xbt_dynar_t smpi_or_values = NULL;
xbt_dynar_t smpi_ois_values = NULL;
+double smpi_wtime_sleep = 0.0;
+double smpi_iprobe_sleep = 1e-4;
+double smpi_test_sleep = 1e-4;
+
+
// Methods used to parse and store the values for timing injections in smpi
// These are taken from surf/network.c and generalized to have more factors
// These methods should be merged with those in surf/network.c (moved somewhere in xbt ?)
return current;
}
+double smpi_mpi_wtime(){
+ double time;
+ if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
+ smpi_bench_end();
+ time = SIMIX_get_clock();
+ //to avoid deadlocks if called too many times
+ if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep);
+ smpi_bench_begin();
+ } else {
+ time = SIMIX_get_clock();
+ }
+ return time;
+}
+
static MPI_Request build_request(void *buf, int count,
MPI_Datatype datatype, int src, int dst,
int tag, MPI_Comm comm, unsigned flags)
if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) {
void *oldbuf = NULL;
request->detached = 1;
+ XBT_DEBUG("Send request %p is detached", request);
request->refcount++;
if(request->old_type->has_subtype == 0){
oldbuf = request->buf;
&& ((char*)request->buf >= start_data_exe)
&& ((char*)request->buf < start_data_exe + size_data_exe )){
XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
- switch_data_segment(request->src);
+ smpi_switch_data_segment(request->src);
}
buf = xbt_malloc(request->size);
memcpy(buf,oldbuf,request->size);
+ XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
}
}
- XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
}
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
&& ((char*)req->old_buf < start_data_exe + size_data_exe )
){
XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
- switch_data_segment(smpi_process_index());
+ smpi_switch_data_segment(smpi_process_index());
}
}
- if(req->flags & RECV) {
-
- if(datatype->has_subtype == 1){
- // This part handles the problem of non-contignous memory
- // the unserialization at the reception
- s_smpi_subtype_t *subtype = datatype->substruct;
- subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
- if(req->detached == 0) free(req->buf);
- }else{//apply op on contiguous buffer for accumulate
- int n =req->real_size/smpi_datatype_size(datatype);
- smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
- }
+
+ if(datatype->has_subtype == 1){
+ // This part handles the problem of non-contignous memory
+ // the unserialization at the reception
+ s_smpi_subtype_t *subtype = datatype->substruct;
+ if(req->flags & RECV)
+ subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct, req->op);
+ if(req->detached == 0) free(req->buf);
+ }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
+ int n =req->real_size/smpi_datatype_size(datatype);
+ smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
}
}
smpi_comm_unuse(req->comm);
int flag;
//assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
+
+ //to avoid deadlocks
+ //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
+ static int nsleeps = 1;
+ if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
+
smpi_empty_status(status);
flag = 1;
if (!((*request)->flags & PREPARED)) {
flag = simcall_comm_test((*request)->action);
if (flag) {
finish_wait(request, status);
+ nsleeps=1;//reset the number of sleeps we will do next time
if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT))
*request = MPI_REQUEST_NULL;
+ }else{
+ nsleeps++;
}
}
return flag;
}
}
if(size > 0) {
+ //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
+ static int nsleeps = 1;
+ if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep);
+
i = simcall_comm_testany(comms);
// not MPI_UNDEFINED, as this is a simix return code
if(i != -1) {
if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
requests[*index] = MPI_REQUEST_NULL;
flag = 1;
+ nsleeps=1;
+ }else{
+ nsleeps++;
}
}else{
//all requests are null or inactive, return true
comm, PERSISTENT | RECV);
//to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
- double sleeptime= sg_cfg_get_double("smpi/iprobe");
//multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
static int nsleeps = 1;
-
- simcall_process_sleep(sleeptime);
-
+ if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
// behave like a receive, but don't do it
smx_rdv_t mailbox;