#include "simgrid/sg_config.h"
#include "smpi/smpi_utils.hpp"
#include "colls/colls.h"
+#include <simgrid/s4u/host.hpp>
#include "src/kernel/activity/SynchroComm.hpp"
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
+extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
+
+
static int match_recv(void* a, void* b, smx_activity_t ignored) {
MPI_Request ref = static_cast<MPI_Request>(a);
MPI_Request req = static_cast<MPI_Request>(b);
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
request->real_size=request->size;
request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
- ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
+ ! smpi_process_get_replaying()? smpi_comm_copy_data_callback
: &smpi_comm_null_copy_buffer_callback, request, -1.0);
XBT_DEBUG("recv simcall posted");
request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
buf, request->real_size, &match_send,
&xbt_free_f, // how to free the userdata if a detached send fails
- !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
+ !smpi_process_get_replaying() ? smpi_comm_copy_data_callback
: &smpi_comm_null_copy_buffer_callback, request,
// detach if msg size < eager/rdv switch limit
request->detached);
// to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
// (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
- // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
+ // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
+ // (This can speed up the execution of certain applications by an order of magnitude, such as HPL)
static int nsleeps = 1;
- if(smpi_iprobe_sleep > 0)
- simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
+ double speed = simgrid::s4u::Actor::self()->host()->speed();
+ double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
+ if (smpi_iprobe_sleep > 0) {
+ smx_activity_t iprobe_sleep = simcall_execution_start("iprobe", /* flops to executek*/nsleeps*smpi_iprobe_sleep*speed*maxrate, /* priority */1.0, /* performance bound */maxrate*speed);
+ simcall_execution_wait(iprobe_sleep);
+ }
// behave like a receive, but don't do it
smx_mailbox_t mailbox;
nsleeps++;
}
smpi_mpi_request_free(&request);
-
- return;
}
void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
*request = MPI_REQUEST_NULL;
}
+static int sort_accumulates(MPI_Request a, MPI_Request b)
+{
+ return (a->tag < b->tag);
+}
+
int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
{
- xbt_dynar_t comms;
+ s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
int i;
int size = 0;
int index = MPI_UNDEFINED;
if(count > 0) {
// Wait for a request to complete
- comms = xbt_dynar_new(sizeof(smx_activity_t), nullptr);
+ xbt_dynar_init(&comms, sizeof(smx_activity_t), nullptr);
map = xbt_new(int, count);
XBT_DEBUG("Wait for one of %d", count);
for(i = 0; i < count; i++) {
if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
if (requests[i]->action != nullptr) {
XBT_DEBUG("Waiting any %p ", requests[i]);
- xbt_dynar_push(comms, &requests[i]->action);
+ xbt_dynar_push(&comms, &requests[i]->action);
map[size] = i;
size++;
- }else{
- //This is a finished detached request, let's return this one
- size=0;//so we free the dynar but don't do the waitany call
- index=i;
- finish_wait(&requests[i], status);//cleanup if refcount = 0
- if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
- requests[i]=MPI_REQUEST_NULL;//set to null
- break;
- }
+ } else {
+ // This is a finished detached request, let's return this one
+ size = 0; // so we free the dynar but don't do the waitany call
+ index = i;
+ finish_wait(&requests[i], status); // cleanup if refcount = 0
+ if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
+ requests[i] = MPI_REQUEST_NULL; // set to null
+ break;
+ }
}
}
if(size > 0) {
- i = simcall_comm_waitany(comms, -1);
+ i = simcall_comm_waitany(&comms, -1);
// not MPI_UNDEFINED, as this is a simix return code
if (i != -1) {
index = map[i];
- finish_wait(&requests[index], status);
- if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
- requests[index] = MPI_REQUEST_NULL;
+ //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
+ if ((requests[index] == MPI_REQUEST_NULL)
+ || (!((requests[index]->flags & ACCUMULATE) && (requests[index]->flags & RECV)))){
+ finish_wait(&requests[index], status);
+ if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
+ requests[index] = MPI_REQUEST_NULL;
+ }else{
+ XBT_WARN("huu?");
+ }
}
}
+
+ xbt_dynar_free_data(&comms);
xbt_free(map);
- xbt_dynar_free(&comms);
}
if (index==MPI_UNDEFINED)
int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
{
- int index, c;
+ std::vector<MPI_Request> accumulates;
+ int index;
MPI_Status stat;
MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
int retvalue = MPI_SUCCESS;
//tag invalid requests in the set
if (status != MPI_STATUSES_IGNORE) {
- for (c = 0; c < count; c++) {
+ for (int c = 0; c < count; c++) {
if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
smpi_empty_status(&status[c]);
} else if (requests[c]->src == MPI_PROC_NULL) {
}
}
}
- for(c = 0; c < count; c++) {
-
+ for (int c = 0; c < count; c++) {
if (MC_is_active() || MC_record_replay_is_active()) {
smpi_mpi_wait(&requests[c], pstat);
index = c;
index = smpi_mpi_waitany(count, requests, pstat);
if (index == MPI_UNDEFINED)
break;
+
+ if (requests[index] != MPI_REQUEST_NULL
+ && (requests[index]->flags & RECV)
+ && (requests[index]->flags & ACCUMULATE))
+ accumulates.push_back(requests[index]);
if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
- requests[index]=MPI_REQUEST_NULL;
+ requests[index] = MPI_REQUEST_NULL;
}
if (status != MPI_STATUSES_IGNORE) {
status[index] = *pstat;
}
}
+ if (!accumulates.empty()) {
+ std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
+ for (auto req : accumulates) {
+ finish_wait(&req, status);
+ }
+ }
+
return retvalue;
}
void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
int system_tag = -888;
- MPI_Aint lb = 0, dataext = 0;
+ MPI_Aint lb = 0;
+ MPI_Aint dataext = 0;
int rank = smpi_comm_rank(comm);
int size = smpi_comm_size(comm);
void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
int system_tag = -888;
- MPI_Aint lb = 0, dataext = 0;
+ MPI_Aint lb = 0;
+ MPI_Aint dataext = 0;
int recvbuf_is_empty=1;
int rank = smpi_comm_rank(comm);
int size = smpi_comm_size(comm);
}
// Wait for completion of all comms.
smpi_mpi_startall(size - 1, requests);
+
if(smpi_op_is_commute(op)){
for (int other = 0; other < size - 1; other++) {
index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
for (int other = 0; other < size - 1; other++) {
smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
if(index < rank) {
- if(recvbuf_is_empty){
- smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
- recvbuf_is_empty=0;
- } else
- smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
+ if (recvbuf_is_empty) {
+ smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
+ recvbuf_is_empty = 0;
+ } else
+ smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
}
}
}
