X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/4f76738343f7a677f81f5ab26a15234a3045ec8d..81b746a8bf1396a9e957ea63d654f0748a466529:/src/smpi/smpi_base.cpp diff --git a/src/smpi/smpi_base.cpp b/src/smpi/smpi_base.cpp index 3b4cdb9019..b7af99977a 100644 --- a/src/smpi/smpi_base.cpp +++ b/src/smpi/smpi_base.cpp @@ -5,7 +5,6 @@ * under the terms of the license (GNU LGPL) which comes with this package. */ #include -#include #include #include "private.h" @@ -17,16 +16,17 @@ #include "src/simix/smx_private.h" #include "surf/surf.h" #include "simgrid/sg_config.h" +#include "smpi/smpi_utils.hpp" #include "colls/colls.h" -#include "src/simix/SynchroComm.hpp" +#include "src/kernel/activity/SynchroComm.hpp" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)"); -static int match_recv(void* a, void* b, smx_synchro_t ignored) { - MPI_Request ref = static_cast(a); - MPI_Request req = static_cast(b); - XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag); +static int match_recv(void* a, void* b, smx_activity_t ignored) { + MPI_Request ref = static_cast(a); + MPI_Request req = static_cast(b); + XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag); xbt_assert(ref, "Cannot match recv against null reference"); xbt_assert(req, "Cannot match recv against null request"); @@ -34,52 +34,44 @@ static int match_recv(void* a, void* b, smx_synchro_t ignored) { && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){ //we match, we can transfer some values if(ref->src == MPI_ANY_SOURCE) - ref->real_src = req->src; + ref->real_src = req->src; if(ref->tag == MPI_ANY_TAG) - ref->real_tag = req->tag; + ref->real_tag = req->tag; if(ref->real_size < req->real_size) - ref->truncated = 1; + ref->truncated = 1; if(req->detached==1) - ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver + ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver XBT_DEBUG("match succeeded"); return 1; }else return 0; } -static int match_send(void* a, void* b,smx_synchro_t ignored) { - MPI_Request ref = static_cast(a); - MPI_Request req = static_cast(b); - XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag); - xbt_assert(ref, "Cannot match send against null reference"); - xbt_assert(req, "Cannot match send against null request"); - - if((req->src == MPI_ANY_SOURCE || req->src == ref->src) - && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag)) - { - if(req->src == MPI_ANY_SOURCE) - req->real_src = ref->src; - if(req->tag == MPI_ANY_TAG) - req->real_tag = ref->tag; - if(req->real_size < ref->real_size) - req->truncated = 1; - if(ref->detached==1) - req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver - XBT_DEBUG("match succeeded"); - return 1; - } else return 0; +static int match_send(void* a, void* b,smx_activity_t ignored) { + MPI_Request ref = static_cast(a); + MPI_Request req = static_cast(b); + XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag); + xbt_assert(ref, "Cannot match send against null reference"); + xbt_assert(req, "Cannot match send against null request"); + + if((req->src == MPI_ANY_SOURCE || req->src == ref->src) + && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag)){ + if(req->src == MPI_ANY_SOURCE) + req->real_src = ref->src; + if(req->tag == MPI_ANY_TAG) + req->real_tag = ref->tag; + if(req->real_size < ref->real_size) + req->truncated = 1; + if(ref->detached==1) + req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver + XBT_DEBUG("match succeeded"); + return 1; + } else + return 0; } -// 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 values for each factor -typedef struct s_smpi_factor_multival *smpi_os_factor_multival_t; -typedef struct s_smpi_factor_multival { // FIXME: this should be merged (deduplicated) with s_smpi_factor defined in network_smpi.c - long factor=0; - std::vector values; /** We allocate arbitrarily 4 elements **/ -} s_smpi_factor_multival_t; - -std::vector smpi_os_values; -std::vector smpi_or_values; -std::vector smpi_ois_values; +std::vector smpi_os_values; +std::vector smpi_or_values; +std::vector smpi_ois_values; static simgrid::config::Flag smpi_wtime_sleep( "smpi/wtime", "Minimum time to inject inside a call to MPI_Wtime", 0.0); @@ -90,79 +82,20 @@ 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); -static int factor_cmp(const s_smpi_factor_multival_t& pa, const s_smpi_factor_multival_t& pb) -{ - return (pa.factor > pb.factor) ? 1 : - (pa.factor < pb.factor) ? -1 : 0; -} - -static std::vector parse_factor(const char *smpi_coef_string) -{ - std::vector smpi_factor; - s_smpi_factor_multival_t fact; - - /** Setup the tokenizer that parses the string **/ - typedef boost::tokenizer> Tokenizer; - boost::char_separator sep(";"); - boost::char_separator factor_separator(":"); - std::string tmp_string(smpi_coef_string); - Tokenizer tokens(tmp_string, sep); - - /** - * Iterate over patterns like A:B:C:D;E:F;G:H - * These will be broken down into: - * A --> B, C, D - * E --> F - * G --> H - */ - for (Tokenizer::iterator token_iter = tokens.begin(); - token_iter != tokens.end(); token_iter++) { - Tokenizer factor_values(*token_iter, factor_separator); - - if (factor_values.begin() == factor_values.end()) { - xbt_die("Malformed radical for smpi factor: '%s'", smpi_coef_string); - } - unsigned int iteration = 0; - for (Tokenizer::iterator factor_iter = factor_values.begin(); - factor_iter != factor_values.end(); factor_iter++, iteration++) { - char *errmsg; - - if (factor_iter == factor_values.begin()) { /* first element */ - errmsg = bprintf("Invalid factor in chunk #%zu: %%s", smpi_factor.size()+1); - fact.factor = xbt_str_parse_int(factor_iter->c_str(), errmsg); - } - else { - errmsg = bprintf("Invalid factor value %d in chunk #%zu: %%s", iteration, smpi_factor.size()+1); - fact.values.push_back(xbt_str_parse_double((*factor_iter).c_str(), errmsg)); - } - xbt_free(errmsg); - } - - smpi_factor.push_back(fact); - XBT_DEBUG("smpi_factor:\t%ld : %zu values, first: %f", fact.factor, smpi_factor.size(), fact.values[0]); - } - std::sort(smpi_factor.begin(), smpi_factor.end(), &factor_cmp); - for (auto& fact : smpi_factor) { - XBT_DEBUG("smpi_factor:\t%ld : %zu values, first: %f", fact.factor, smpi_factor.size() ,fact.values[0]); - } - return smpi_factor; -} - -static double smpi_os(double size) +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=0.0; + 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 dynar we iterate over! + // 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 : %f <= %ld return %f", size, fact.factor, current); + 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. @@ -170,23 +103,23 @@ static double smpi_os(double size) current = fact.values[0]+fact.values[1]*size; } } - XBT_DEBUG("Searching for smpi/os: %f is larger than the largest boundary, return %f", size, current); + XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current); return current; } -static double smpi_ois(double size) +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=0.0; + 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 dynar we iterate over! + // 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 : %f <= %ld return %f", size, fact.factor, 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. @@ -194,24 +127,25 @@ static double smpi_ois(double size) current = fact.values[0]+fact.values[1]*size; } } - XBT_DEBUG("Searching for smpi/ois: %f is larger than the largest boundary, return %f", size, current); + XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current); return current; } -static double smpi_or(double size) +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=0.0; + + 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 dynar we iterate over! + // 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 : %f <= %ld return %f", size, fact.factor, current); + 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. @@ -219,7 +153,7 @@ static double smpi_or(double size) current=fact.values[0]+fact.values[1]*size; } } - XBT_DEBUG("smpi_or: %f is larger than largest boundary, return %f", size, current); + XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current); return current; } @@ -231,9 +165,7 @@ void smpi_mpi_init() { double smpi_mpi_wtime(){ double time; - if (smpi_process_initialized() != 0 && - smpi_process_finalized() == 0 && - smpi_process_get_sampling() == 0) { + if (smpi_process_initialized() != 0 && smpi_process_finalized() == 0 && smpi_process_get_sampling() == 0) { smpi_bench_end(); time = SIMIX_get_clock(); // to avoid deadlocks if used as a break condition, such as @@ -319,8 +251,7 @@ static void smpi_mpi_request_free_voidp(void* request) } /* MPI Low level calls */ -MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, - int dst, int tag, MPI_Comm comm) +MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(), @@ -328,8 +259,7 @@ MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, return request; } -MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype, - int dst, int tag, MPI_Comm comm) +MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(), @@ -337,8 +267,7 @@ MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype, return request; } -MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, - int src, int tag, MPI_Comm comm) +MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */ request = build_request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, @@ -373,8 +302,9 @@ void smpi_mpi_start(MPI_Request request) //begin with the more appropriate one : the small one. mailbox = smpi_process_mailbox_small(); XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox); - smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast(request)); - + smx_activity_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, + static_cast(request)); + if (action == nullptr) { mailbox = smpi_process_mailbox(); XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox); @@ -383,21 +313,18 @@ void smpi_mpi_start(MPI_Request request) XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox); mailbox = smpi_process_mailbox_small(); } - } - else { + } else { XBT_DEBUG("yes there was something for us in the large mailbox"); } - } - else { + } else { mailbox = smpi_process_mailbox_small(); XBT_DEBUG("Is there a corresponding send already posted the small mailbox?"); - smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); - + smx_activity_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); + if (action == nullptr) { XBT_DEBUG("No, nothing in the permanent receive mailbox"); mailbox = smpi_process_mailbox(); - } - else { + } else { XBT_DEBUG("yes there was something for us in the small mailbox"); } } @@ -407,23 +334,22 @@ void smpi_mpi_start(MPI_Request request) 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_comm_null_copy_buffer_callback, request, -1.0); - XBT_DEBUG("recv simcall posted"); + XBT_DEBUG("recv simcall posted"); if (async_small_thresh != 0 || (request->flags & RMA) != 0 ) xbt_mutex_release(mut); - } - else { /* the RECV flag was not set, so this is a send */ + } else { /* the RECV flag was not set, so this is a send */ int receiver = request->dst; int rank = request->src; if (TRACE_smpi_view_internals()) { - TRACE_smpi_send(rank, rank, receiver,request->size); + TRACE_smpi_send(rank, rank, receiver, request->tag, request->size); } print_request("New send", request); void* buf = request->buf; - if ( (request->flags & SSEND) == 0 - && ( (request->flags & RMA) != 0 || static_cast(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) { + if ((request->flags & SSEND) == 0 && ( (request->flags & RMA) != 0 + || static_cast(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) { void *oldbuf = nullptr; request->detached = 1; XBT_DEBUG("Send request %p is detached", request); @@ -452,9 +378,9 @@ void smpi_mpi_start(MPI_Request request) } if(sleeptime > 0.0){ - simcall_process_sleep(sleeptime); - XBT_DEBUG("sending size of %zu : sleep %f ", request->size, sleeptime); - } + simcall_process_sleep(sleeptime); + XBT_DEBUG("sending size of %zu : sleep %f ", request->size, sleeptime); + } int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh"); @@ -465,17 +391,16 @@ void smpi_mpi_start(MPI_Request request) if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) { mailbox = smpi_process_remote_mailbox(receiver); - } - else if (((request->flags & RMA) != 0) || static_cast(request->size) < async_small_thresh) { // eager mode + } else if (((request->flags & RMA) != 0) || static_cast(request->size) < async_small_thresh) { // eager mode mailbox = smpi_process_remote_mailbox(receiver); XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox); - smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast(request)); + smx_activity_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, + static_cast(request)); if (action == nullptr) { if ((request->flags & SSEND) == 0){ mailbox = smpi_process_remote_mailbox_small(receiver); XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox); - } - else { + } else { mailbox = smpi_process_remote_mailbox_small(receiver); XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox); action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast(request)); @@ -484,12 +409,10 @@ void smpi_mpi_start(MPI_Request request) mailbox = smpi_process_remote_mailbox(receiver); } } - } - else { + } else { XBT_DEBUG("Yes there was something for us in the large mailbox"); } - } - else { + } else { mailbox = smpi_process_remote_mailbox(receiver); XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf); } @@ -537,10 +460,10 @@ void smpi_mpi_request_free(MPI_Request * request) xbt_free(*request); *request = MPI_REQUEST_NULL; }else{ - print_request("Decrementing", (*request)); + print_request("Decrementing", (*request)); } }else{ - xbt_die("freeing an already free request"); + xbt_die("freeing an already free request"); } } @@ -722,16 +645,15 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) if (TRACE_smpi_view_internals() && ((req->flags & RECV) != 0)){ int rank = smpi_process_index(); int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src); - TRACE_smpi_recv(rank, src_traced, rank); + TRACE_smpi_recv(rank, src_traced, rank,req->tag); } if(req->detached_sender != nullptr){ - //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0 double sleeptime = smpi_or(req->real_size); if(sleeptime > 0.0){ - simcall_process_sleep(sleeptime); - XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size, sleeptime); + simcall_process_sleep(sleeptime); + XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size, sleeptime); } smpi_mpi_request_free(&(req->detached_sender)); } @@ -764,7 +686,7 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { nsleeps=1;//reset the number of sleeps we will do next time if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0) *request = MPI_REQUEST_NULL; - }else{ + } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){ nsleeps++; } } @@ -773,47 +695,44 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status) { - xbt_dynar_t comms; + std::vector comms; + comms.reserve(count); + int i; - int* map; int flag = 0; - int size = 0; *index = MPI_UNDEFINED; - comms = xbt_dynar_new(sizeof(smx_synchro_t), nullptr); - map = xbt_new(int, count); + + std::vector map; /** Maps all matching comms back to their location in requests **/ for(i = 0; i < count; i++) { if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) { - xbt_dynar_push(comms, &requests[i]->action); - map[size] = i; - size++; + comms.push_back(requests[i]->action); + map.push_back(i); } } - if(size > 0) { + if(!map.empty()) { //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) { - *index = map[i]; + i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches! + if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches) + *index = map[i]; finish_wait(&requests[*index], status); - if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT)) - requests[*index] = MPI_REQUEST_NULL; flag = 1; nsleeps = 1; - }else{ + if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT)) { + requests[*index] = MPI_REQUEST_NULL; + } + } else { nsleeps++; } - }else{ + } else { //all requests are null or inactive, return true - flag=1; + flag = 1; smpi_empty_status(status); } - xbt_free(map); - xbt_dynar_free(&comms); return flag; } @@ -823,8 +742,7 @@ int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[]) MPI_Status stat; MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat; int flag=1; - int i; - for(i=0; iflags & PREPARED)) { if (smpi_mpi_test(&requests[i], pstat)!=1){ flag=0; @@ -852,7 +770,6 @@ void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){ } void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){ - MPI_Request request = build_request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV); @@ -870,18 +787,20 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){ mailbox = smpi_process_mailbox_small(); XBT_DEBUG("Trying to probe the perm recv mailbox"); - request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, static_cast(request)); + request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, + static_cast(request)); } if (request->action == nullptr){ mailbox = smpi_process_mailbox(); XBT_DEBUG("trying to probe the other mailbox"); - request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast(request)); + request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, + static_cast(request)); } if (request->action != nullptr){ - simgrid::simix::Comm *sync_comm = static_cast(request->action); - MPI_Request req = static_cast(sync_comm->src_data); + simgrid::kernel::activity::Comm *sync_comm = static_cast(request->action); + MPI_Request req = static_cast(sync_comm->src_data); *flag = 1; if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED) == 0) { status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src); @@ -893,7 +812,8 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* } else { *flag = 0; - nsleeps++; + if (xbt_cfg_get_boolean("smpi/grow-injected-times")) + nsleeps++; } smpi_mpi_request_free(&request); @@ -914,7 +834,7 @@ void smpi_mpi_wait(MPI_Request * request, MPI_Status * status) finish_wait(request, status); if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0)) - *request = MPI_REQUEST_NULL; + *request = MPI_REQUEST_NULL; } int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status) @@ -927,7 +847,7 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status) if(count > 0) { // Wait for a request to complete - comms = xbt_dynar_new(sizeof(smx_synchro_t), nullptr); + comms = xbt_dynar_new(sizeof(smx_activity_t), nullptr); map = xbt_new(int, count); XBT_DEBUG("Wait for one of %d", count); for(i = 0; i < count; i++) { @@ -949,7 +869,7 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status) } } if(size > 0) { - i = simcall_comm_waitany(comms); + i = simcall_comm_waitany(comms, -1); // not MPI_UNDEFINED, as this is a simix return code if (i != -1) { @@ -1064,35 +984,35 @@ int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Sta void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) { - smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm); + smpi_coll_tuned_bcast_binomial_tree(buf, count, datatype, root, comm); } void smpi_mpi_barrier(MPI_Comm comm) { - smpi_coll_tuned_barrier_ompi_basic_linear(comm); + smpi_coll_tuned_barrier_ompi_basic_linear(comm); } void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) { int system_tag = COLL_TAG_GATHER; - int rank, size, src, index; - MPI_Aint lb = 0, recvext = 0; - MPI_Request *requests; + MPI_Aint lb = 0; + MPI_Aint recvext = 0; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); if(rank != root) { // Send buffer to root smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm); } else { smpi_datatype_extent(recvtype, &lb, &recvext); // Local copy from root - smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast(recvbuf) + root * recvcount * recvext, recvcount, recvtype); + smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast(recvbuf) + root * recvcount * recvext, + recvcount, recvtype); // Receive buffers from senders - requests = xbt_new(MPI_Request, size - 1); - index = 0; - for(src = 0; src < size; src++) { + MPI_Request *requests = xbt_new(MPI_Request, size - 1); + int index = 0; + for (int src = 0; src < size; src++) { if(src != root) { requests[index] = smpi_irecv_init(static_cast(recvbuf) + src * recvcount * recvext, recvcount, recvtype, src, system_tag, comm); @@ -1102,7 +1022,7 @@ void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, // Wait for completion of irecv's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); - for(src = 0; src < size-1; src++) { + for (int src = 0; src < size-1; src++) { smpi_mpi_request_free(&requests[src]); } xbt_free(requests); @@ -1112,38 +1032,34 @@ void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { - int i, size, count; - int *displs; - int rank = smpi_comm_rank(comm); - void *tmpbuf; - - /* arbitrarily choose root as rank 0 */ - size = smpi_comm_size(comm); - count = 0; - displs = xbt_new(int, size); - for (i = 0; i < size; i++) { - displs[i] = count; - count += recvcounts[i]; - } - tmpbuf=static_cast(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype))); + int rank = smpi_comm_rank(comm); - mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm); - smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm); - xbt_free(displs); - smpi_free_tmp_buffer(tmpbuf); + /* arbitrarily choose root as rank 0 */ + int size = smpi_comm_size(comm); + int count = 0; + int *displs = xbt_new(int, size); + for (int i = 0; i < size; i++) { + displs[i] = count; + count += recvcounts[i]; + } + void *tmpbuf = static_cast(smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype))); + + mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm); + smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm); + xbt_free(displs); + smpi_free_tmp_buffer(tmpbuf); } void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, int root, MPI_Comm comm) { int system_tag = COLL_TAG_GATHERV; - int rank, size, src, index; - MPI_Aint lb = 0, recvext = 0; - MPI_Request *requests; + MPI_Aint lb = 0; + MPI_Aint recvext = 0; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); - if(rank != root) { + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); + if (rank != root) { // Send buffer to root smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm); } else { @@ -1152,9 +1068,9 @@ void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast(recvbuf) + displs[root] * recvext, recvcounts[root], recvtype); // Receive buffers from senders - requests = xbt_new(MPI_Request, size - 1); - index = 0; - for(src = 0; src < size; src++) { + MPI_Request *requests = xbt_new(MPI_Request, size - 1); + int index = 0; + for (int src = 0; src < size; src++) { if(src != root) { requests[index] = smpi_irecv_init(static_cast(recvbuf) + displs[src] * recvext, recvcounts[src], recvtype, src, system_tag, comm); @@ -1164,7 +1080,7 @@ void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void // Wait for completion of irecv's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); - for(src = 0; src < size-1; src++) { + for (int src = 0; src < size-1; src++) { smpi_mpi_request_free(&requests[src]); } xbt_free(requests); @@ -1175,32 +1091,33 @@ void smpi_mpi_allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,int recvcount, MPI_Datatype recvtype, MPI_Comm comm) { int system_tag = COLL_TAG_ALLGATHER; - int rank, size, other, index; - MPI_Aint lb = 0, recvext = 0; + MPI_Aint lb = 0; + MPI_Aint recvext = 0; MPI_Request *requests; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); // FIXME: check for errors smpi_datatype_extent(recvtype, &lb, &recvext); // Local copy from self - smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast(recvbuf) + rank * recvcount * recvext, recvcount, recvtype); + smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast(recvbuf) + rank * recvcount * recvext, recvcount, + recvtype); // Send/Recv buffers to/from others; requests = xbt_new(MPI_Request, 2 * (size - 1)); - index = 0; - for(other = 0; other < size; other++) { + int index = 0; + for (int other = 0; other < size; other++) { if(other != rank) { requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,comm); index++; - requests[index] = smpi_irecv_init(static_cast(recvbuf) + other * recvcount * recvext, recvcount, recvtype, other, - system_tag, comm); + requests[index] = smpi_irecv_init(static_cast(recvbuf) + other * recvcount * recvext, recvcount, recvtype, + other, system_tag, comm); index++; } } // Wait for completion of all comms. smpi_mpi_startall(2 * (size - 1), requests); smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); - for(other = 0; other < 2*(size-1); other++) { + for (int other = 0; other < 2*(size-1); other++) { smpi_mpi_request_free(&requests[other]); } xbt_free(requests); @@ -1210,19 +1127,19 @@ void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, vo int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm) { int system_tag = COLL_TAG_ALLGATHERV; - int rank, size, other, index; - MPI_Aint lb = 0, recvext = 0; - MPI_Request *requests; + MPI_Aint lb = 0; + MPI_Aint recvext = 0; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); smpi_datatype_extent(recvtype, &lb, &recvext); // Local copy from self - smpi_datatype_copy(sendbuf, sendcount, sendtype, static_cast(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype); + smpi_datatype_copy(sendbuf, sendcount, sendtype, + static_cast(recvbuf) + displs[rank] * recvext,recvcounts[rank], recvtype); // Send buffers to others; - requests = xbt_new(MPI_Request, 2 * (size - 1)); - index = 0; - for(other = 0; other < size; other++) { + MPI_Request *requests = xbt_new(MPI_Request, 2 * (size - 1)); + int index = 0; + for (int other = 0; other < size; other++) { if(other != rank) { requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm); @@ -1235,7 +1152,7 @@ void smpi_mpi_allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, vo // Wait for completion of all comms. smpi_mpi_startall(2 * (size - 1), requests); smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); - for(other = 0; other < 2*(size-1); other++) { + for (int other = 0; other < 2*(size-1); other++) { smpi_mpi_request_free(&requests[other]); } xbt_free(requests); @@ -1245,12 +1162,12 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) { int system_tag = COLL_TAG_SCATTER; - int rank, size, dst, index; - MPI_Aint lb = 0, sendext = 0; + MPI_Aint lb = 0; + MPI_Aint sendext = 0; MPI_Request *requests; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); if(rank != root) { // Recv buffer from root smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE); @@ -1263,18 +1180,18 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, } // Send buffers to receivers requests = xbt_new(MPI_Request, size - 1); - index = 0; - for(dst = 0; dst < size; dst++) { + int index = 0; + for(int dst = 0; dst < size; dst++) { if(dst != root) { - requests[index] = smpi_isend_init(static_cast(sendbuf) + dst * sendcount * sendext, sendcount, sendtype, dst, - system_tag, comm); + requests[index] = smpi_isend_init(static_cast(sendbuf) + dst * sendcount * sendext, sendcount, sendtype, + dst, system_tag, comm); index++; } } // Wait for completion of isend's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); - for(dst = 0; dst < size-1; dst++) { + for (int dst = 0; dst < size-1; dst++) { smpi_mpi_request_free(&requests[dst]); } xbt_free(requests); @@ -1285,12 +1202,11 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype MPI_Datatype recvtype, int root, MPI_Comm comm) { int system_tag = COLL_TAG_SCATTERV; - int rank, size, dst, index; - MPI_Aint lb = 0, sendext = 0; - MPI_Request *requests; + MPI_Aint lb = 0; + MPI_Aint sendext = 0; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); if(rank != root) { // Recv buffer from root smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE); @@ -1302,10 +1218,10 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, recvbuf, recvcount, recvtype); } // Send buffers to receivers - requests = xbt_new(MPI_Request, size - 1); - index = 0; - for(dst = 0; dst < size; dst++) { - if(dst != root) { + MPI_Request *requests = xbt_new(MPI_Request, size - 1); + int index = 0; + for (int dst = 0; dst < size; dst++) { + if (dst != root) { requests[index] = smpi_isend_init(static_cast(sendbuf) + displs[dst] * sendext, sendcounts[dst], sendtype, dst, system_tag, comm); index++; @@ -1314,7 +1230,7 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, MPI_Datatype // Wait for completion of isend's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); - for(dst = 0; dst < size-1; dst++) { + for (int dst = 0; dst < size-1; dst++) { smpi_mpi_request_free(&requests[dst]); } xbt_free(requests); @@ -1325,16 +1241,13 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datat MPI_Comm comm) { int system_tag = COLL_TAG_REDUCE; - int rank, size, src, index; - MPI_Aint lb = 0, dataext = 0; - MPI_Request *requests; - void **tmpbufs; + MPI_Aint lb = 0; + MPI_Aint dataext = 0; char* sendtmpbuf = static_cast(sendbuf); - - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); //non commutative case, use a working algo from openmpi if(!smpi_op_is_commute(op)){ smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, datatype, op, root, comm); @@ -1355,11 +1268,11 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datat if (sendtmpbuf != nullptr && recvbuf != nullptr) smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype); // Receive buffers from senders - requests = xbt_new(MPI_Request, size - 1); - tmpbufs = xbt_new(void *, size - 1); - index = 0; - for(src = 0; src < size; src++) { - if(src != root) { + MPI_Request *requests = xbt_new(MPI_Request, size - 1); + void **tmpbufs = xbt_new(void *, size - 1); + int index = 0; + for (int src = 0; src < size; src++) { + if (src != root) { if (!smpi_process_get_replaying()) tmpbufs[index] = xbt_malloc(count * dataext); else @@ -1371,7 +1284,7 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datat } // Wait for completion of irecv's. smpi_mpi_startall(size - 1, requests); - for(src = 0; src < size - 1; src++) { + for (int src = 0; src < size - 1; src++) { index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); XBT_DEBUG("finished waiting any request with index %d", index); if(index == MPI_UNDEFINED) { @@ -1403,13 +1316,10 @@ void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype da void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { int system_tag = -888; - int rank, size, other, index; MPI_Aint lb = 0, dataext = 0; - MPI_Request *requests; - void **tmpbufs; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); smpi_datatype_extent(datatype, &lb, &dataext); @@ -1417,15 +1327,15 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatyp smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype); // Send/Recv buffers to/from others; - requests = xbt_new(MPI_Request, size - 1); - tmpbufs = xbt_new(void *, rank); - index = 0; - for(other = 0; other < rank; other++) { + MPI_Request *requests = xbt_new(MPI_Request, size - 1); + void **tmpbufs = xbt_new(void *, rank); + int index = 0; + for (int other = 0; other < rank; other++) { tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext); requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm); index++; } - for(other = rank + 1; other < size; other++) { + for (int other = rank + 1; other < size; other++) { requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm); index++; } @@ -1433,7 +1343,7 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatyp smpi_mpi_startall(size - 1, requests); if(smpi_op_is_commute(op)){ - for(other = 0; other < size - 1; other++) { + for (int other = 0; other < size - 1; other++) { index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); if(index == MPI_UNDEFINED) { break; @@ -1445,7 +1355,7 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatyp } }else{ //non commutative case, wait in order - for(other = 0; other < size - 1; other++) { + for (int other = 0; other < size - 1; other++) { smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE); if(index < rank) { smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype); @@ -1465,35 +1375,30 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatyp void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { int system_tag = -888; - int rank, size, other, index; MPI_Aint lb = 0, dataext = 0; - MPI_Request *requests; - void **tmpbufs; int recvbuf_is_empty=1; - rank = smpi_comm_rank(comm); - size = smpi_comm_size(comm); + int rank = smpi_comm_rank(comm); + int size = smpi_comm_size(comm); smpi_datatype_extent(datatype, &lb, &dataext); // Send/Recv buffers to/from others; - requests = xbt_new(MPI_Request, size - 1); - tmpbufs = xbt_new(void *, rank); - index = 0; - for(other = 0; other < rank; other++) { + MPI_Request *requests = xbt_new(MPI_Request, size - 1); + void **tmpbufs = xbt_new(void *, rank); + int index = 0; + for (int other = 0; other < rank; other++) { tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext); - requests[index] = - smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm); + requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm); index++; } - for(other = rank + 1; other < size; other++) { - requests[index] = - smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm); + for (int other = rank + 1; other < size; other++) { + requests[index] = smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm); index++; } // Wait for completion of all comms. smpi_mpi_startall(size - 1, requests); if(smpi_op_is_commute(op)){ - for(other = 0; other < size - 1; other++) { + for (int other = 0; other < size - 1; other++) { index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); if(index == MPI_UNDEFINED) { break; @@ -1502,20 +1407,21 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datat if(recvbuf_is_empty){ smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype); recvbuf_is_empty=0; - }else - // #Request is below rank: it's a irecv - smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + } else + // #Request is below rank: it's a irecv + smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); } } }else{ //non commutative case, wait in order - for(other = 0; other < size - 1; other++) { + 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); + } else + smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype); } } }