Updating network workloads: adding dumpi reader

Makefile.am

@@ -47,3 +47,9 @@ if USE_RECORDER
 AM_CPPFLAGS += ${RECORDER_CPPFLAGS}
 src_libcodes_base_a_SOURCES += src/workload/codes-recorder-io-wrkld.c
 endif
+
+if USE_DUMPI
+AM_CPPFLAGS += ${DUMPI_CFLAGS} -DUSE_DUMPI=1
+src_libcodes_base_a_SOURCES += src/network-workload/codes-dumpi-trace-nw-wrkld.c 
+AM_LIBS += ${DUMPI_LIBS}
+endif

codes/codes-nw-workload.h

@@ -9,21 +9,35 @@
 
 #include "ross.h"
 
-#define MAX_NAME_LENGTH 256
+#define MAX_NAME_LENGTH 512
 
 /* struct to hold the actual data from a single MPI event*/
 typedef struct mpi_event_list mpi_event_list;
 typedef struct scala_trace_params scala_trace_params;
 
+#ifdef USE_DUMPI
+typedef struct dumpi_trace_params dumpi_trace_params;
+#endif
+
 struct scala_trace_params
 {
    char offset_file_name[MAX_NAME_LENGTH];
    char nw_wrkld_file_name[MAX_NAME_LENGTH];
 };
 
+#ifdef USE_DUMPI
+struct dumpi_trace_params
+{
+   char file_name[MAX_NAME_LENGTH];
+};
+#endif
+
 enum NW_WORKLOADS
 {
    SCALA_TRACE = 1,
+#ifdef USE_DUMPI
+   DUMPI,
+#endif
    OTHERS, /* add the names of other workload generators here */
 };
 enum mpi_workload_type
@@ -32,10 +46,30 @@ enum mpi_workload_type
      CODES_NW_END = 1,
     /* sleep/delay to simulate computation or other activity */
      CODES_NW_DELAY,
-    /* MPI send operation */
+    /* MPI blocking send operation */
      CODES_NW_SEND,
-    /* MPI recv operation */
-     CODES_NW_RECV
+    /* MPI blocking recv operation */
+     CODES_NW_RECV,
+    /* MPI non-blocking send operation */
+     CODES_NW_ISEND,
+    /* MPI non-blocking receive operation */
+     CODES_NW_IRECV,
+    /* MPI broadcast operation */
+     CODES_NW_BCAST,
+    /* MPI Allgather operation */
+     CODES_NW_ALLGATHER,
+     /* MPI Allgatherv operation */
+     CODES_NW_ALLGATHERV,
+    /* MPI Alltoall operation */
+     CODES_NW_ALLTOALL,
+    /* MPI Alltoallv operation */
+     CODES_NW_ALLTOALLV,
+    /* MPI Reduce operation */
+     CODES_NW_REDUCE,
+    /* MPI Allreduce operation */
+     CODES_NW_ALLREDUCE,
+    /* Generic collective operation */
+    CODES_NW_COL
 };
 
 /* data structure for holding data from a MPI event (coming through scala-trace) 
@@ -44,26 +78,35 @@ struct mpi_event_list
 {
     /* what type of operation this is */
     enum mpi_workload_type op_type;
+    double start_time;
+    double end_time;
 
    /* parameters for each operation type */
     union
     {
   	struct
   	{
-      	   long seconds;
+	   double nsecs;
+	   double seconds;
   	} delay;
         struct
   	{
       	    int source_rank;/* source rank of MPI send message */
             int dest_rank; /* dest rank of MPI send message */
+	    int num_bytes;
             int blocking; /* boolean value to indicate if message is blocking or non-blocking*/
-        } send;
+	} send;
        struct
        {
      	    int source_rank;/* source rank of MPI recv message */
      	    int dest_rank;/* dest rank of MPI recv message */
+	    int num_bytes;
      	    int blocking;/* boolean value to indicate if message is blocking or non-blocking*/
-       } recv;  
+       } recv; 
+      struct
+      {
+	  int num_bytes;
+      } collective;
     }u;
 };
 

configure.ac

@@ -112,6 +112,24 @@ AM_CONDITIONAL(USE_RECORDER, true)
 RECORDER_CPPFLAGS="-DUSE_RECORDER=1"
 AC_SUBST(RECORDER_CPPFLAGS)
 
+#check for Dumpi
+AC_ARG_WITH([dumpi],[AS_HELP_STRING([--with-dumpi@<:@=DIR@:>@],
+                        [location of Dumpi installation])])
+if test "x${with_dumpi}" != "x" ; then
+	AC_CHECK_FILE([${with_dumpi}/lib/libundumpi.la],
+                  AM_CONDITIONAL(USE_DUMPI, true),
+                  AC_MSG_ERROR(Could not find libundumpi.la))
+	DUMPI_CFLAGS="-I${with_dumpi}/include"
+#	DUMPI_CFLAGS+=" -I${with_dumpi}/include/dumpi/common" 
+#	DUMPI_CFLAGS+=" -I${with_dumpi}/include/dumpi/libdumpi"
+#	DUMPI_CFLAGS+=" -I${with_dumpi}/include/dumpi/libundumpi"
+	DUMPI_LIBS="-L${with_dumpi}/lib/ -ldumpi -lundumpi"
+	AC_SUBST(DUMPI_LIBS)
+	AC_SUBST(DUMPI_CFLAGS)
+else
+	AM_CONDITIONAL(USE_DUMPI, false)
+fi
+
 dnl ======================================================================
 dnl  Try harder to be valgrind safe
 dnl ======================================================================

src/Makefile.subdir

@@ -108,7 +108,8 @@ src_libcodes_base_a_SOURCES =  \
     codes/codes-nw-workload.h \
     src/network-workload/codes-nw-workload.c \
     src/network-workload/codes-nw-workload-method.h \
-    src/network-workload/codes-scala-trace-nw-wrkld.c
+    src/network-workload/codes-scala-trace-nw-wrkld.c \
+    src/network-workload/codes-dumpi-trace-nw-wrkld.c 
 
 # stealth testing of the template code (actual test is not run, just compiled as
 # a program - Make error signifies test failure)
@@ -133,6 +134,13 @@ bin_PROGRAMS += src/network-workload/codes-nw-test
 
 src_network_workload_codes_nw_test_SOURCES = \
  src/network-workload/codes-nw-test.c
-src_network_workload_codes_nw_test_LDADD = $(testlib) ${ROSS_LIBS}
+src_network_workload_codes_nw_test_LDADD = $(testlib) ${ROSS_LIBS} ${DUMPI_LIBS}
 src_network_workload_codes_nw_test_LDFLAGS =  ${ROSS_LDFLAGS}
 
+#bin_PROGRAMS += src/network-workload/codes-dumpi-wrkld
+
+#src_network_workload_codes_dumpi_wrkld_SOURCES = \
+ src/network-workload/codes-dumpi-wrkld.c
+#src_network_workload_codes_dumpi_wrkld_LDADD = $(testlib) ${ROSS_LIBS} ${DUMPI_LIBS}
+#src_network_workload_codes_dumpi_wrkld_LDFLAGS =  ${ROSS_LDFLAGS}
+

src/network-workload/codes-nw-test.c

@@ -8,7 +8,7 @@
 
 char workload_file[8192];
 char offset_file[8192];
-static int total_nw_lps = 16;
+static int total_nw_lps = 8;
 static int nlp_per_pe;
 static int wrkld_id;
 
@@ -36,10 +36,12 @@ tw_peid nw_test_map(tw_lpid gid)
 void nw_test_init(nw_state* s, tw_lp* lp)
 {
    /* initialize the LP's and load the data */
-   scala_trace_params params;
-   strcpy(params.offset_file_name, offset_file);
-   strcpy(params.nw_wrkld_file_name, workload_file);
-   wrkld_id = codes_nw_workload_load("scala-trace-workload", (char*)&params, (int)lp->gid);
+   //scala_trace_params params;
+   //strcpy(params.offset_file_name, offset_file);
+   //strcpy(params.nw_wrkld_file_name, workload_file);
+   dumpi_trace_params params;
+   strcpy(params.file_name, workload_file);
+   wrkld_id = codes_nw_workload_load("dumpi-trace-workload", (char*)&params, (int)lp->gid);
    
    tw_event *e;
    tw_stime kickoff_time;
@@ -108,10 +110,10 @@ int main( int argc, char** argv )
   tw_opt_add(app_opt);
   tw_init(&argc, &argv);
 
-  if(strlen(offset_file) == 0 || strlen(workload_file) == 0 || total_nw_lps == 0)
+  if(strlen(workload_file) == 0 || total_nw_lps == 0)
     {
 	if(tw_ismaster())
-		printf("\n Usage: mpirun -np n ./codes-nw-test --sync=1/2/3 --total_nw_lps=n --workload_file=workload-file-name --offset_file=offset-file-name ");
+		printf("\n Usage: mpirun -np n ./codes-nw-test --sync=1/2/3 --total_nw_lps=n --workload_file=workload-file-name");
 	tw_end();
 	return -1;
     }

src/network-workload/codes-nw-workload.c

@@ -14,10 +14,14 @@
  * could make generators optional via autoconf tests etc. if needed
  */
 extern struct codes_nw_workload_method scala_trace_workload_method;
+extern struct codes_nw_workload_method dumpi_trace_workload_method;
 
 static struct codes_nw_workload_method *method_array[] =
 {
     &scala_trace_workload_method,
+#ifdef USE_DUMPI
+    &dumpi_trace_workload_method,
+#endif
     NULL};
 
 /* This shim layer is responsible for queueing up reversed operations and
@@ -153,21 +157,39 @@ void codes_nw_workload_print_op(FILE *f, struct mpi_event_list *op, int rank){
             fprintf(f, "op: rank:%d type:end\n", rank);
             break;
         case CODES_NW_DELAY:
-            fprintf(f, "op: rank:%d type:delay seconds:%ld\n",
-                    rank, op->u.delay.seconds);
+            fprintf(f, "op: rank:%d type:delay seconds:%f \n",
+                    rank, op->u.delay.nsecs);
             break;
         case CODES_NW_SEND:
+	case CODES_NW_ISEND:
             fprintf(f, "op: rank:%d type:send "
-                       "sender: %d receiver: %d blocking: %d \n",
+                       "sender: %d receiver: %d blocking: %d number of bytes: %d "
+			"start time: %f end time: %f \n",
                     rank, op->u.send.source_rank, op->u.send.dest_rank,
-                    op->u.send.blocking);
+                    op->u.send.blocking, op->u.send.num_bytes,
+		    op->start_time, op->end_time);
             break;
         case CODES_NW_RECV:
+	case CODES_NW_IRECV:
             fprintf(f, "op: rank:%d type:recv "
-                       "sender: %d receiver: %d blocking: %d \n",
+                       "sender: %d receiver: %d blocking: %d number of bytes: %d "
+			"start time: %f end time: %f \n",
                     rank, op->u.recv.source_rank, op->u.recv.dest_rank,
-                    op->u.recv.blocking);
+                    op->u.recv.blocking, op->u.recv.num_bytes,
+		    op->start_time, op->end_time);
             break;
+	case CODES_NW_COL:
+	case CODES_NW_BCAST:
+	case CODES_NW_ALLGATHER:
+	case CODES_NW_ALLGATHERV:
+	case CODES_NW_ALLTOALL:
+	case CODES_NW_ALLTOALLV:
+	case CODES_NW_REDUCE:
+	case CODES_NW_ALLREDUCE:
+            fprintf(f, "op: rank:%d type:collective "
+                       "count: %d \n",
+                    rank, op->u.collective.num_bytes);
+            break;	    
     }
 }