Merge pull request #174 from ROSS-org/release-7.2.1

Release 7.2.1
- Includes Random Cleanup (#169)

Author: gonsie

CMakeLists.txt

@@ -47,18 +47,11 @@ if (COVERALLS)
     coveralls_turn_on_coverage()
 endif()
 
-# Old way to do this
-# Data Structure for Unprocessed Event List
-#SET(QUEUE calendar) #Calendar Queue
-#SET(QUEUE heap) #Push/Down Heap
-#SET(QUEUE splay) #Splay Tree
-#SET(QUEUE kp_splay) #Splay Tree in KPs
-
-# New way as of CMake 2.8
-# The default value for the QUEUE variable is splay
-# The other options are presented at config time by cmake-gui
-SET(QUEUE splay CACHE STRING "Queue type chosen by the user at configure time")
-SET_PROPERTY(CACHE QUEUE PROPERTY STRINGS splay calendar heap kp_splay)
+# Priority Queue Implementation
+SET(QUEUE splay)
+# Other queue implementations are no longer supported.
+# SET(QUEUE splay CACHE STRING "Queue type chosen by the user at configure time")
+# SET_PROPERTY(CACHE QUEUE PROPERTY STRINGS splay calendar heap kp_splay)
 
 # Random Library
 SET(RAND clcg4)

core/CMakeLists.txt

@@ -53,7 +53,6 @@ tw-timing.c
 
 tw-sched.c
 tw-setup.c
-tw-signal.c
 tw-stats.c
 tw-util.c
 
@@ -94,7 +93,7 @@ set(VERSION_SHORT "${VERSION_MAJOR}.${VERSION_MINOR}.${VERSION_PATCH}")
 
 # Data Structure for remote Events
 # If AVL_TREE is OFF, ROSS reverts to hashing
-OPTION(AVL_TREE "Use AVL trees for optimistic mode events? (hash tabels otherwise)" ON)
+OPTION(AVL_TREE "Use AVL trees for optimistic mode events? (hash tables otherwise)" ON)
 IF(AVL_TREE)
   SET(ross_srcs ${ross_srcs} avl_tree.h avl_tree.c)
 ENDIF(AVL_TREE)

core/gvt/mpi_allreduce.c

@@ -13,8 +13,7 @@ static const tw_optdef gvt_opts [] =
 {
 	TWOPT_GROUP("ROSS MPI GVT"),
 	TWOPT_UINT("gvt-interval", g_tw_gvt_interval, "GVT Interval: Iterations through scheduling loop (synch=1,2,3,4), or ms between GVTs (synch=5)"),
-	TWOPT_DOUBLE("report-interval", gvt_print_interval,
-			"percent of runtime to print GVT"),
+	TWOPT_DOUBLE("report-interval", gvt_print_interval, "percent of runtime to print GVT"),
 	TWOPT_END()
 };
 

core/network-mpi.c

@@ -7,10 +7,15 @@ int custom_communicator = 0;
 /**
  * @struct act_q
  * @brief Keeps track of posted send or recv operations.
+ *
+ * This list structure is used *only* by the network mpi layer (this
+ * file). Within this file, two lists are used, for MPI Irecv and
+ * Isend requests. The MPI requests and statusus are linked with an
+ * event buffer through this struct.
  */
 struct act_q
 {
-  const char *name;
+  const char *name;         /**< name of the list, used in error printouts */
 
   tw_event **event_list;    /**< list of event pointers in this queue */
   MPI_Request *req_list;    /**< list of MPI request handles */
@@ -27,8 +32,8 @@ static struct act_q posted_sends;
 static struct act_q posted_recvs;
 static tw_eventq outq;
 
-static unsigned int read_buffer = 16;
-static unsigned int send_buffer = 1024;
+static unsigned int read_buffer = 16;   /**< Number of Irecv's to buffer, length of posted_recvs queue */
+static unsigned int send_buffer = 1024; /**< Number of Isend's to buffer, length of posted_sends queue */
 static int world_size = 1;
 
 static const tw_optdef mpi_opts[] = {
@@ -85,20 +90,13 @@ tw_net_init(int *argc, char ***argv)
  * @param[in] name name of the queue
  */
 static void
-init_q(struct act_q *q, const char *name)
+init_q(struct act_q *q, const char *name, unsigned int size)
 {
-  unsigned int n;
-
-  if(q == &posted_sends)
-    n = send_buffer;
-  else
-    n = read_buffer;
-
   q->name = name;
-  q->event_list = (tw_event **) tw_calloc(TW_LOC, name, sizeof(*q->event_list), n);
-  q->req_list = (MPI_Request *) tw_calloc(TW_LOC, name, sizeof(*q->req_list), n);
-  q->idx_list = (int *) tw_calloc(TW_LOC, name, sizeof(*q->idx_list), n);
-  q->status_list = (MPI_Status *) tw_calloc(TW_LOC, name, sizeof(*q->status_list), n);
+  q->event_list = (tw_event **) tw_calloc(TW_LOC, name, sizeof(tw_event *), size);
+  q->req_list = (MPI_Request *) tw_calloc(TW_LOC, name, sizeof(MPI_Request), size);
+  q->idx_list = (int *) tw_calloc(TW_LOC, name, sizeof(int), size);
+  q->status_list = (MPI_Status *) tw_calloc(TW_LOC, name, sizeof(MPI_Status), size);
 }
 
 unsigned int
@@ -110,6 +108,7 @@ tw_nnodes(void)
 void
 tw_net_start(void)
 {
+  // sets value of tw_nnodes
   if (MPI_Comm_size(MPI_COMM_ROSS, &world_size) != MPI_SUCCESS)
     tw_error(TW_LOC, "Cannot get MPI_Comm_size(MPI_COMM_ROSS)");
 
@@ -131,22 +130,12 @@ tw_net_start(void)
 
   tw_pe_init();
 
-  //If we're in (some variation of) optimistic mode, we need this hash
-  if (g_tw_synchronization_protocol == OPTIMISTIC ||
-      g_tw_synchronization_protocol == OPTIMISTIC_DEBUG ||
-      g_tw_synchronization_protocol == OPTIMISTIC_REALTIME) {
-    g_tw_pe->hash_t = tw_hash_create();
-  } else {
-    g_tw_pe->hash_t = NULL;
-  }
-
-  if (send_buffer < 1)
-    tw_error(TW_LOC, "network send buffer must be >= 1");
-  if (read_buffer < 1)
-    tw_error(TW_LOC, "network read buffer must be >= 1");
+  // these values are command line options
+  if (send_buffer < 1) tw_error(TW_LOC, "network send buffer must be >= 1");
+  if (read_buffer < 1) tw_error(TW_LOC, "network read buffer must be >= 1");
 
-  init_q(&posted_sends, "MPI send queue");
-  init_q(&posted_recvs, "MPI recv queue");
+  init_q(&posted_sends, "MPI send queue", send_buffer);
+  init_q(&posted_recvs, "MPI recv queue", read_buffer);
 
   g_tw_net_device_size = read_buffer;
 

core/queue/tw-queue.h

@@ -13,4 +13,3 @@ unsigned int tw_pq_max_size(tw_pq *);
 #ifdef ROSS_QUEUE_kp_splay
 tw_eventpq * tw_eventpq_create(void);
 #endif
-

core/ross-extern.h

@@ -143,12 +143,6 @@ extern void tw_scheduler_optimistic(tw_pe * me);
 extern void tw_scheduler_optimistic_debug(tw_pe * me);
 extern void tw_scheduler_optimistic_realtime(tw_pe * me);
 
-/*
- * tw-signal.c
- */
-extern void     tw_sigsegv(int sig);
-extern void     tw_sigterm(int sig);
-
 /*
  * tw-state.c
  */

core/ross-types.h

@@ -379,7 +379,6 @@ struct tw_pe {
     tw_eventq event_q; /**< @brief Linked list of events sent to this PE */
     tw_event *cancel_q; /**< @brief List of canceled events */
     tw_pq *pq; /**< @brief Priority queue used to sort events */
-    tw_kp *kp_list; /**< @brief */
 
     tw_eventq free_q; /**< @brief Linked list of free tw_events */
     tw_event *abort_event; /**< @brief Placeholder event for when free_q is empty */

core/tw-event.c

@@ -31,7 +31,7 @@ void tw_event_send(tw_event * event) {
 #endif
 
      // moved from network-mpi.c in order to give all events a seq_num
-	event->event_id = (tw_eventid) ++send_pe->seq_num;
+     event->event_id = (tw_eventid) ++send_pe->seq_num;
 
     // call LP remote mapping function to get dest_pe
     dest_peid = (*src_lp->type->map) ((tw_lpid) event->dest_lp);

core/tw-eventq.h

@@ -5,6 +5,9 @@
 
 #include <ross.h>
 
+/**
+ * debug assitant fuction
+ */
 static inline void
 tw_eventq_debug(tw_eventq * q)
 {
@@ -39,6 +42,9 @@ tw_eventq_debug(tw_eventq * q)
 #endif
 }
 
+/**
+ * push the contents of one list onto another??
+ */
 static inline void
 tw_eventq_push_list(tw_eventq * q, tw_event * h, tw_event * t, long cnt)
 {
@@ -115,6 +121,15 @@ tw_eventq_push_list(tw_eventq * q, tw_event * h, tw_event * t, long cnt)
     tw_eventq_debug(q);
 }
 
+/**
+ * Given a list, move the portion of its contents that is older than GVT to
+ * the free list.
+ *
+ * Assumptions:
+ * - The provided q is not the free_q
+ * - The head of the list has the maximum time stamp in the list. Therefore,
+ *   if the head is older than GVT, everything in the list is as well.
+ */
 static inline void
 tw_eventq_fossil_collect(tw_eventq *q, tw_pe *pe)
 {
@@ -164,6 +179,9 @@ tw_eventq_fossil_collect(tw_eventq *q, tw_pe *pe)
   }
 }
 
+/**
+ * allocate a events into a given tw_eventq
+ */
 static inline void
 tw_eventq_alloc(tw_eventq * q, unsigned int cnt)
 {
@@ -198,7 +216,6 @@ tw_eventq_alloc(tw_eventq * q, unsigned int cnt)
   g_tw_event_msg_sz = event_len;
 
   // compute number of events needed for the network.
-  g_tw_gvt_threshold = (int) ceil(g_tw_net_device_size / g_tw_event_msg_sz);
   g_tw_gvt_threshold = g_tw_net_device_size;
   g_tw_events_per_pe += g_tw_gvt_threshold;
   cnt += g_tw_gvt_threshold;
@@ -233,6 +250,9 @@ tw_eventq_alloc(tw_eventq * q, unsigned int cnt)
   q->tail = event;
 }
 
+/**
+ * push to tail of list
+ */
 static inline void
 tw_eventq_push(tw_eventq *q, tw_event *e)
 {
@@ -253,12 +273,18 @@ tw_eventq_push(tw_eventq *q, tw_event *e)
   tw_eventq_debug(q);
 }
 
+/**
+ * peek to tail of list
+ */
 static inline tw_event *
 tw_eventq_peek(tw_eventq *q)
 {
   return q->tail;
 }
 
+/**
+ * pop to tail of list
+ */
 static inline tw_event *
 tw_eventq_pop(tw_eventq * q)
 {
@@ -287,6 +313,9 @@ tw_eventq_pop(tw_eventq * q)
   return t;
 }
 
+/**
+ * push to head of list
+ */
 static inline void
 tw_eventq_unshift(tw_eventq *q, tw_event *e)
 {
@@ -308,12 +337,18 @@ tw_eventq_unshift(tw_eventq *q, tw_event *e)
   tw_eventq_debug(q);
 }
 
+/**
+ * peek at head of list
+ */
 static inline tw_event *
 tw_eventq_peek_head(tw_eventq *q)
 {
   return q->head;
 }
 
+/**
+ * pop from head of list
+ */
 static inline tw_event *
 tw_eventq_shift(tw_eventq *q)
 {
@@ -342,6 +377,9 @@ tw_eventq_shift(tw_eventq *q)
   return h;
 }
 
+/**
+ * delete an event from anywhere in the list
+ */
 static inline void
 tw_eventq_delete_any(tw_eventq *q, tw_event *e)
 {
@@ -366,6 +404,10 @@ tw_eventq_delete_any(tw_eventq *q, tw_event *e)
   tw_eventq_debug(q);
 }
 
+/**
+ * pop the entire list.
+ * After this operation, the size of the provided q is 0.
+ */
 static inline tw_event *
 tw_eventq_pop_list(tw_eventq * q)
 {
@@ -377,7 +419,7 @@ tw_eventq_pop_list(tw_eventq * q)
   return h;
 }
 
-/*
+/**
  * The purpose of this function is to be able to remove some
  * part of a list.. could be all of list, from head to some inner
  * buffer, or from some inner buffer to tail.  I only care about the

core/tw-pe.c

@@ -22,29 +22,39 @@ tw_pe_settype(const tw_petype * type)
 #undef copy_pef
 }
 
-/*
- * tw_pe_init: initialize individual PE structs
+/**
+ * initialize individual PE structs
+ *
+ * must be called after tw_nnodes / MPI world size is set.
  *
  */
 void
 tw_pe_init(void)
 {
-    if (g_tw_pe)
-		tw_error(TW_LOC, "PE %u already initialized", g_tw_mynode);
+    if (g_tw_pe) tw_error(TW_LOC, "PE %u already initialized", g_tw_mynode);
 
     g_tw_pe = (tw_pe*)tw_calloc(TW_LOC, "PE Struct", sizeof(*g_tw_pe), 1);
-	tw_petype no_type;
 
-	memset(&no_type, 0, sizeof(no_type));
+    g_tw_pe->id = g_tw_mynode;
 
-	g_tw_pe->id = g_tw_mynode;
-	tw_pe_settype(&no_type);
+    tw_petype no_type;
+    memset(&no_type, 0, sizeof(no_type));
+    tw_pe_settype(&no_type);
 
-	g_tw_pe->trans_msg_ts = TW_STIME_MAX;
-	g_tw_pe->gvt_status = 0;
+    g_tw_pe->trans_msg_ts = DBL_MAX;
+    g_tw_pe->gvt_status = 0;
 
     // TODO is the PE RNG ever actually used?
-	g_tw_pe->rng = tw_rand_init(31, 41);
+    g_tw_pe->rng = tw_rand_init(31, 41);
+
+    //If we're in (some variation of) optimistic mode, we need this hash
+    if (g_tw_synchronization_protocol == OPTIMISTIC ||
+        g_tw_synchronization_protocol == OPTIMISTIC_DEBUG ||
+        g_tw_synchronization_protocol == OPTIMISTIC_REALTIME) {
+        g_tw_pe->hash_t = tw_hash_create();
+    } else {
+        g_tw_pe->hash_t = NULL;
+    }
 
 }