X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/48e24b849b73cfd56e92cdfac43b16eb06067b3a..25d1e57ecf83941d9ad4b361e16e6e354c87d585:/src/smpi/smpi_request.cpp diff --git a/src/smpi/smpi_request.cpp b/src/smpi/smpi_request.cpp index f7ca04bc9f..40ae6cde98 100644 --- a/src/smpi/smpi_request.cpp +++ b/src/smpi/smpi_request.cpp @@ -22,94 +22,18 @@ static simgrid::config::Flag smpi_iprobe_sleep( static simgrid::config::Flag smpi_test_sleep( "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4); -std::vector smpi_os_values; -std::vector smpi_or_values; std::vector smpi_ois_values; extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t); -static double smpi_os(size_t size) -{ - if (smpi_os_values.empty()) { - smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os")); - } - double current=smpi_os_values.empty()?0.0:smpi_os_values[0].values[0]+smpi_os_values[0].values[1]*size; - // Iterate over all the sections that were specified and find the right - // value. (fact.factor represents the interval sizes; we want to find the - // section that has fact.factor <= size and no other such fact.factor <= size) - // Note: parse_factor() (used before) already sorts the vector we iterate over! - for (auto& fact : smpi_os_values) { - if (size <= fact.factor) { // Values already too large, use the previously computed value of current! - XBT_DEBUG("os : %zu <= %zu return %.10f", size, fact.factor, current); - return current; - }else{ - // If the next section is too large, the current section must be used. - // Hence, save the cost, as we might have to use it. - current = fact.values[0]+fact.values[1]*size; - } - } - XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current); - - return current; -} - -static double smpi_ois(size_t size) -{ - if (smpi_ois_values.empty()) { - smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois")); - } - double current=smpi_ois_values.empty()?0.0:smpi_ois_values[0].values[0]+smpi_ois_values[0].values[1]*size; - // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval - // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size) - // Note: parse_factor() (used before) already sorts the vector we iterate over! - for (auto& fact : smpi_ois_values) { - if (size <= fact.factor) { // Values already too large, use the previously computed value of current! - XBT_DEBUG("ois : %zu <= %zu return %.10f", size, fact.factor, current); - return current; - }else{ - // If the next section is too large, the current section must be used. - // Hence, save the cost, as we might have to use it. - current = fact.values[0]+fact.values[1]*size; - } - } - XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current); - - return current; -} - -static double smpi_or(size_t size) -{ - if (smpi_or_values.empty()) { - smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or")); - } - - double current=smpi_or_values.empty()?0.0:smpi_or_values.front().values[0]+smpi_or_values.front().values[1]*size; - - // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval - // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size) - // Note: parse_factor() (used before) already sorts the vector we iterate over! - for (auto fact : smpi_or_values) { - if (size <= fact.factor) { // Values already too large, use the previously computed value of current! - XBT_DEBUG("or : %zu <= %zu return %.10f", size, fact.factor, current); - return current; - } else { - // If the next section is too large, the current section must be used. - // Hence, save the cost, as we might have to use it. - current=fact.values[0]+fact.values[1]*size; - } - } - XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current); - - return current; -} - namespace simgrid{ namespace smpi{ Request::Request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags) : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags) { void *old_buf = nullptr; - if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(buf_))){ +// FIXME Handle the case of a partial shared malloc. + if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) { // This part handles the problem of non-contiguous memory old_buf = buf; if (count==0){ @@ -473,7 +397,7 @@ void Request::start() if(!(old_type_->flags() & DT_FLAG_DERIVED)){ oldbuf = buf_; if (!process->replaying() && oldbuf != nullptr && size_!=0){ - if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) + if((smpi_privatize_global_variables != 0) && (static_cast(buf_) >= smpi_start_data_exe) && (static_cast(buf_) < smpi_start_data_exe + smpi_size_data_exe )){ XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment "); @@ -488,9 +412,11 @@ void Request::start() //if we are giving back the control to the user without waiting for completion, we have to inject timings double sleeptime = 0.0; - if(detached_ != 0 || ((flags_ & (ISEND|SSEND)) != 0)){// issend should be treated as isend - //isend and send timings may be different - sleeptime = ((flags_ & ISEND) != 0) ? smpi_ois(size_) : smpi_os(size_); + if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend + // isend and send timings may be different + sleeptime = ((flags_ & ISEND) != 0) + ? simgrid::s4u::Actor::self()->host()->extension()->oisend(size_) + : simgrid::s4u::Actor::self()->host()->extension()->osend(size_); } if(sleeptime > 0.0){ @@ -764,10 +690,12 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status) req->print_request("Finishing"); MPI_Datatype datatype = req->old_type_; - if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(req->old_buf_))){ +// FIXME Handle the case of a partial shared malloc. + if (((req->flags_ & ACCUMULATE) != 0) || + (datatype->flags() & DT_FLAG_DERIVED)) { // && (!smpi_is_shared(req->old_buf_))){ if (!smpi_process()->replaying()){ - if( smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP && (static_cast(req->old_buf_) >= smpi_start_data_exe) + if( smpi_privatize_global_variables != 0 && (static_cast(req->old_buf_) >= smpi_start_data_exe) && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){ XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment "); smpi_switch_data_segment(smpi_process()->index());