diff --git a/src/apex/address_resolution.cpp b/src/apex/address_resolution.cpp
index 87780c0e..795bb82b 100644
--- a/src/apex/address_resolution.cpp
+++ b/src/apex/address_resolution.cpp
@@ -118,8 +118,16 @@ namespace apex {
         if (rc == 0) {
         } else {
             node->info.probeAddr = ip;
-            node->info.filename = strdup(info.dli_fname);
-            node->info.funcname = strdup(info.dli_sname);
+            if (info.dli_fname == nullptr) {
+                node->info.filename = strdup("unknown");
+            } else {
+                node->info.filename = strdup(info.dli_fname);
+            }
+            if (info.dli_sname == nullptr) {
+                node->info.funcname = strdup("unknown");
+            } else {
+                node->info.funcname = strdup(info.dli_sname);
+            }
         }
 #endif // no APEX_HAVE_BFD
 #endif // no __APPLE__
diff --git a/src/apex/apex.cpp b/src/apex/apex.cpp
index f3066179..100e775f 100644
--- a/src/apex/apex.cpp
+++ b/src/apex/apex.cpp
@@ -114,7 +114,7 @@ using namespace std;
 namespace apex
 {
 
-bool& apex::get_program_over() {
+bool& get_program_over() {
     static bool _program_over{false};
     return _program_over;
 }
@@ -671,7 +671,7 @@ inline std::shared_ptr<task_wrapper> _new_task(
 }
 
 void debug_print(const char * event, std::shared_ptr<task_wrapper> tt_ptr) {
-    if (apex::get_program_over()) return;
+    if (get_program_over()) return;
     static std::mutex this_mutex;
     std::unique_lock<std::mutex> l(this_mutex);
     std::stringstream ss;
@@ -877,6 +877,7 @@ void start(std::shared_ptr<task_wrapper> tt_ptr) {
                 return;
             }
         }
+        tt_ptr->prof->thread_id = thread_instance::instance().get_id();
         // If we are allowing untied timers, clear the timer stack on this thread
         if (apex_options::untied_timers() == true) {
             thread_instance::instance().clear_current_profiler();
@@ -1833,7 +1834,7 @@ void finalize(void)
 void cleanup(void) {
     in_apex prevent_deadlocks;
     FUNCTION_ENTER
-    apex::get_program_over() = true;
+    get_program_over() = true;
 #ifdef APEX_HAVE_HPX
     // prevent crash at shutdown.
     return;
diff --git a/src/apex/apex_api.hpp b/src/apex/apex_api.hpp
index 992f5a6e..2783e95a 100644
--- a/src/apex/apex_api.hpp
+++ b/src/apex/apex_api.hpp
@@ -357,6 +357,11 @@ APEX_EXPORT std::shared_ptr<task_wrapper> new_task(
     const uint64_t task_id = UINTMAX_MAX,
     const std::shared_ptr<apex::task_wrapper> parent_task = null_task_wrapper);
 
+APEX_EXPORT std::shared_ptr<task_wrapper> new_task(
+    const std::string &name,
+    const uint64_t task_id = UINTMAX_MAX,
+    const std::vector<std::shared_ptr<apex::task_wrapper>> parent_tasks);
+
 /**
  \brief Create a new task (dependency).
 
@@ -856,6 +861,8 @@ APEX_EXPORT void send (uint64_t tag, uint64_t size, uint64_t target);
 APEX_EXPORT void recv (uint64_t tag, uint64_t size,
     uint64_t source_rank, uint64_t source_thread);
 
+APEX_EXPORT bool& get_program_over(void);
+
 /**
  \brief A convenience class for using APEX in C++ applications.
 
diff --git a/src/apex/profiler.hpp b/src/apex/profiler.hpp
index 4f35ae9b..21eb163e 100644
--- a/src/apex/profiler.hpp
+++ b/src/apex/profiler.hpp
@@ -86,7 +86,8 @@ class profiler {
         guid(task->guid),
         is_counter(false),
         is_resume(resume),
-        is_reset(reset), stopped(false) {
+        is_reset(reset), stopped(false),
+        thread_id(task->thread_id) {
             task->prof = this;
             task->start_ns = start_ns;
         }
diff --git a/src/apex/taskstubs_implementation.cpp b/src/apex/taskstubs_implementation.cpp
index 65512af6..8658f9b0 100644
--- a/src/apex/taskstubs_implementation.cpp
+++ b/src/apex/taskstubs_implementation.cpp
@@ -74,6 +74,8 @@ extern "C" {
         const tasktimer_guid_t timer_guid,
         const tasktimer_guid_t* parent_guids,
         const uint64_t parent_count) {
+        static bool& over = apex::get_program_over();
+        if (over) return nullptr;
         // TODO: need to handle multiple parents!
         // need to look up the parent shared pointers?
         std::vector<std::shared_ptr<apex::task_wrapper>> parent_tasks;
@@ -110,6 +112,8 @@ extern "C" {
         tasktimer_timer_t timer,
         tasktimer_argument_value_p arguments,
         uint64_t argument_count) {
+        static bool& over = apex::get_program_over();
+        if (over) return;
         // TODO: handle the schedule event, somehow
         APEX_UNUSED(timer);
         APEX_UNUSED(arguments);
@@ -117,6 +121,8 @@ extern "C" {
     }
 
 #define MAP_TASK(_timer, _apex_timer) \
+    static bool& over_{apex::get_program_over()}; \
+    if (over_) return; \
     uint64_t _tmp = (uint64_t)(_timer); \
     auto _apex_timer = safeLookup(_tmp);
 
@@ -129,6 +135,8 @@ extern "C" {
     }
     void tasktimer_yield_impl(
         tasktimer_timer_t timer) {
+        static bool& over = apex::get_program_over();
+        if (over) return;
         MAP_TASK(timer, apex_timer);
         apex::yield(apex_timer);
     }
@@ -170,5 +178,46 @@ extern "C" {
         APEX_UNUSED(children);
         APEX_UNUSED(child_count);
     }
+
+    void timerStack(
+        std::shared_ptr<apex::task_wrapper> apex_timer,
+        bool start) {
+        static thread_local std::stack<std::shared_ptr<apex::task_wrapper>> theStack;
+        if (start) {
+            apex::start(apex_timer);
+            theStack.push(apex_timer);
+        } else {
+            auto timer = theStack.top();
+            apex::stop(timer);
+            theStack.pop();
+        }
+    }
+
+    void tasktimer_data_transfer_start_impl(
+        tasktimer_guid_t guid,
+        tasktimer_execution_space_p source_type,
+        const char* source_name,
+        const void* source_ptr,
+        tasktimer_execution_space_p dest_type,
+        const char* dest_name,
+        const void* dest_ptr) {
+        std::shared_ptr<apex::task_wrapper> parent = safeLookup(guid);
+        auto task = apex::new_task("data xfer", 0, parent);
+        timerStack(task, true);
+    }
+
+    void tasktimer_data_transfer_stop_impl(tasktimer_guid_t guid) {
+        timerStack(nullptr, false);
+    }
+
+    void tasktimer_command_start_impl(const char* type_name) {
+        auto task = apex::new_task(type_name);
+        timerStack(task, true);
+    }
+
+    void tasktimer_command_stop_impl(void) {
+        timerStack(nullptr, false);
+    }
+
 }
 
diff --git a/src/apex/thread_instance.cpp b/src/apex/thread_instance.cpp
index 25e8bee1..035923c3 100644
--- a/src/apex/thread_instance.cpp
+++ b/src/apex/thread_instance.cpp
@@ -88,7 +88,7 @@ thread_instance::~thread_instance(void) {
         }
         std::cout << "thread " << _id << " exiting... " << __APEX_FUNCTION__ << std::endl;
     }
-    if (apex::get_program_over()) { return; }
+    if (get_program_over()) { return; }
     if (_top_level_timer != nullptr) {
         stop(_top_level_timer);
         _top_level_timer = nullptr;
diff --git a/src/apex/trace_event_listener.cpp b/src/apex/trace_event_listener.cpp
index 5506e2cf..214e46ff 100644
--- a/src/apex/trace_event_listener.cpp
+++ b/src/apex/trace_event_listener.cpp
@@ -198,11 +198,13 @@ void write_flow_event(std::stringstream& ss, double ts, char ph,
 
 inline void trace_event_listener::_common_stop(std::shared_ptr<profiler> &p) {
     static APEX_NATIVE_TLS long unsigned int tid = get_thread_id_metadata();
+    static auto main_wrapper = task_wrapper::get_apex_main_wrapper();
     // With HPX, the APEX MAIN timer will be stopped on a different thread
     // than the one that started it. So... make sure we get the right TID
     // But don't worry, the thread metadata will have been written at the
     // event start.
-    long unsigned int _tid = (p->tt_ptr->explicit_trace_start ? p->thread_id : tid);
+    //long unsigned int _tid = (p->tt_ptr->explicit_trace_start ? p->thread_id : tid);
+    long unsigned int _tid = p->thread_id;
     if (!_terminate) {
         std::stringstream ss;
         ss.precision(3);
@@ -212,7 +214,8 @@ inline void trace_event_listener::_common_stop(std::shared_ptr<profiler> &p) {
             pguid = p->tt_ptr->parent->guid;
         }
         // if the parent tid is not the same, create a flow event BEFORE the single event
-        if (p->tt_ptr->parent != nullptr
+        if (p->tt_ptr->parent != nullptr &&
+            p->tt_ptr->parent != main_wrapper
 #ifndef APEX_HAVE_HPX // ...except for HPX - make the flow event regardless
             && p->tt_ptr->parent->thread_id != _tid
 #endif
diff --git a/src/scripts/apex_exec b/src/scripts/apex_exec
index 4341c07d..e8b49781 100755
--- a/src/scripts/apex_exec
+++ b/src/scripts/apex_exec
@@ -29,6 +29,7 @@ $(basename "$0") <APEX options> executable <executable options>
 where APEX options are zero or more of:
     --apex:help                   show this usage message
     --apex:debug                  run with APEX in debugger
+    --apex:debugger               debugger executable name (default: gdb)
     --apex:verbose                enable verbose list of APEX environment variables
     --apex:screen                 enable screen text output (on by default)
     --apex:screen-detail          enable detailed text output (off by default)
@@ -185,6 +186,9 @@ while (( "$#" )); do
       ;;
     --apex:debug)
       debug=yes
+      if [ "$debugger" == "" ] ; then
+        debugger=gdb
+      fi
       shift
       ;;
     --apex:pthread)
@@ -422,6 +426,16 @@ while (( "$#" )); do
       postprocess=yes
       shift
       ;;
+    --apex:debugger)
+      debug=yes
+      if [ -n "$2" ] && [ ${2:0:1} != "-" ]; then
+        debugger=$2
+        shift 2
+      else
+        echo "Error: Argument for $1 is missing" >&2
+        usage
+      fi
+      ;;
     --apex:concur_max|--apex:concur-max)
       if [ -n "$2" ] && [ ${2:0:1} != "-" ]; then
         export APEX_MEASURE_CONCURRENCY_MAX_TIMERS=$2
@@ -710,11 +724,7 @@ else
             gdbargs="-batch -q"
         fi
         #echo "set env LD_AUDIT=${APEX_LD_AUDIT}" >> ./.gdbcmds
-        if [ $hip = yes ] ; then
-            debugger="rocgdb -x ${gdbcmds} ${gdbargs} --args"
-        else
-            debugger="gdb -x ${gdbcmds} ${gdbargs} --args"
-        fi
+        debugger="${debugger} -x ${gdbcmds} ${gdbargs} --args"
     else
         export LD_LIBRARY_PATH=${APEX_LD_LIBRARY_PATH}
         export LD_PRELOAD=${APEX_LD_PRELOAD}