Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
moving smpi into the mainstream build...
authormarkls <markls@48e7efb5-ca39-0410-a469-dd3cf9ba447f>
Sun, 5 Aug 2007 09:41:13 +0000 (09:41 +0000)
committermarkls <markls@48e7efb5-ca39-0410-a469-dd3cf9ba447f>
Sun, 5 Aug 2007 09:41:13 +0000 (09:41 +0000)
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/simgrid/simgrid/trunk@3963 48e7efb5-ca39-0410-a469-dd3cf9ba447f

src/Makefile.am
src/smpi/private.h [new file with mode: 0644]
src/smpi/smpi.h [new file with mode: 0644]
src/smpi/smpi_base.c [new file with mode: 0644]
src/smpi/smpi_mpi.c [new file with mode: 0644]

index d5229da..96f1f21 100644 (file)
@@ -63,7 +63,9 @@ EXTRA_DIST= \
        amok/amok_modinter.h
        \
        include/simix/simix.h include/simix/datatypes.h \
-       simix/msg_simix_private.h
+       simix/msg_simix_private.h \
+       \
+       smpi/private.h smpi/smpi.h
 
 #LIBRARY_VERSION= 0:0:0
 #                 | | |
@@ -179,6 +181,10 @@ SIMIX_SRC= \
   simix/smx_action.c \
   simix/smx_synchro.c
 
+SMPI_SRC= \
+  smpi/smpi_base.c \
+  smpi/smpi_mpi.c
+
 MSG_SRC=  msg/msg_config.c \
   msg/task.c msg/host.c msg/m_process.c msg/gos.c \
   msg/global.c msg/environment.c msg/deployment.c 
@@ -276,7 +282,7 @@ endif
 ### Declare the library content
 ###
 
-lib_LTLIBRARIES= libsimgrid.la libgras.la 
+lib_LTLIBRARIES= libsimgrid.la libgras.la libsmpi.la
 
 libgras_la_SOURCES=  $(XBT_SRC) $(GRAS_COMMON_SRC) $(GRAS_RL_SRC) $(AMOK_SRC)
 libgras_la_LDFLAGS = -no-undefined $(VERSION_INFO)                     @GRAS_DEP@ @LD_DYNAMIC_FLAGS@ -lm 
@@ -287,6 +293,11 @@ libsimgrid_la_SOURCES = $(XBT_SRC) $(SURF_SRC) $(GTNETS_USED) $(SDP_SRC) \
                        $(GRAS_COMMON_SRC) $(GRAS_SG_SRC) $(AMOK_SRC)
 libsimgrid_la_LDFLAGS = -no-undefined $(VERSION_INFO) @SIMGRID_DEP@ @LD_DYNAMIC_FLAGS@ -lm
 
+libsmpi_la_LIBADD = libsimgrid.la
+libsmpi_la_SOURCES = $(SMPI_SRC)
+#libsmpi_la_LDFLAGS = -no-undefined $(VERSION_INFO) @SMPI_DEP@ @LD_DYNAMIC_FLAGS@ -lm
+libsmpi_la_LDFLAGS = -no-undefined $(VERSION_INFO) @LD_DYNAMIC_FLAGS@ -lm
+
 ##
 ## Java cruft
 
diff --git a/src/smpi/private.h b/src/smpi/private.h
new file mode 100644 (file)
index 0000000..fa42b5c
--- /dev/null
@@ -0,0 +1,84 @@
+#include "smpi.h"
+#include "xbt/mallocator.h"
+#include "xbt/xbt_os_time.h"
+
+#define SMPI_DEFAULT_SPEED 100
+#define SMPI_REQUEST_MALLOCATOR_SIZE 100
+#define SMPI_MESSAGE_MALLOCATOR_SIZE 100
+
+typedef struct SMPI_Global {
+
+       // config vars
+       double            reference_speed;
+
+       // state vars
+       int               root_ready:1;
+       int               ready_process_count;
+       smx_mutex_t       start_stop_mutex;
+       smx_cond_t        start_stop_cond;
+
+       xbt_mallocator_t  request_mallocator;
+       xbt_mallocator_t  message_mallocator;
+
+       xbt_fifo_t       *pending_send_request_queues;
+       smx_mutex_t      *pending_send_request_queues_mutexes;
+
+       xbt_fifo_t       *pending_recv_request_queues;
+       smx_mutex_t      *pending_recv_request_queues_mutexes;
+
+       xbt_fifo_t       *received_message_queues;
+       smx_mutex_t      *received_message_queues_mutexes;
+
+       smx_process_t    *sender_processes;
+       smx_process_t    *receiver_processes;
+
+       int               running_hosts_count;
+       smx_mutex_t       running_hosts_count_mutex;
+
+       xbt_os_timer_t   *timers;
+       smx_mutex_t      *timers_mutexes;
+
+
+} s_SMPI_Global_t, *SMPI_Global_t;
+
+extern SMPI_Global_t smpi_global;
+
+struct smpi_received_message_t {
+       smpi_mpi_communicator_t *comm;
+       int src;
+       int dst;
+       int tag;
+       void *buf;
+};
+
+typedef struct smpi_received_message_t smpi_received_message_t;
+
+// function prototypes
+int smpi_mpi_comm_size(smpi_mpi_communicator_t *comm);
+int smpi_mpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host);
+int smpi_mpi_comm_rank_self(smpi_mpi_communicator_t *comm);
+int smpi_mpi_comm_world_rank_self(void);
+int smpi_sender(int argc, char **argv);
+int smpi_receiver(int argc, char **argv);
+void *smpi_request_new(void);
+void smpi_request_free(void *pointer);
+void smpi_request_reset(void *pointer);
+void *smpi_message_new(void);
+void smpi_message_free(void *pointer);
+void smpi_message_reset(void *pointer);
+void smpi_global_init(void);
+void smpi_global_destroy(void);
+int smpi_run_simulation(int argc, char **argv);
+void smpi_mpi_land_func(void *x, void *y, void *z);
+void smpi_mpi_sum_func(void *x, void *y, void *z);
+void smpi_mpi_init(void);
+void smpi_mpi_finalize(void);
+void smpi_bench_begin(void);
+void smpi_bench_end(void);
+void smpi_barrier(smpi_mpi_communicator_t *comm);
+int smpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host);
+int smpi_create_request(void *buf, int count, smpi_mpi_datatype_t *datatype,
+       int src, int dst, int tag, smpi_mpi_communicator_t *comm, smpi_mpi_request_t **request);
+int smpi_isend(smpi_mpi_request_t *request);
+int smpi_irecv(smpi_mpi_request_t *request);
+void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status);
diff --git a/src/smpi/smpi.h b/src/smpi/smpi.h
new file mode 100644 (file)
index 0000000..5061bef
--- /dev/null
@@ -0,0 +1,114 @@
+#include <xbt/function_types.h>
+#include <simix/simix.h>
+
+#define SMPI_RAND_SEED 5
+
+#define MPI_ANY_SOURCE -1
+
+// errorcodes
+#define MPI_SUCCESS     0
+#define MPI_ERR_COMM    1
+#define MPI_ERR_ARG     2
+#define MPI_ERR_TYPE    3
+#define MPI_ERR_REQUEST 4
+#define MPI_ERR_INTERN  5
+#define MPI_ERR_COUNT   6
+#define MPI_ERR_RANK    7
+#define MPI_ERR_TAG     8
+
+// MPI_Comm
+struct smpi_mpi_communicator_t {
+  int            size;
+  smx_host_t    *hosts;
+
+  smx_process_t *processes;
+  int            barrier_count;
+  smx_mutex_t    barrier_mutex;
+  smx_cond_t     barrier_cond;
+};
+typedef struct smpi_mpi_communicator_t smpi_mpi_communicator_t;
+typedef smpi_mpi_communicator_t *MPI_Comm;
+
+// MPI_Status
+struct smpi_mpi_status_t {
+  int MPI_SOURCE;
+};
+typedef struct smpi_mpi_status_t smpi_mpi_status_t;
+typedef smpi_mpi_status_t MPI_Status;
+
+// MPI_Datatype
+struct smpi_mpi_datatype_t {
+  size_t size;
+};
+typedef struct smpi_mpi_datatype_t smpi_mpi_datatype_t;
+typedef smpi_mpi_datatype_t *MPI_Datatype;
+
+// MPI_Request
+struct smpi_mpi_request_t {
+       smpi_mpi_communicator_t *comm;
+       int src;
+       int dst;
+       int tag;
+
+       void *buf;
+       smpi_mpi_datatype_t *datatype;
+       int count;
+
+       short int completed :1;
+       smx_mutex_t mutex;
+       smx_cond_t  cond;
+};
+typedef struct smpi_mpi_request_t smpi_mpi_request_t;
+typedef smpi_mpi_request_t *MPI_Request;
+
+// MPI_Op
+struct smpi_mpi_op_t {
+  void (*func)(void *x, void *y, void *z);
+};
+typedef struct smpi_mpi_op_t smpi_mpi_op_t;
+typedef smpi_mpi_op_t *MPI_Op;
+
+// global SMPI data structure
+typedef struct SMPI_MPI_Global {
+
+       smpi_mpi_communicator_t *mpi_comm_world;
+
+       smpi_mpi_datatype_t     *mpi_byte;
+       smpi_mpi_datatype_t     *mpi_int;
+       smpi_mpi_datatype_t     *mpi_double;
+
+       smpi_mpi_op_t           *mpi_land;
+       smpi_mpi_op_t           *mpi_sum;
+
+} s_SMPI_MPI_Global_t, *SMPI_MPI_Global_t;
+extern SMPI_MPI_Global_t smpi_mpi_global;
+
+#define MPI_COMM_WORLD    (smpi_mpi_global->mpi_comm_world)
+
+#define MPI_STATUS_IGNORE NULL
+
+#define MPI_BYTE          (smpi_mpi_global->mpi_byte)
+#define MPI_DOUBLE        (smpi_mpi_global->mpi_double)
+#define MPI_INT           (smpi_mpi_global->mpi_int)
+
+#define MPI_LAND          (smpi_mpi_global->mpi_land)
+#define MPI_SUM           (smpi_mpi_glboal->mpi_sum)
+
+// MPI Functions
+int MPI_Init(int *argc, char ***argv);
+int MPI_Finalize(void);
+int MPI_Abort(MPI_Comm comm, int errorcode);
+int MPI_Comm_size(MPI_Comm comm, int *size);
+int MPI_Comm_rank(MPI_Comm comm, int *rank);
+int MPI_Type_size(MPI_Datatype datatype, size_t *size);
+int MPI_Barrier(MPI_Comm comm);
+int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Request *request);
+int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status *status);
+int MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request *request);
+int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm);
+
+// smpi functions
+extern int smpi_simulated_main(int argc, char **argv);
+unsigned int smpi_sleep(unsigned int);
+void smpi_exit(int);
+int smpi_gettimeofday(struct timeval *tv, struct timezone *tz);
diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c
new file mode 100644 (file)
index 0000000..c2d0939
--- /dev/null
@@ -0,0 +1,816 @@
+#include <stdio.h>
+#include <signal.h>
+#include <sys/time.h>
+
+#include "private.h"
+
+SMPI_Global_t     smpi_global     = NULL;
+
+SMPI_MPI_Global_t smpi_mpi_global = NULL;
+
+XBT_LOG_NEW_DEFAULT_CATEGORY(smpi, "SMPI");
+
+int inline smpi_mpi_comm_size(smpi_mpi_communicator_t *comm)
+{
+       return comm->size;
+}
+
+// FIXME: smarter algorithm?
+int smpi_mpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host)
+{
+       int i;
+
+       for(i = comm->size - 1; i > 0 && host != comm->hosts[i]; i--);
+
+       return i;
+}
+
+int inline smpi_mpi_comm_rank_self(smpi_mpi_communicator_t *comm)
+{
+       return smpi_mpi_comm_rank(comm, SIMIX_host_self());
+}
+
+//int smpi_mpi_comm_world_rank_self()
+//{
+//     return smpi_mpi_comm_rank(smpi_mpi_global->mpi_comm_world, SIMIX_host_self());
+//}
+
+int smpi_sender(int argc, char **argv)
+{
+       smx_process_t self;
+       smx_host_t shost;
+       int rank;
+
+       xbt_fifo_t request_queue;
+       smx_mutex_t request_queue_mutex;
+       int size;
+
+       int running_hosts_count;
+
+       smpi_mpi_request_t *request;
+
+       smx_host_t dhost;
+
+       smx_action_t communicate_action;
+
+       smpi_received_message_t *message;
+
+       int drank;
+
+       smx_process_t receiver_process;
+
+       self  = SIMIX_process_self();
+       shost = SIMIX_host_self();
+       rank  = smpi_mpi_comm_rank(smpi_mpi_global->mpi_comm_world, shost);
+
+       // make sure root is done before own initialization
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       if (!smpi_global->root_ready) {
+               SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex);
+       }
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       request_queue       = smpi_global->pending_send_request_queues[rank];
+       request_queue_mutex = smpi_global->pending_send_request_queues_mutexes[rank];
+       size                = smpi_mpi_comm_size(smpi_mpi_global->mpi_comm_world);
+
+       smpi_global->sender_processes[rank] = self;
+
+       // wait for all nodes to signal initializatin complete
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       smpi_global->ready_process_count++;
+       if (smpi_global->ready_process_count < 3 * size) {
+               SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex);
+       } else {
+               SIMIX_cond_broadcast(smpi_global->start_stop_cond);
+       }
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       do {
+
+               SIMIX_mutex_lock(request_queue_mutex);
+               request = xbt_fifo_shift(request_queue);
+               SIMIX_mutex_unlock(request_queue_mutex);
+
+               if (NULL == request) {
+                       SIMIX_process_suspend(self);
+               } else {
+
+                       SIMIX_mutex_lock(request->mutex);
+
+                       // copy request to appropriate received queue
+                       message       = xbt_mallocator_get(smpi_global->message_mallocator);
+                       message->comm = request->comm;
+                       message->src  = request->src;
+                       message->dst  = request->dst;
+                       message->tag  = request->tag;
+                       message->buf  = xbt_malloc(request->datatype->size * request->count);
+                       memcpy(message->buf, request->buf, request->datatype->size * request->count);
+
+                       dhost = request->comm->hosts[request->dst];
+                       drank = smpi_mpi_comm_rank(smpi_mpi_global->mpi_comm_world, dhost);
+
+                       SIMIX_mutex_lock(smpi_global->received_message_queues_mutexes[drank]);
+                       xbt_fifo_push(smpi_global->received_message_queues[drank], message);
+                       SIMIX_mutex_unlock(smpi_global->received_message_queues_mutexes[drank]);
+
+                       request->completed = 1;
+
+                       communicate_action = SIMIX_action_communicate(shost, dhost,
+                               NULL, request->datatype->size * request->count * 1.0, -1.0);
+
+                       SIMIX_register_condition_to_action(communicate_action, request->cond);
+                       SIMIX_register_action_to_condition(communicate_action, request->cond);
+
+                       SIMIX_cond_wait(request->cond, request->mutex);
+
+                       SIMIX_mutex_unlock(request->mutex);
+
+                       // wake up receiver if necessary
+                       receiver_process = smpi_global->receiver_processes[drank];
+
+                       if (SIMIX_process_is_suspended(receiver_process)) {
+                               SIMIX_process_resume(receiver_process);
+                       }
+
+               }
+
+               SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex);
+               running_hosts_count = smpi_global->running_hosts_count;
+               SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex);
+
+       } while (0 < running_hosts_count);
+
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       smpi_global->ready_process_count--;
+       if (smpi_global->ready_process_count == 0) {
+               SIMIX_cond_broadcast(smpi_global->start_stop_cond);
+       } else if (smpi_global->ready_process_count < 0) {
+               // FIXME: can't happen! abort!
+       }
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       return 0;
+}
+
+int smpi_receiver(int argc, char **argv)
+{
+       smx_process_t self;
+       int rank;
+
+       xbt_fifo_t request_queue;
+       smx_mutex_t request_queue_mutex;
+       xbt_fifo_t message_queue;
+       smx_mutex_t message_queue_mutex;
+       int size;
+
+       int running_hosts_count;
+
+       smpi_mpi_request_t *request;
+       smpi_received_message_t *message;
+
+       xbt_fifo_item_t request_item;
+       xbt_fifo_item_t message_item;
+
+       self  = SIMIX_process_self();
+       rank  = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world);
+
+       // make sure root is done before own initialization
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       if (!smpi_global->root_ready) {
+               SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex);
+       }
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       request_queue       = smpi_global->pending_recv_request_queues[rank];
+       request_queue_mutex = smpi_global->pending_recv_request_queues_mutexes[rank];
+       message_queue       = smpi_global->received_message_queues[rank];
+       message_queue_mutex = smpi_global->received_message_queues_mutexes[rank];
+       size                = smpi_mpi_comm_size(smpi_mpi_global->mpi_comm_world);
+
+       smpi_global->receiver_processes[rank] = self;
+
+       // wait for all nodes to signal initializatin complete
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       smpi_global->ready_process_count++;
+       if (smpi_global->ready_process_count < 3 * size) {
+               SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex);
+       } else {
+               SIMIX_cond_broadcast(smpi_global->start_stop_cond);
+       }
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       do {
+               request = NULL;
+               message = NULL;
+
+               // FIXME: better algorithm, maybe some kind of balanced tree? or a heap?
+
+               // FIXME: not the best way to request multiple locks...
+               SIMIX_mutex_lock(request_queue_mutex);
+               SIMIX_mutex_lock(message_queue_mutex);
+               for (request_item = xbt_fifo_get_first_item(request_queue);
+                       NULL != request_item;
+                       request_item = xbt_fifo_get_next_item(request_item)) {
+                       request = xbt_fifo_get_item_content(request_item);
+                       for (message_item = xbt_fifo_get_first_item(message_queue);
+                               NULL != message_item;
+                               message_item = xbt_fifo_get_next_item(message_item)) {
+                               message = xbt_fifo_get_item_content(message_item);
+                               if (request->comm == message->comm &&
+                                               (MPI_ANY_SOURCE == request->src || request->src == message->src) &&
+                                               request->tag == message->tag) {
+                                       xbt_fifo_remove_item(request_queue, request_item);
+                                       xbt_fifo_remove_item(message_queue, message_item);
+                                       goto stopsearch;
+                               }
+                       }
+               }
+stopsearch:
+               SIMIX_mutex_unlock(message_queue_mutex);
+               SIMIX_mutex_unlock(request_queue_mutex);
+
+               if (NULL == request || NULL == message) {
+                       SIMIX_process_suspend(self);
+               } else {
+                       SIMIX_mutex_lock(request->mutex);
+
+                       memcpy(request->buf, message->buf, request->datatype->size * request->count);
+                       request->src = message->src;
+                       request->completed = 1;
+                       SIMIX_cond_broadcast(request->cond);
+
+                       SIMIX_mutex_unlock(request->mutex);
+
+                       xbt_free(message->buf);
+                       xbt_mallocator_release(smpi_global->message_mallocator, message);
+               }
+
+               SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex);
+               running_hosts_count = smpi_global->running_hosts_count;
+               SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex);
+
+       } while (0 < running_hosts_count);
+
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       smpi_global->ready_process_count--;
+       if (smpi_global->ready_process_count == 0) {
+               SIMIX_cond_broadcast(smpi_global->start_stop_cond);
+       } else if (smpi_global->ready_process_count < 0) {
+               // FIXME: can't happen, abort!
+       }
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       return 0;
+}
+
+void *smpi_request_new()
+{
+       smpi_mpi_request_t *request = xbt_new(smpi_mpi_request_t, 1);
+
+       request->completed = 0;
+       request->mutex     = SIMIX_mutex_init();
+       request->cond      = SIMIX_cond_init();
+
+       return request;
+}
+
+void smpi_request_free(void *pointer)
+{
+
+       smpi_mpi_request_t *request = pointer;
+
+       if (NULL != request) {
+               SIMIX_cond_destroy(request->cond);
+               SIMIX_mutex_destroy(request->mutex);
+               xbt_free(request);
+       }
+
+       return;
+}
+
+void smpi_request_reset(void *pointer)
+{
+       return;
+}
+
+
+void *smpi_message_new()
+{
+       return xbt_new(smpi_received_message_t, 1);
+}
+
+void smpi_message_free(void *pointer)
+{
+       if (NULL != pointer) {
+               xbt_free(pointer);
+       }
+
+       return;
+}
+
+void smpi_message_reset(void *pointer)
+{
+       return;
+}
+
+void smpi_global_init()
+{
+       int i;
+
+       int size = SIMIX_host_get_number();
+
+       smpi_global = xbt_new(s_SMPI_Global_t, 1);
+
+       // config variable
+       smpi_global->reference_speed                     = SMPI_DEFAULT_SPEED;
+
+       smpi_global->root_ready                          = 0;
+       smpi_global->ready_process_count                 = 0;
+
+       // start/stop
+       smpi_global->start_stop_mutex                    = SIMIX_mutex_init();
+       smpi_global->start_stop_cond                     = SIMIX_cond_init();
+
+       // processes
+       smpi_global->sender_processes                    = xbt_new(smx_process_t, size);
+       smpi_global->receiver_processes                  = xbt_new(smx_process_t, size);
+
+       // running hosts
+       smpi_global->running_hosts_count_mutex           = SIMIX_mutex_init();
+        smpi_global->running_hosts_count                 = 0;
+
+       // mallocators
+       smpi_global->request_mallocator                  = xbt_mallocator_new(SMPI_REQUEST_MALLOCATOR_SIZE,
+                                                            smpi_request_new, smpi_request_free, smpi_request_reset);
+       smpi_global->message_mallocator                  = xbt_mallocator_new(SMPI_MESSAGE_MALLOCATOR_SIZE,
+                                                            smpi_message_new, smpi_message_free, smpi_message_reset);
+
+       //
+       smpi_global->pending_send_request_queues         = xbt_new(xbt_fifo_t,  size);
+       smpi_global->pending_send_request_queues_mutexes = xbt_new(smx_mutex_t, size);
+       smpi_global->pending_recv_request_queues         = xbt_new(xbt_fifo_t,  size);
+       smpi_global->pending_recv_request_queues_mutexes = xbt_new(smx_mutex_t, size);
+       smpi_global->received_message_queues             = xbt_new(xbt_fifo_t,  size);
+       smpi_global->received_message_queues_mutexes     = xbt_new(smx_mutex_t, size);
+       smpi_global->timers                              = xbt_new(xbt_os_timer_t, size);
+       smpi_global->timers_mutexes                      = xbt_new(smx_mutex_t, size);
+
+       for(i = 0; i < size; i++) {
+               smpi_global->pending_send_request_queues[i]         = xbt_fifo_new();
+               smpi_global->pending_send_request_queues_mutexes[i] = SIMIX_mutex_init();
+               smpi_global->pending_recv_request_queues[i]         = xbt_fifo_new();
+               smpi_global->pending_recv_request_queues_mutexes[i] = SIMIX_mutex_init();
+               smpi_global->received_message_queues[i]             = xbt_fifo_new();
+               smpi_global->received_message_queues_mutexes[i]     = SIMIX_mutex_init();
+               smpi_global->timers[i]                              = xbt_os_timer_new();
+               smpi_global->timers_mutexes[i]                      = SIMIX_mutex_init();
+       }
+
+}
+
+void smpi_global_destroy()
+{
+       int i;
+
+       int size = SIMIX_host_get_number();
+
+       // start/stop
+       SIMIX_mutex_destroy(smpi_global->start_stop_mutex);
+       SIMIX_cond_destroy(smpi_global->start_stop_cond);
+
+       // processes
+       xbt_free(smpi_global->sender_processes);
+       xbt_free(smpi_global->receiver_processes);
+
+       // running hosts
+       SIMIX_mutex_destroy(smpi_global->running_hosts_count_mutex);
+
+       // mallocators
+       xbt_mallocator_free(smpi_global->request_mallocator);
+       xbt_mallocator_free(smpi_global->message_mallocator);
+
+       for(i = 0; i < size; i++) {
+               xbt_fifo_free(smpi_global->pending_send_request_queues[i]);
+               SIMIX_mutex_destroy(smpi_global->pending_send_request_queues_mutexes[i]);
+               xbt_fifo_free(smpi_global->pending_recv_request_queues[i]);
+               SIMIX_mutex_destroy(smpi_global->pending_recv_request_queues_mutexes[i]);
+               xbt_fifo_free(smpi_global->received_message_queues[i]);
+               SIMIX_mutex_destroy(smpi_global->received_message_queues_mutexes[i]);
+               xbt_os_timer_free(smpi_global->timers[i]);
+               SIMIX_mutex_destroy(smpi_global->timers_mutexes[i]);
+       }
+
+       xbt_free(smpi_global->pending_send_request_queues);
+       xbt_free(smpi_global->pending_send_request_queues_mutexes);
+       xbt_free(smpi_global->pending_recv_request_queues);
+       xbt_free(smpi_global->pending_recv_request_queues_mutexes);
+       xbt_free(smpi_global->received_message_queues);
+       xbt_free(smpi_global->received_message_queues_mutexes);
+       xbt_free(smpi_global->timers);
+       xbt_free(smpi_global->timers_mutexes);
+
+       xbt_free(smpi_global);
+}
+
+int smpi_run_simulation(int argc, char **argv)
+{
+       smx_cond_t   cond           = NULL;
+       smx_action_t action         = NULL;
+
+       xbt_fifo_t   actions_failed = xbt_fifo_new();
+       xbt_fifo_t   actions_done   = xbt_fifo_new();
+
+       srand(SMPI_RAND_SEED);
+
+       SIMIX_global_init(&argc, argv);
+
+       SIMIX_function_register("smpi_simulated_main", smpi_simulated_main);
+       SIMIX_function_register("smpi_sender",         smpi_sender);
+       SIMIX_function_register("smpi_receiver",       smpi_receiver);
+
+       // FIXME: ought to verify these files...
+       SIMIX_create_environment(argv[1]);
+
+       // must initialize globals between creating environment and launching app....
+       smpi_global_init();
+
+       SIMIX_launch_application(argv[2]);
+
+       /* Prepare to display some more info when dying on Ctrl-C pressing */
+       // FIXME: doesn't work
+       //signal(SIGINT, inthandler);
+
+       /* Clean IO before the run */
+       fflush(stdout);
+       fflush(stderr);
+
+       while (SIMIX_solve(actions_done, actions_failed) != -1.0) {
+               while ((action = xbt_fifo_pop(actions_failed))) {
+                       DEBUG1("** %s failed **", action->name);
+                       while ((cond = xbt_fifo_pop(action->cond_list))) {
+                               SIMIX_cond_broadcast(cond);
+                       }
+                       SIMIX_action_destroy(action);
+               }
+               while ((action = xbt_fifo_pop(actions_done))) {
+                       DEBUG1("** %s done **",action->name);
+                       while ((cond = xbt_fifo_pop(action->cond_list))) {
+                               SIMIX_cond_broadcast(cond);
+                       }
+                       SIMIX_action_destroy(action);
+               }
+       }
+
+       xbt_fifo_free(actions_failed);
+       xbt_fifo_free(actions_done);
+
+       INFO1("simulation time %g", SIMIX_get_clock());
+
+       smpi_global_destroy();
+
+       SIMIX_clean();
+
+       return 0;
+}
+
+void smpi_mpi_land_func(void *x, void *y, void *z)
+{
+       *(int *)z = *(int *)x && *(int *)y;
+}
+
+void smpi_mpi_sum_func(void *x, void *y, void *z)
+{
+       *(int *)z = *(int *)x + *(int *)y;
+}
+
+
+void smpi_mpi_init()
+{
+       smx_process_t process;
+       smx_host_t host;
+       smx_host_t *hosts;
+       int size;
+
+       SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex);
+       smpi_global->running_hosts_count++;
+       SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex);
+
+       // initialize some local variables
+       process = SIMIX_process_self();
+       host    = SIMIX_host_self();
+       hosts   = SIMIX_host_get_table();
+       size    = SIMIX_host_get_number();
+
+       // node 0 sets the globals
+       if (host == hosts[0]) {
+
+               smpi_mpi_global                                = xbt_new(s_SMPI_MPI_Global_t, 1);
+
+               // global communicator
+               smpi_mpi_global->mpi_comm_world                = xbt_new(smpi_mpi_communicator_t, 1);
+               smpi_mpi_global->mpi_comm_world->size          = size;
+               smpi_mpi_global->mpi_comm_world->barrier_count = 0;
+               smpi_mpi_global->mpi_comm_world->barrier_mutex = SIMIX_mutex_init();
+               smpi_mpi_global->mpi_comm_world->barrier_cond  = SIMIX_cond_init();
+               smpi_mpi_global->mpi_comm_world->hosts         = hosts;
+               smpi_mpi_global->mpi_comm_world->processes     = xbt_new(smx_process_t, size);
+               smpi_mpi_global->mpi_comm_world->processes[0]  = process;
+
+               // mpi datatypes
+               smpi_mpi_global->mpi_byte                      = xbt_new(smpi_mpi_datatype_t, 1);
+               smpi_mpi_global->mpi_byte->size                = (size_t)1;
+               smpi_mpi_global->mpi_int                       = xbt_new(smpi_mpi_datatype_t, 1);
+               smpi_mpi_global->mpi_int->size                 = sizeof(int);
+               smpi_mpi_global->mpi_double                    = xbt_new(smpi_mpi_datatype_t, 1);
+               smpi_mpi_global->mpi_double->size              = sizeof(double);
+
+               // mpi operations
+               smpi_mpi_global->mpi_land                      = xbt_new(smpi_mpi_op_t, 1);
+               smpi_mpi_global->mpi_land->func                = smpi_mpi_land_func;
+               smpi_mpi_global->mpi_sum                       = xbt_new(smpi_mpi_op_t, 1);
+               smpi_mpi_global->mpi_sum->func                 = smpi_mpi_sum_func;
+
+               // signal all nodes to perform initialization
+               SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+               smpi_global->root_ready = 1;
+               SIMIX_cond_broadcast(smpi_global->start_stop_cond);
+               SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       } else {
+
+               // make sure root is done before own initialization
+               SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+               if (!smpi_global->root_ready) {
+                       SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex);
+               }
+               SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+               smpi_mpi_global->mpi_comm_world->processes[smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world)] = process;
+       }
+
+       // wait for all nodes to signal initializatin complete
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       smpi_global->ready_process_count++;
+       if (smpi_global->ready_process_count < 3 * size) {
+               SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex);
+       } else {
+               SIMIX_cond_broadcast(smpi_global->start_stop_cond);
+       }
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       return;
+}
+
+void smpi_mpi_finalize()
+{
+       int i;
+
+       SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex);
+       i = --smpi_global->running_hosts_count;
+       SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex);
+
+       SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+       smpi_global->ready_process_count--;
+       SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+       if (0 >= i) {
+
+               // wake up senders/receivers
+               for (i = 0; i < smpi_mpi_global->mpi_comm_world->size; i++) {
+                       if (SIMIX_process_is_suspended(smpi_global->sender_processes[i])) {
+                               SIMIX_process_resume(smpi_global->sender_processes[i]);
+                       }
+                       if (SIMIX_process_is_suspended(smpi_global->receiver_processes[i])) {
+                               SIMIX_process_resume(smpi_global->receiver_processes[i]);
+                       }
+               }
+
+               // wait for senders/receivers to exit...
+               SIMIX_mutex_lock(smpi_global->start_stop_mutex);
+               if (smpi_global->ready_process_count > 0) {
+                       SIMIX_cond_wait(smpi_global->start_stop_cond, smpi_global->start_stop_mutex);
+               }
+               SIMIX_mutex_unlock(smpi_global->start_stop_mutex);
+
+               SIMIX_mutex_destroy(smpi_mpi_global->mpi_comm_world->barrier_mutex);
+               SIMIX_cond_destroy(smpi_mpi_global->mpi_comm_world->barrier_cond);
+               xbt_free(smpi_mpi_global->mpi_comm_world->processes);
+               xbt_free(smpi_mpi_global->mpi_comm_world);
+
+               xbt_free(smpi_mpi_global->mpi_byte);
+               xbt_free(smpi_mpi_global->mpi_int);
+               xbt_free(smpi_mpi_global->mpi_double);
+
+               xbt_free(smpi_mpi_global->mpi_land);
+               xbt_free(smpi_mpi_global->mpi_sum);
+
+               xbt_free(smpi_mpi_global);
+       }
+
+}
+
+// FIXME: could cause trouble with multithreaded procs on same host...
+void smpi_bench_begin()
+{
+       int rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world);
+       SIMIX_mutex_lock(smpi_global->timers_mutexes[rank]);
+       xbt_os_timer_start(smpi_global->timers[rank]);
+       return;
+}
+
+void smpi_bench_end()
+{
+       int rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world);
+       double duration;
+       smx_host_t host;
+       smx_action_t compute_action;
+       smx_mutex_t mutex;
+       smx_cond_t cond;
+
+       xbt_os_timer_stop(smpi_global->timers[rank]);
+
+       duration       = xbt_os_timer_elapsed(smpi_global->timers[rank]);
+       SIMIX_mutex_unlock(smpi_global->timers_mutexes[rank]);
+
+       host           = smpi_mpi_global->mpi_comm_world->hosts[rank];
+       compute_action = SIMIX_action_execute(host, NULL, duration * SMPI_DEFAULT_SPEED);
+       mutex          = SIMIX_mutex_init();
+       cond           = SIMIX_cond_init();
+
+       SIMIX_register_condition_to_action(compute_action, cond);
+       SIMIX_register_action_to_condition(compute_action, cond);
+       SIMIX_mutex_lock(mutex);
+       SIMIX_cond_wait(cond, mutex);
+       SIMIX_mutex_unlock(mutex);
+
+       SIMIX_mutex_destroy(mutex);
+       SIMIX_cond_destroy(cond);
+
+       // FIXME: check for success/failure?
+
+       return;
+}
+
+void smpi_barrier(smpi_mpi_communicator_t *comm)
+{
+
+       SIMIX_mutex_lock(comm->barrier_mutex);
+       if(++comm->barrier_count < comm->size) {
+               SIMIX_cond_wait(comm->barrier_cond, comm->barrier_mutex);
+       } else {
+               comm->barrier_count = 0;
+               SIMIX_cond_broadcast(comm->barrier_cond);
+       }
+       SIMIX_mutex_unlock(comm->barrier_mutex);
+
+       return;
+}
+
+// FIXME: smarter algorithm...
+int smpi_comm_rank(smpi_mpi_communicator_t *comm, smx_host_t host)
+{
+       int i;
+       for(i = 0; i < comm->size && host != comm->hosts[i]; i++);
+       if (i >= comm->size) i = -1;
+       return i;
+}
+
+int smpi_create_request(void *buf, int count, smpi_mpi_datatype_t *datatype,
+       int src, int dst, int tag, smpi_mpi_communicator_t *comm, smpi_mpi_request_t **request)
+{
+       int retval = MPI_SUCCESS;
+
+       *request = NULL;
+
+       if (0 > count) {
+               retval = MPI_ERR_COUNT;
+       } else if (NULL == buf) {
+               retval = MPI_ERR_INTERN;
+       } else if (NULL == datatype) {
+               retval = MPI_ERR_TYPE;
+       } else if (NULL == comm) {
+               retval = MPI_ERR_COMM;
+       } else if (MPI_ANY_SOURCE != src && (0 > src || comm->size <= src)) {
+               retval = MPI_ERR_RANK;
+       } else if (0 > dst || comm->size <= dst) {
+               retval = MPI_ERR_RANK;
+       } else if (0 > tag) {
+               retval = MPI_ERR_TAG;
+       } else {
+               *request = xbt_mallocator_get(smpi_global->request_mallocator);
+               (*request)->comm       = comm;
+               (*request)->src        = src;
+               (*request)->dst        = dst;
+               (*request)->tag        = tag;
+               (*request)->buf        = buf;
+               (*request)->count      = count;
+               (*request)->datatype   = datatype;
+       }
+       return retval;
+}
+
+int smpi_isend(smpi_mpi_request_t *request)
+{
+       int retval = MPI_SUCCESS;
+       int rank   = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world);
+
+       if (NULL != request) {
+               SIMIX_mutex_lock(smpi_global->pending_send_request_queues_mutexes[rank]);
+               xbt_fifo_push(smpi_global->pending_send_request_queues[rank], request);
+               SIMIX_mutex_unlock(smpi_global->pending_send_request_queues_mutexes[rank]);
+       }
+
+       if (SIMIX_process_is_suspended(smpi_global->sender_processes[rank])) {
+               SIMIX_process_resume(smpi_global->sender_processes[rank]);
+       }
+
+       return retval;
+}
+
+int smpi_irecv(smpi_mpi_request_t *request)
+{
+       int retval = MPI_SUCCESS;
+       int rank = smpi_mpi_comm_rank_self(smpi_mpi_global->mpi_comm_world);
+
+       if (NULL != request) {
+               SIMIX_mutex_lock(smpi_global->pending_recv_request_queues_mutexes[rank]);
+               xbt_fifo_push(smpi_global->pending_recv_request_queues[rank], request);
+               SIMIX_mutex_unlock(smpi_global->pending_recv_request_queues_mutexes[rank]);
+       }
+
+       if (SIMIX_process_is_suspended(smpi_global->receiver_processes[rank])) {
+               SIMIX_process_resume(smpi_global->receiver_processes[rank]);
+       }
+
+       return retval;
+}
+
+void smpi_wait(smpi_mpi_request_t *request, smpi_mpi_status_t *status)
+{
+       if (NULL != request) {
+               SIMIX_mutex_lock(request->mutex);
+               if (!request->completed) {
+                       SIMIX_cond_wait(request->cond, request->mutex);
+               }
+               if (NULL != status) {
+                       status->MPI_SOURCE = request->src;
+               }
+               SIMIX_mutex_unlock(request->mutex);
+       }
+}
+
+// FIXME: move into own file
+int smpi_gettimeofday(struct timeval *tv, struct timezone *tz)
+{
+       double now;
+       int retval = 0;
+       smpi_bench_end();
+       if (NULL == tv) {
+               retval = -1;
+       } else {
+               now = SIMIX_get_clock();
+               tv->tv_sec  = now;
+               tv->tv_usec = ((now - (double)tv->tv_sec) * 1000000.0);
+       }
+       smpi_bench_begin();
+       return retval;
+}
+
+unsigned int smpi_sleep(unsigned int seconds)
+{
+       smx_mutex_t mutex;
+       smx_cond_t cond;
+       smx_host_t host;
+       smx_action_t sleep_action;
+
+       smpi_bench_end();
+       host         = SIMIX_host_self();
+       sleep_action = SIMIX_action_sleep(host, seconds);
+       mutex        = SIMIX_mutex_init();
+       cond         = SIMIX_cond_init();
+
+       SIMIX_register_condition_to_action(sleep_action, cond);
+       SIMIX_register_action_to_condition(sleep_action, cond);
+       SIMIX_mutex_lock(mutex);
+       SIMIX_cond_wait(cond, mutex);
+       SIMIX_mutex_unlock(mutex);
+
+       SIMIX_mutex_destroy(mutex);
+       SIMIX_cond_destroy(cond);
+
+       // FIXME: check for success/failure?
+
+       smpi_bench_begin();
+       return 0;
+}
+
+void smpi_exit(int status)
+{
+       smpi_bench_end();
+       SIMIX_mutex_lock(smpi_global->running_hosts_count_mutex);
+       smpi_global->running_hosts_count--;
+       SIMIX_mutex_unlock(smpi_global->running_hosts_count_mutex);
+       SIMIX_process_kill(SIMIX_process_self());
+       return;
+}
diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c
new file mode 100644 (file)
index 0000000..432e57f
--- /dev/null
@@ -0,0 +1,155 @@
+#include <stdio.h>
+#include <sys/time.h>
+#include "private.h"
+
+int MPI_Init(int *argc, char ***argv)
+{
+       smpi_mpi_init();
+       smpi_bench_begin();
+       return MPI_SUCCESS;
+}
+
+int MPI_Finalize()
+{
+       smpi_bench_end();
+       smpi_mpi_finalize();
+       return MPI_SUCCESS;
+}
+
+// right now this just exits the current node, should send abort signal to all
+// hosts in the communicator;
+int MPI_Abort(MPI_Comm comm, int errorcode)
+{
+       smpi_exit(errorcode);
+
+       return 0;
+}
+
+int MPI_Comm_size(MPI_Comm comm, int *size)
+{
+       int retval = MPI_SUCCESS;
+
+       smpi_bench_end();
+
+       if (NULL == comm) {
+               retval = MPI_ERR_COMM;
+       } else if (NULL == size) {
+               retval = MPI_ERR_ARG;
+       } else {
+               *size = comm->size;
+       }
+
+       smpi_bench_begin();
+
+       return retval;
+}
+
+int MPI_Comm_rank(MPI_Comm comm, int *rank)
+{
+       int retval = MPI_SUCCESS;
+
+       smpi_bench_end();
+
+       if (NULL == comm) {
+               retval = MPI_ERR_COMM;
+       } else if (NULL == rank) {
+               retval = MPI_ERR_ARG;
+       } else {
+               *rank = smpi_comm_rank(comm, SIMIX_host_self());
+       }
+
+       smpi_bench_begin();
+
+       return retval;
+}
+
+int MPI_Type_size(MPI_Datatype datatype, size_t *size)
+{
+       int retval = MPI_SUCCESS;
+
+       smpi_bench_end();
+
+       if (NULL == datatype) {
+               retval = MPI_ERR_TYPE;
+       } else if (NULL == size) {
+               retval = MPI_ERR_ARG;
+       } else {
+               *size = datatype->size;
+       }
+
+       smpi_bench_begin();
+
+       return retval;
+}
+
+int MPI_Barrier(MPI_Comm comm)
+{
+       smpi_bench_end();
+       smpi_barrier(comm);
+       smpi_bench_begin();
+       return MPI_SUCCESS;
+}
+
+int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Request *request)
+{
+       int retval = MPI_SUCCESS;
+       int dst;
+       smpi_bench_end();
+       dst = smpi_mpi_comm_rank_self(comm);
+       retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, request);
+       if (NULL != *request) {
+               smpi_irecv(*request);
+       }
+       smpi_bench_begin();
+       return retval;
+}
+
+int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status *status)
+{
+       int retval = MPI_SUCCESS;
+       int dst;
+       smpi_mpi_request_t *request;
+       smpi_bench_end();
+       dst = smpi_mpi_comm_rank_self(comm);
+       retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &request);
+       if (NULL != request) {
+               smpi_irecv(request);
+               smpi_wait(request, status);
+               // FIXME: mallocator
+               //xbt_free(request);
+       }
+       smpi_bench_begin();
+       return retval;
+}
+
+int MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request *request)
+{
+       int retval = MPI_SUCCESS;
+       int src;
+       smpi_bench_end();
+       src = smpi_mpi_comm_rank_self(comm);
+       retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, request);
+       if (NULL != *request) {
+               smpi_isend(*request);
+       }
+       smpi_bench_begin();
+       return retval;
+}
+
+int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+{
+       int retval = MPI_SUCCESS;
+       int src;
+       smpi_mpi_request_t *request;
+       smpi_bench_end();
+       src = smpi_mpi_comm_rank_self(comm);
+       retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &request);
+       if (NULL != request) {
+               smpi_isend(request);
+               smpi_wait(request, MPI_STATUS_IGNORE);
+               // FIXME: mallocator
+               //xbt_free(request)
+       }
+       smpi_bench_begin();
+       return retval;
+}