diff --git a/.gitmodules b/.gitmodules
index 1ab918920..f46086ada 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -7,3 +7,6 @@
 [submodule "models/ROSS-Models"]
 	path = models/ROSS-Models
 	url = https://github.com/carothersc/ROSS-Models
+[submodule "damaris"]
+	path = core/damaris
+	url = https://github.com/caitlinross/ROSS-damaris
diff --git a/CMakeLists.txt b/CMakeLists.txt
index be9d8999e..f257e1318 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -128,3 +128,4 @@ IF(DOXYGEN_FOUND)
     ADD_SUBDIRECTORY(docs)
   ENDIF(ROSS_BUILD_DOXYGEN)
 ENDIF(DOXYGEN_FOUND)
+
diff --git a/core/CMakeLists.txt b/core/CMakeLists.txt
index 8f3cdc356..4316c269c 100644
--- a/core/CMakeLists.txt
+++ b/core/CMakeLists.txt
@@ -100,6 +100,14 @@ IF(USE_RIO)
 	SET(ross_srcs ${ross_srcs} ${RIO_SOURCE_DIR}/io.h)
 ENDIF(USE_RIO)
 
+# Damaris I/O and data management
+OPTION(USE_DAMARIS "Build with Damaris library (for in situ vis/analysis)?" OFF)
+if (USE_DAMARIS)
+    ADD_SUBDIRECTORY(damaris)
+    INCLUDE_DIRECTORIES(${DAMARIS_INCLUDE})
+    SET(ross_srcs ${ross_srcs} ${ROSS_Damaris_SOURCE_DIR}/core/damaris.h)
+ENDIF(USE_DAMARIS)
+
 # Use debugging-friendly memory allocation
 OPTION(ROSS_ALLOC_DEBUG "Use naive allocator to be more friendly to memory debugging tools" OFF)
 
diff --git a/core/config.h.in b/core/config.h.in
index aa7f8e65f..a397082cc 100644
--- a/core/config.h.in
+++ b/core/config.h.in
@@ -17,3 +17,4 @@
 #cmakedefine ROSS_runtime_checks
 #cmakedefine ROSS_ALLOC_DEBUG
 #cmakedefine USE_RIO
+#cmakedefine USE_DAMARIS
diff --git a/core/damaris b/core/damaris
new file mode 160000
index 000000000..f4a1739c7
--- /dev/null
+++ b/core/damaris
@@ -0,0 +1 @@
+Subproject commit f4a1739c74233be0926841107913d891a3208b4a
diff --git a/core/gvt/mpi_allreduce.c b/core/gvt/mpi_allreduce.c
index 56ce9c0c3..b0b9a0854 100644
--- a/core/gvt/mpi_allreduce.c
+++ b/core/gvt/mpi_allreduce.c
@@ -211,8 +211,26 @@ tw_gvt_step2(tw_pe *me)
     if ((g_st_engine_stats == GVT_STATS || g_st_engine_stats == ALL_STATS) && 
             g_tw_gvt_done % g_st_num_gvt == 0 && gvt <= g_tw_ts_end)
     {
+#ifdef USE_DAMARIS
+        if (g_st_damaris_enabled)
+        {
+            st_damaris_expose_data(me, gvt, GVT_COL);
+            st_damaris_end_iteration();
+        }
+        else
+            st_collect_engine_data(me, GVT_COL);
+#else
 		st_collect_engine_data(me, GVT_COL);
+#endif
     }
+#ifdef USE_DAMARIS
+    // need to make sure damaris_end_iteration is called if GVT instrumentation not turned on
+    //if (!g_st_stats_enabled && g_st_real_time_samp) //need to make sure if one PE enters this, all do; otherwise deadlock
+    if (g_st_damaris_enabled && (g_st_engine_stats == RT_STATS || g_st_engine_stats == VT_STATS))
+    {
+        st_damaris_end_iteration();
+    }
+#endif
 
     if ((g_st_model_stats == GVT_STATS || g_st_model_stats == ALL_STATS) && g_tw_gvt_done % g_st_num_gvt == 0)
         st_collect_model_data(me, (tw_stime)tw_clock_read() / g_tw_clock_rate, GVT_STATS);
diff --git a/core/instrumentation/ross-lps/analysis-lp.c b/core/instrumentation/ross-lps/analysis-lp.c
index 34e558870..1eb1fe7fe 100644
--- a/core/instrumentation/ross-lps/analysis-lp.c
+++ b/core/instrumentation/ross-lps/analysis-lp.c
@@ -125,8 +125,18 @@ void analysis_event(analysis_state *s, tw_bf *bf, analysis_msg *m, tw_lp *lp)
     }
 
     // sim engine sampling
-    if (g_tw_synchronization_protocol != SEQUENTIAL && (g_st_engine_stats == VT_STATS || g_st_engine_stats == ALL_STATS))
+    if (g_tw_synchronization_protocol != SEQUENTIAL && 
+            (g_st_engine_stats == VT_STATS || g_st_engine_stats == ALL_STATS))
+    {
+#ifdef USE_DAMARIS
+        if (g_st_damaris_enabled)
+            st_damaris_expose_data(lp->pe, tw_now(lp), ANALYSIS_LP);
+        else
+            st_collect_engine_data(lp->pe, ANALYSIS_LP);
+#else
         st_collect_engine_data(lp->pe, ANALYSIS_LP);
+#endif
+    }
         //collect_sim_engine_data(lp->pe, lp, s, (tw_stime) tw_clock_read() / g_tw_clock_rate);
     
     // create next sampling event
diff --git a/core/instrumentation/st-instrumentation.c b/core/instrumentation/st-instrumentation.c
index 914bef937..75aad87b1 100644
--- a/core/instrumentation/st-instrumentation.c
+++ b/core/instrumentation/st-instrumentation.c
@@ -3,7 +3,9 @@
 
 char g_st_stats_out[INST_MAX_LENGTH] = {0};
 char g_st_stats_path[4096] = {0};
-int g_st_granularity = 0;
+int g_st_pe_data = 1;
+int g_st_kp_data = 0;
+int g_st_lp_data = 0;
 int g_st_disable_out = 0;
 
 int g_st_model_stats = 0;
@@ -19,6 +21,8 @@ tw_clock g_st_rt_samp_start_cycles = 0;
 tw_stime g_st_vt_interval = 1000000;
 tw_stime g_st_sampling_end = 0;
 
+
+
 static const tw_optdef inst_options[] = {
     TWOPT_GROUP("ROSS Instrumentation"),
     TWOPT_UINT("engine-stats", g_st_engine_stats, "Collect sim engine level stats; 0 don't collect, 1 GVT-sampling, 2 RT sampling, 3 VT sampling, 4 All sampling modes"),
@@ -27,7 +31,9 @@ static const tw_optdef inst_options[] = {
     TWOPT_ULONGLONG("rt-interval", g_st_rt_interval, "real time sampling interval in ms"),
     TWOPT_STIME("vt-interval", g_st_vt_interval, "Virtual time sampling interval"),
     TWOPT_STIME("vt-samp-end", g_st_sampling_end, "End time for virtual time sampling (if different from g_tw_ts_end)"),
-    TWOPT_UINT("granularity", g_st_granularity, "for sim engine instrumentation; 0 = PE only, 1 = KP only, 2 = LP only, 3 = All levels"),
+    TWOPT_UINT("pe-data", g_st_pe_data, "Turn on/off collection of sim engine data at PE level"),
+    TWOPT_UINT("kp-data", g_st_kp_data, "Turn on/off collection of sim engine data at KP level"),
+    TWOPT_UINT("lp-data", g_st_lp_data, "Turn on/off collection of sim engine data at LP level"),
     TWOPT_UINT("event-trace", g_st_ev_trace, "collect detailed data on all events for specified LPs; 0, no trace, 1 full trace, 2 only events causing rollbacks, 3 only committed events"),
     TWOPT_CHAR("stats-prefix", g_st_stats_out, "prefix for filename(s) for stats output"),
     TWOPT_CHAR("stats-path", g_st_stats_path, "path to directory to save instrumentation output"),
diff --git a/core/instrumentation/st-instrumentation.h b/core/instrumentation/st-instrumentation.h
index ce9366a83..91ac3514c 100644
--- a/core/instrumentation/st-instrumentation.h
+++ b/core/instrumentation/st-instrumentation.h
@@ -70,7 +70,9 @@ struct sample_metadata
 
 extern char g_st_stats_out[INST_MAX_LENGTH];
 extern char g_st_stats_path[INST_MAX_LENGTH];
-extern int g_st_granularity;
+extern int g_st_pe_data;
+extern int g_st_kp_data;
+extern int g_st_lp_data;
 extern int g_st_disable_out;
 
 extern int g_st_model_stats;
diff --git a/core/instrumentation/st-sim-engine.c b/core/instrumentation/st-sim-engine.c
index 44ac10a62..1d625cb19 100644
--- a/core/instrumentation/st-sim-engine.c
+++ b/core/instrumentation/st-sim-engine.c
@@ -21,9 +21,9 @@ void st_collect_engine_data(tw_pe *pe, int col_type)
     sample_md.ts = pe->GVT;
     sample_md.real_time = (double)tw_clock_read() / g_tw_clock_rate;
 
-    if (g_st_granularity == GRAN_PE || g_st_granularity == GRAN_ALL)
+    if (g_st_pe_data)
         st_collect_engine_data_pes(pe, &sample_md, &s, col_type);
-    if (g_st_granularity == GRAN_KP || g_st_granularity == GRAN_ALL)
+    if (g_st_kp_data)
     {
         for (i = 0; i < g_tw_nkp; i++)
         {
@@ -31,7 +31,7 @@ void st_collect_engine_data(tw_pe *pe, int col_type)
             st_collect_engine_data_kps(pe, kp, &sample_md, &s, col_type);
         }
     }
-    if (g_st_granularity == GRAN_LP || g_st_granularity == GRAN_ALL)
+    if (g_st_lp_data)
     {
         for (i = 0; i < g_tw_nlp; i++)
         {
@@ -144,7 +144,6 @@ void st_collect_engine_data_kps(tw_pe *pe, tw_kp *kp, sample_metadata *sample_md
 
 void st_collect_engine_data_lps(tw_pe *pe, tw_lp *lp, sample_metadata *sample_md, tw_statistics *s, int col_type)
 {
-    tw_kp *kp;
     st_lp_stats lp_stats;
     int buf_size = sizeof(*sample_md) + sizeof(lp_stats);
     char buffer[buf_size];
diff --git a/core/network-mpi.c b/core/network-mpi.c
index 5ccf0adee..f98b6b9c8 100644
--- a/core/network-mpi.c
+++ b/core/network-mpi.c
@@ -67,7 +67,6 @@ const tw_optdef *
 tw_net_init(int *argc, char ***argv)
 {
   int my_rank;
-
   int initialized;
   MPI_Initialized(&initialized);
 
@@ -75,12 +74,13 @@ tw_net_init(int *argc, char ***argv)
     if (MPI_Init(argc, argv) != MPI_SUCCESS)
       tw_error(TW_LOC, "MPI_Init failed.");
   }
+
   if (MPI_Comm_rank(MPI_COMM_ROSS, &my_rank) != MPI_SUCCESS)
     tw_error(TW_LOC, "Cannot get MPI_Comm_rank(MPI_COMM_ROSS)");
 
   g_tw_masternode = 0;
   g_tw_mynode = my_rank;
-
+  
   return mpi_opts;
 }
 
@@ -180,10 +180,22 @@ tw_net_abort(void)
 void
 tw_net_stop(void)
 {
+#ifdef USE_DAMARIS
+    if (g_st_damaris_enabled)
+        st_damaris_ross_finalize();
+    else
+    {
+        if (!custom_communicator) {
+            if (MPI_Finalize() != MPI_SUCCESS)
+                tw_error(TW_LOC, "Failed to finalize MPI");
+        }
+    }
+#else
   if (!custom_communicator) {
     if (MPI_Finalize() != MPI_SUCCESS)
       tw_error(TW_LOC, "Failed to finalize MPI");
   }
+#endif
 }
 
 void
diff --git a/core/ross-extern.h b/core/ross-extern.h
index e0226f25c..9d7c491da 100644
--- a/core/ross-extern.h
+++ b/core/ross-extern.h
@@ -188,6 +188,7 @@ extern size_t	tw_memory_allocate(tw_memoryq *);
 #define	TW_LOC	__FILE__,__LINE__
 extern int tw_output(tw_lp *lp, const char *fmt, ...);
 extern void tw_error(const char *file, int line, const char *fmt, ...) NORETURN;
+extern void tw_warning(const char *file, int line, const char *fmt, ...);
 extern void tw_printf(const char *file, int line, const char *fmt, ...);
 extern void tw_calloc_stats(size_t *alloc, size_t *waste);
 extern void* tw_calloc(const char *file, int line, const char *for_who, size_t e_sz, size_t n);
diff --git a/core/ross.h b/core/ross.h
index 21beccb4b..07c956c98 100644
--- a/core/ross.h
+++ b/core/ross.h
@@ -213,6 +213,10 @@ typedef uint64_t tw_lpid;
 
 #include "instrumentation/st-instrumentation.h"
 
+#ifdef USE_DAMARIS
+#include "damaris/core/damaris.h"
+#endif
+
 #include "tw-eventq.h"
 
 #ifdef USE_RIO
diff --git a/core/tw-sched.c b/core/tw-sched.c
index d8a5f4d00..0ec9d726e 100644
--- a/core/tw-sched.c
+++ b/core/tw-sched.c
@@ -258,11 +258,20 @@ static void tw_sched_batch(tw_pe * me) {
                 tw_clock_read() - g_st_rt_samp_start_cycles > g_st_rt_interval)
         {
             tw_clock current_rt = tw_clock_read();
+#ifdef USE_DAMARIS
+            if (g_st_engine_stats == RT_STATS || g_st_engine_stats == ALL_STATS)
+            {
+                if (g_st_damaris_enabled)
+                    st_damaris_expose_data(me, me->GVT, RT_COL);
+                else
+                    st_collect_engine_data(me, RT_COL);
+            }
+#else
             if (g_st_engine_stats == RT_STATS || g_st_engine_stats == ALL_STATS)
                 st_collect_engine_data(me, RT_COL);
             if (g_st_model_stats == RT_STATS || g_st_model_stats == ALL_STATS)
                 st_collect_model_data(me, (tw_stime)current_rt / g_tw_clock_rate, RT_STATS);
-
+#endif
             g_st_rt_samp_start_cycles = tw_clock_read();
         }
 
diff --git a/core/tw-setup.c b/core/tw-setup.c
index cf490591c..10f8db559 100644
--- a/core/tw-setup.c
+++ b/core/tw-setup.c
@@ -38,7 +38,7 @@ void tw_init(int *argc, char ***argv) {
 #endif
 
     tw_opt_add(tw_net_init(argc, argv));
-
+    
     // Print out the command-line so we know what we passed in
     if (tw_ismaster()) {
         for (i = 0; i < *argc; i++) {
@@ -62,6 +62,9 @@ void tw_init(int *argc, char ***argv) {
     tw_opt_add(tw_gvt_setup());
     tw_opt_add(tw_clock_setup());
     tw_opt_add(st_inst_opts());
+#ifdef USE_DAMARIS
+    tw_opt_add(st_damaris_opts());
+#endif
     tw_opt_add(st_special_lp_opts());
 #ifdef USE_RIO
     tw_opt_add(io_opts);
@@ -70,6 +73,14 @@ void tw_init(int *argc, char ***argv) {
     // by now all options must be in
     tw_opt_parse(argc, argv);
 
+#ifdef USE_DAMARIS
+    st_damaris_ross_init();
+    if (!g_st_ross_rank) // Damaris ranks only
+        return; 
+    else 
+    {
+#endif
+
     if(tw_ismaster())
     {
         struct stat buffer;
@@ -103,7 +114,9 @@ void tw_init(int *argc, char ***argv) {
 
     tw_net_start();
     tw_gvt_start();
-
+#ifdef USE_DAMARIS
+    } // end of if(g_st_ross_rank)
+#endif
 }
 
 static void early_sanity_check(void) {
@@ -279,6 +292,7 @@ void tw_define_lps(tw_lpid nlp, size_t msg_sz) {
             tw_rand_init_streams(g_tw_lp[i], g_tw_nRNG_per_lp);
         }
     }
+
 }
 
 static void late_sanity_check(void) {
@@ -350,6 +364,10 @@ void tw_run(void) {
 
     // init instrumentation
     st_inst_init(); 
+#ifdef USE_DAMARIS
+    if (g_st_damaris_enabled)
+        st_damaris_inst_init();
+#endif
 
 #ifdef USE_BGPM
     Bgpm_Init(BGPM_MODE_SWDISTRIB);
@@ -419,10 +437,17 @@ void tw_run(void) {
 }
 
 void tw_end(void) {
+#ifdef USE_DAMARIS
+    if(g_st_ross_rank)
+    {
+#endif
     if(tw_ismaster()) {
         fprintf(g_tw_csv, "\n");
         fclose(g_tw_csv);
     }
+#ifdef USE_DAMARIS
+    } // end if(g_st_ross_rank)
+#endif
 
     tw_net_stop();
 }
@@ -518,6 +543,10 @@ static tw_pe * setup_pes(void) {
         }
         printf("\n");
 
+#ifdef USE_DAMARIS
+    st_damaris_init_print();
+#endif
+
         // moved these so ross.csv stays consistent
         fprintf(g_tw_csv, "%d,", num_events_per_pe);
         fprintf(g_tw_csv, "%d,", g_tw_gvt_threshold);
diff --git a/core/tw-util.c b/core/tw-util.c
index a5722713b..f8c3dcec9 100644
--- a/core/tw-util.c
+++ b/core/tw-util.c
@@ -86,6 +86,20 @@ tw_error(const char *file, int line, const char *fmt, ...)
 	tw_net_abort();
 }
 
+void
+tw_warning(const char *file, int line, const char *fmt, ...)
+{
+	va_list	ap;
+
+	va_start(ap, fmt);
+	fprintf(stdout, "node: %ld: warning: %s:%i: ", g_tw_mynode, file, line);
+	vfprintf(stdout, fmt, ap);
+	fprintf(stdout, "\n");
+	fflush(stdout);
+	fflush(stdout);
+	va_end(ap);
+}
+
 struct mem_pool
 {
 	struct mem_pool *next_pool;
diff --git a/models/CMakeLists.txt b/models/CMakeLists.txt
index fb23b215f..ca8011854 100644
--- a/models/CMakeLists.txt
+++ b/models/CMakeLists.txt
@@ -18,13 +18,13 @@ ENDFUNCTION(ROSS_TEST_SCHEDULERS)
 ## INSTRUMENTATION TESTS
 FUNCTION(ROSS_TEST_INSTRUMENTATION target_name)
 	ADD_TEST(${target_name}_INST_OptPE mpirun -np 2 ./${target_name} --synch=3 --engine-stats=4 --event-trace=2 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
-	ADD_TEST(${target_name}_INST_OptLP mpirun -np 2 ./${target_name} --synch=3 --engine-stats=4 --event-trace=2 --granularity=3 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
+	ADD_TEST(${target_name}_INST_OptLP mpirun -np 2 ./${target_name} --synch=3 --engine-stats=4 --event-trace=2 --kp-data=1 --lp-data=1 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
 
 	ADD_TEST(${target_name}_INST_RTOptPE mpirun -np 2 ./${target_name} --synch=5 --engine-stats=4 --event-trace=2 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
-	ADD_TEST(${target_name}_INST_RTOptLP mpirun -np 2 ./${target_name} --synch=5 --engine-stats=4 --event-trace=2 --granularity=3 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
+	ADD_TEST(${target_name}_INST_RTOptLP mpirun -np 2 ./${target_name} --synch=5 --engine-stats=4 --event-trace=2 --kp-data=1 --lp-data=1 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
 
 	ADD_TEST(${target_name}_INST_ConsPE mpirun -np 2 ./${target_name} --synch=2 --engine-stats=4 --event-trace=2 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
-	ADD_TEST(${target_name}_INST_ConsLP mpirun -np 2 ./${target_name} --synch=2 --engine-stats=4 --event-trace=2 --granularity=3 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
+	ADD_TEST(${target_name}_INST_ConsLP mpirun -np 2 ./${target_name} --synch=2 --engine-stats=4 --event-trace=2 --kp-data=1 --lp-data=1 --extramem=100000 --vt-interval=10000 --vt-samp-end=100000)
 
     ADD_TEST(${target_name}_INST_Seq ./${target_name} --synch=1 --event-trace=1 --extramem=100000)
 
diff --git a/models/phold/CMakeLists.txt b/models/phold/CMakeLists.txt
index b6c458c46..100cfb411 100644
--- a/models/phold/CMakeLists.txt
+++ b/models/phold/CMakeLists.txt
@@ -1,4 +1,7 @@
 INCLUDE_DIRECTORIES(${ROSS_BINARY_DIR})
+IF(USE_DAMARIS)
+    INCLUDE_DIRECTORIES(${DAMARIS_INCLUDE}) 
+ENDIF(USE_DAMARIS)
 IF(BGPM)
 	INCLUDE_DIRECTORIES(${ROSS_SOURCE_DIR} ${BGPM_LIB})
 ELSE(NOT(BGPM))
@@ -8,6 +11,7 @@ ENDIF(BGPM)
 SET(phold_srcs
 phold.c		phold.h)
 
+
 ADD_EXECUTABLE(phold ${phold_srcs})
 ADD_EXECUTABLE(phold_test ${phold_srcs})
 
@@ -15,8 +19,13 @@ IF(BGPM)
 	TARGET_LINK_LIBRARIES(phold ROSS imp_bgpm m)
 	TARGET_LINK_LIBRARIES(phold_test ROSS imp_bgpm m)
 ELSE(NOT(BGPM))
-	TARGET_LINK_LIBRARIES(phold ROSS m)
-	TARGET_LINK_LIBRARIES(phold_test ROSS m)
+    IF(USE_DAMARIS)
+        TARGET_LINK_LIBRARIES(phold ROSS ROSS_Damaris m)
+        TARGET_LINK_LIBRARIES(phold_test ROSS ROSS_Damaris m)
+    ELSE(NOT(USE_DAMARIS))
+        TARGET_LINK_LIBRARIES(phold ROSS m)
+        TARGET_LINK_LIBRARIES(phold_test ROSS m)
+    ENDIF(USE_DAMARIS)
 ENDIF(BGPM)
 
 ROSS_TEST_SCHEDULERS(phold)
diff --git a/models/phold/phold.c b/models/phold/phold.c
index f1e7e7df4..961d8a435 100644
--- a/models/phold/phold.c
+++ b/models/phold/phold.c
@@ -175,6 +175,10 @@ main(int argc, char **argv, char **env)
 	tw_opt_add(app_opt);
 	tw_init(&argc, &argv);
 
+#ifdef USE_DAMARIS
+    if(g_st_ross_rank)
+    { // only ross ranks should run code between here and tw_run()
+#endif
 	if( lookahead > 1.0 )
 	  tw_error(TW_LOC, "Lookahead > 1.0 .. needs to be less\n");
 
@@ -213,6 +217,9 @@ main(int argc, char **argv, char **env)
 	  }
 
 	tw_run();
+#ifdef USE_DAMARIS
+    } // end if(g_st_ross_rank)
+#endif
 	tw_end();
 
 	return 0;