRegisterGroupIrr(int chareIndex,int isIrr){
 // TODO give a unique name to entry methods when calling CkRegisterEp
 // (no need has appeared for this so this is very low priority)
 
-void CkRegisterGroupExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx) {
+void CkRegisterGroupExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx) {
   int __idx = CkRegisterChare(s, sizeof(GroupExt), TypeGroup);
   CkRegisterBase(__idx, CkIndex_IrrGroup::__idx);
 
   int epIdxCtor = CkRegisterEp(s, GroupExt::__GroupExt, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
   CkRegisterDefaultCtor(__idx, epIdxCtor);
 
-  for (int i=1; i < numEntryMethods; i++)
-    CkRegisterEp(s, GroupExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+  for (int i=emNamesStart + 1; i < emNamesStart+numEntryMethods; i++)
+    {
+      CkRegisterEp(emNames[i], GroupExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+    }
 
   *chareIdx = __idx;
   *startEpIdx = epIdxCtor;
 }
 
-void CkRegisterSectionManagerExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx) {
+void CkRegisterSectionManagerExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx) {
   int __idx = CkRegisterChare(s, sizeof(SectionManagerExt), TypeGroup);
   CkRegisterBase(__idx, CkIndex_IrrGroup::__idx);
 
   int epIdxCtor = CkRegisterEp(s, SectionManagerExt::__SectionManagerExt, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
   CkRegisterDefaultCtor(__idx, epIdxCtor);
 
-  CkRegisterEp(s, SectionManagerExt::__sendToSection, CkMarshallMsg::__idx, __idx, 0);
+  CkRegisterEp("sendToSection", SectionManagerExt::__sendToSection, CkMarshallMsg::__idx, __idx, 0);
 
-  for (int i=2; i < numEntryMethods; i++)
-    CkRegisterEp(s, SectionManagerExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+  for (int i=emNamesStart+2; i < numEntryMethods+emNamesStart; i++)
+    {
+      CkRegisterEp(emNames[i], SectionManagerExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+    }
 
   *chareIdx = __idx;
   *startEpIdx = epIdxCtor;
 }
 
-void CkRegisterArrayMapExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx) {
+void CkRegisterArrayMapExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx) {
   int __idx = CkRegisterChare(s, sizeof(ArrayMapExt), TypeGroup);
   CkRegisterBase(__idx, CkIndex_IrrGroup::__idx);
 
@@ -139,13 +144,13 @@ void CkRegisterArrayMapExt(const char *s, int numEntryMethods, int *chareIdx, in
   CkRegisterDefaultCtor(__idx, epIdxCtor);
 
   for (int i=1; i < numEntryMethods; i++)
-    CkRegisterEp(s, ArrayMapExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+    CkRegisterEp(emNames[i], ArrayMapExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
 
   *chareIdx = __idx;
   *startEpIdx = epIdxCtor;
 }
 
-void CkRegisterArrayExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx) {
+void CkRegisterArrayExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx) {
   int __idx = CkRegisterChare(s, sizeof(ArrayElemExt), TypeArray);
   CkRegisterBase(__idx, CkIndex_ArrayElement::__idx);
 
@@ -157,8 +162,10 @@ void CkRegisterArrayExt(const char *s, int numEntryMethods, int *chareIdx, int *
 
   CkRegisterEp(s, ArrayElemExt::__AtSyncEntryMethod, 0, __idx, 0);
 
-  for (int i=3; i < numEntryMethods; i++)
-    CkRegisterEp(s, ArrayElemExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+  for (int i=emNamesStart+3; i < emNamesStart+numEntryMethods; i++)
+    {
+      CkRegisterEp(emNames[i], ArrayElemExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+    }
 
   *chareIdx = __idx;
   *startEpIdx = epIdxCtor;
@@ -184,7 +191,7 @@ int CkRegisterMainChare(int chareIdx, int entryIdx)
 
 #if CMK_CHARM4PY
 
-void CkRegisterMainChareExt(const char *s, int numEntryMethods, int *chareIdx, int *startEpIdx) {
+void CkRegisterMainChareExt(const char *s, const char **emNames, int emNamesStart, int numEntryMethods, int *chareIdx, int *startEpIdx) {
   int __idx = CkRegisterChare(s, sizeof(MainchareExt), TypeMainChare);
   CkRegisterBase(__idx, CkIndex_Chare::__idx);
 
@@ -192,8 +199,10 @@ void CkRegisterMainChareExt(const char *s, int numEntryMethods, int *chareIdx, i
   CkRegisterMessagePupFn(epIdxCtor, (CkMessagePupFn)CkArgMsg::ckDebugPup);
   CkRegisterMainChare(__idx, epIdxCtor);
 
-  for (int i=1; i < numEntryMethods; i++)
-    CkRegisterEp(s, MainchareExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+  for (int i=emNamesStart+1; i < emNamesStart+numEntryMethods; i++)
+    {
+      CkRegisterEp(emNames[i], MainchareExt::__entryMethod, CkMarshallMsg::__idx, __idx, 0+CK_EP_NOKEEP);
+    }
 
   *chareIdx = __idx;
   *startEpIdx = epIdxCtor;
diff --git a/src/ck-perf/trace-common.C b/src/ck-perf/trace-common.C
index a4a310db36..541a14e58a 100644
--- a/src/ck-perf/trace-common.C
+++ b/src/ck-perf/trace-common.C
@@ -296,8 +296,8 @@ void traceWriteSTS(FILE *stsfp,int nUserEvents) {
   for(i=0;i<_chareTable.size();i++)
     fprintf(stsfp, "CHARE %d \"%s\" %d\n", (int)i, _chareTable[i]->name, _chareTable[i]->ndims);
   for(i=0;i<_entryTable.size();i++)
-    fprintf(stsfp, "ENTRY CHARE %d \"%s\" %d %d\n", (int)i, _entryTable[i]->name,
-                 (int)_entryTable[i]->chareIdx, (int)_entryTable[i]->msgIdx);
+      fprintf(stsfp, "ENTRY CHARE %d \"%s\" %d %d\n", (int)i, _entryTable[i]->name,
+              (int)_entryTable[i]->chareIdx, (int)_entryTable[i]->msgIdx);
   for(i=0;i<_msgTable.size();i++)
     fprintf(stsfp, "MESSAGE %d %u\n", (int)i, (int)_msgTable[i]->size);
 }
@@ -487,10 +487,8 @@ static inline void _traceInit(char **argv)
   // check if trace is turned on/off for this pe
   CkpvAccess(traceOnPe) = checkTraceOnPe(argv);
 
-#if !CMK_CHARM4PY
   // defined in moduleInit.C
   _createTraces(argv);
-#endif
 
   // Now setup the control point tracing module if desired. It is always compiled/linked in, but is not always enabled
   // FIXME: make sure it is safe to use argv in SMP version