teshsuite/smpi/coll-reduce-scatter/coll-reduce-scatter
teshsuite/smpi/coll-scatter/coll-scatter
teshsuite/smpi/macro-shared/macro-shared
+teshsuite/smpi/macro-partial-shared/macro-partial-shared
+teshsuite/smpi/macro-partial-shared-communication/macro-partial-shared-communication
teshsuite/smpi/type-struct/type-struct
teshsuite/smpi/type-vector/type-vector
teshsuite/s4u/actor/actor
#define SMPI_SAMPLE_DELAY(duration) for(smpi_execute(duration); 0; )
#define SMPI_SAMPLE_FLOPS(flops) for(smpi_execute_flops(flops); 0; )
-XBT_PUBLIC(int) smpi_is_shared(void *buf);
XBT_PUBLIC(void *) smpi_shared_malloc(size_t size, const char *file, int line);
#define SMPI_SHARED_MALLOC(size) smpi_shared_malloc(size, __FILE__, __LINE__)
+XBT_PUBLIC(void *) smpi_shared_malloc_global__(size_t size, const char *file, int line, size_t *shared_block_offsets, int nb_shared_blocks);
+#define SMPI_PARTIAL_SHARED_MALLOC(size, shared_block_offsets, nb_shared_blocks)\
+ smpi_shared_malloc_global__(size, __FILE__, __LINE__, shared_block_offsets, nb_shared_blocks)
XBT_PUBLIC(void) smpi_shared_free(void *data);
#define SMPI_SHARED_FREE(data) smpi_shared_free(data)
--- /dev/null
+#ifndef SMPI_SHARED_HPP
+#define SMPI_SHARED_HPP
+#include <xbt/function_types.h>
+#include <xbt/misc.h>
+#include <vector>
+
+
+/*
+ * We cannot put this declaration in smpi.h, since we use C++ features.
+ */
+
+
+XBT_PUBLIC(int) smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset);
+
+std::vector<std::pair<size_t, size_t>> shift_and_frame_private_blocks(const std::vector<std::pair<size_t, size_t>> vec, size_t offset, size_t buff_size);
+std::vector<std::pair<size_t, size_t>> merge_private_blocks(std::vector<std::pair<size_t, size_t>> src, std::vector<std::pair<size_t, size_t>> dst);
+
+#endif
void *recvbuf, int recvcount, MPI_Datatype recvtype){
int count;
+// FIXME Handle the case of a partial shared malloc.
+#if 0
if(smpi_is_shared(sendbuf)){
XBT_DEBUG("Copy input buf %p is shared. Let's ignore it.", sendbuf);
}else if(smpi_is_shared(recvbuf)){
XBT_DEBUG("Copy output buf %p is shared. Let's ignore it.", recvbuf);
}
+#endif
if(smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP){
smpi_switch_data_segment(smpi_process()->index());
#include "private.h"
#include "private.hpp"
#include "simgrid/s4u/Mailbox.hpp"
+#include "smpi/smpi_shared_malloc.hpp"
#include "simgrid/sg_config.h"
#include "src/kernel/activity/SynchroComm.hpp"
#include "src/mc/mc_record.h"
smpi_comm_copy_data_callback = callback;
}
+static void print(std::vector<std::pair<size_t, size_t>> vec) {
+ fprintf(stderr, "{");
+ for(auto elt: vec) {
+ fprintf(stderr, "(0x%lx, 0x%lx),", elt.first, elt.second);
+ }
+ fprintf(stderr, "}\n");
+}
+static void memcpy_private(void *dest, const void *src, size_t n, std::vector<std::pair<size_t, size_t>> &private_blocks) {
+ for(auto block : private_blocks) {
+ memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
+ }
+}
+
+static void check_blocks(std::vector<std::pair<size_t, size_t>> &private_blocks, size_t buff_size) {
+ for(auto block : private_blocks) {
+ xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc.");
+ }
+}
+
void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
{
-
simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
-
+ int src_shared=0, dst_shared=0;
+ size_t src_offset=0, dst_offset=0;
+ std::vector<std::pair<size_t, size_t>> src_private_blocks;
+ std::vector<std::pair<size_t, size_t>> dst_private_blocks;
XBT_DEBUG("Copy the data over");
- if(smpi_is_shared(buff)){
+ if((src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset))) {
XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
- }else if(smpi_is_shared((char*)comm->dst_buff)){
+ src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size);
+ }
+ else {
+ src_private_blocks.clear();
+ src_private_blocks.push_back(std::make_pair(0, buff_size));
+ }
+ if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) {
XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
- }else{
- void* tmpbuff=buff;
- if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_start_data_exe)
- && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
- ){
- XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
-
- smpi_switch_data_segment(
- (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
- tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
- memcpy(tmpbuff, buff, buff_size);
- }
-
- if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
- && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
- XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
- smpi_switch_data_segment(
- (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
- }
-
- XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
- memcpy(comm->dst_buff, tmpbuff, buff_size);
+ dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size);
+ }
+ else {
+ dst_private_blocks.clear();
+ dst_private_blocks.push_back(std::make_pair(0, buff_size));
+ }
+ check_blocks(src_private_blocks, buff_size);
+ check_blocks(dst_private_blocks, buff_size);
+ auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
+ check_blocks(private_blocks, buff_size);
+ void* tmpbuff=buff;
+ if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast<char*>(buff) >= smpi_start_data_exe)
+ && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )
+ ){
+ XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !");
+
+ smpi_switch_data_segment(
+ (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
+ tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
+ memcpy_private(tmpbuff, buff, buff_size, private_blocks);
+ }
- if (comm->detached) {
- // if this is a detached send, the source buffer was duplicated by SMPI
- // sender to make the original buffer available to the application ASAP
- xbt_free(buff);
- //It seems that the request is used after the call there this should be free somewhere else but where???
- //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
- comm->src_buff = nullptr;
- }
- if(tmpbuff!=buff)xbt_free(tmpbuff);
+ if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe)
+ && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){
+ XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment");
+ smpi_switch_data_segment(
+ (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->dst_proc->data)->data))->index()));
+ }
+ XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
+ memcpy_private(comm->dst_buff, tmpbuff, buff_size, private_blocks);
+
+ if (comm->detached) {
+ // if this is a detached send, the source buffer was duplicated by SMPI
+ // sender to make the original buffer available to the application ASAP
+ xbt_free(buff);
+ //It seems that the request is used after the call there this should be free somewhere else but where???
+ //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
+ comm->src_buff = nullptr;
}
+ if(tmpbuff!=buff)xbt_free(tmpbuff);
}
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))) { // && (!smpi_is_shared(buf_))){
// This part handles the problem of non-contiguous memory
old_buf = buf;
if (count==0){
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 != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
#include "private.h"
#include "private.hpp"
+#include "smpi/smpi_shared_malloc.hpp"
#include "xbt/dict.h"
#include <errno.h>
typedef struct {
size_t size;
+ std::vector<std::pair<size_t, size_t>> private_blocks;
shared_data_key_type* data;
} shared_metadata_t;
return mem;
}
-void *smpi_shared_malloc(size_t size, const char *file, int line)
+static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
{
void* mem;
- if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) {
- smpi_source_location loc(file, line);
- auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
- auto data = res.first;
- if (res.second) {
- // The insertion did not take place.
- // Generate a shared memory name from the address of the shared_data:
- char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise)
- snprintf(shmname, 31, "/shmalloc%p", &*data);
- int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
- if (fd < 0) {
- if (errno == EEXIST)
- xbt_die("Please cleanup /dev/shm/%s", shmname);
- else
- xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno));
- }
- data->second.fd = fd;
- data->second.count = 1;
- mem = shm_map(fd, size, &*data);
- if (shm_unlink(shmname) < 0) {
- XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
- }
- XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd);
- } else {
- mem = shm_map(data->second.fd, size, &*data);
- data->second.count++;
+ smpi_source_location loc(file, line);
+ auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
+ auto data = res.first;
+ if (res.second) {
+ // The insertion did not take place.
+ // Generate a shared memory name from the address of the shared_data:
+ char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise)
+ snprintf(shmname, 31, "/shmalloc%p", &*data);
+ int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
+ if (fd < 0) {
+ if (errno == EEXIST)
+ xbt_die("Please cleanup /dev/shm/%s", shmname);
+ else
+ xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno));
}
- XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data);
-
- } else if (smpi_cfg_shared_malloc == shmalloc_global) {
- /* First reserve memory area */
- mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
-
- xbt_assert(mem != MAP_FAILED, "Failed to allocate %luMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
- "to allow big allocations.\n",
- (unsigned long)(size >> 20));
-
- /* Create bogus file if not done already */
- if (smpi_shared_malloc_bogusfile == -1) {
- /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
- * It still exists in memory but not in the file system (thus it cannot be leaked). */
- smpi_shared_malloc_blocksize = static_cast<unsigned long>(xbt_cfg_get_double("smpi/shared-malloc-blocksize"));
- XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
- char* name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX");
- smpi_shared_malloc_bogusfile = mkstemp(name);
- unlink(name);
- xbt_free(name);
- char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize);
- ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize);
- if(err<0)
- xbt_die("Could not write bogus file for shared malloc");
- xbt_free(dumb);
+ data->second.fd = fd;
+ data->second.count = 1;
+ mem = shm_map(fd, size, &*data);
+ if (shm_unlink(shmname) < 0) {
+ XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
}
+ XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd);
+ } else {
+ mem = shm_map(data->second.fd, size, &*data);
+ data->second.count++;
+ }
+ XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data);
+ return mem;
+}
- /* Map the bogus file in place of the anonymous memory */
+// Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
+#define PAGE_SIZE 0x1000
+#define ALIGN_UP(n, align) (((n) + (align)-1) & -(align))
+#define ALIGN_DOWN(n, align) ((n) & -(align))
+
+void *smpi_shared_malloc_global__(size_t size, const char *file, int line, size_t *shared_block_offsets, int nb_shared_blocks) {
+ void *mem;
+ xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "The block size of shared malloc should be a multiple of the page size.");
+ /* First reserve memory area */
+ mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+
+ xbt_assert(mem != MAP_FAILED, "Failed to allocate %luMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
+ "to allow big allocations.\n",
+ (unsigned long)(size >> 20));
+
+ /* Create bogus file if not done already */
+ if (smpi_shared_malloc_bogusfile == -1) {
+ /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
+ * It still exists in memory but not in the file system (thus it cannot be leaked). */
+ smpi_shared_malloc_blocksize = static_cast<unsigned long>(xbt_cfg_get_double("smpi/shared-malloc-blocksize"));
+ XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
+ char* name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX");
+ smpi_shared_malloc_bogusfile = mkstemp(name);
+ unlink(name);
+ xbt_free(name);
+ char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize);
+ ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize);
+ if(err<0)
+ xbt_die("Could not write bogus file for shared malloc");
+ xbt_free(dumb);
+ }
+
+ /* Map the bogus file in place of the anonymous memory */
+ for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
+ size_t start_offset = shared_block_offsets[2*i_block];
+ size_t stop_offset = shared_block_offsets[2*i_block+1];
+ xbt_assert(start_offset < stop_offset, "start_offset (%lu) should be lower than stop offset (%lu)", start_offset, stop_offset);
+ xbt_assert(stop_offset <= size, "stop_offset (%lu) should be lower than size (%lu)", stop_offset, size);
+ if(i_block < nb_shared_blocks-1)
+ xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
+ "stop_offset (%lu) should be lower than its successor start offset (%lu)", stop_offset, shared_block_offsets[2*i_block+2]);
+ size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
+ size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
unsigned int i;
- for (i = 0; i < size / smpi_shared_malloc_blocksize; i++) {
+ for (i = start_block_offset / smpi_shared_malloc_blocksize; i < stop_block_offset / smpi_shared_malloc_blocksize; i++) {
void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize);
void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED | MAP_POPULATE,
smpi_shared_malloc_bogusfile, 0);
"You can also try using the sysctl vm.max_map_count",
strerror(errno));
}
- if (size % smpi_shared_malloc_blocksize) {
- void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize);
- void* res = mmap(pos, size % smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE,
- MAP_FIXED | MAP_SHARED | MAP_POPULATE, smpi_shared_malloc_bogusfile, 0);
+ size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE);
+ size_t low_page_stop_offset = start_block_offset < ALIGN_DOWN(stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN(stop_offset, PAGE_SIZE);
+ if(low_page_start_offset < low_page_stop_offset) {
+ void* pos = (void*)((unsigned long)mem + low_page_start_offset);
+ void* res = mmap(pos, low_page_stop_offset-low_page_start_offset, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED | MAP_POPULATE,
+ smpi_shared_malloc_bogusfile, 0);
xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
"size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
"You can also try using the sysctl vm.max_map_count",
strerror(errno));
}
+ if(low_page_stop_offset <= stop_block_offset) {
+ size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
+ if(high_page_stop_offset > stop_block_offset) {
+ void* pos = (void*)((unsigned long)mem + stop_block_offset);
+ void* res = mmap(pos, high_page_stop_offset-stop_block_offset, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED | MAP_POPULATE,
+ smpi_shared_malloc_bogusfile, 0);
+ xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
+ "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
+ "You can also try using the sysctl vm.max_map_count",
+ strerror(errno));
+ }
+ }
+ }
- shared_metadata_t newmeta;
- //register metadata for memcpy avoidance
- shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
- data->second.fd = -1;
- data->second.count = 1;
- newmeta.size = size;
- newmeta.data = data;
- allocs_metadata[mem] = newmeta;
+ shared_metadata_t newmeta;
+ //register metadata for memcpy avoidance
+ shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
+ data->second.fd = -1;
+ data->second.count = 1;
+ newmeta.size = size;
+ newmeta.data = data;
+ if(shared_block_offsets[0] > 0) {
+ newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0]));
+ }
+ int i_block;
+ for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
+ newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2]));
+ }
+ if(shared_block_offsets[2*i_block+1] < size) {
+ newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], size));
+ }
+ allocs_metadata[mem] = newmeta;
+
+ return mem;
+}
+
+/*
+ * When nb_shared_blocks == -1, default behavior of smpi_shared_malloc: everything is shared.
+ * Otherwise, only the blocks described by shared_block_offsets are shared.
+ * This array contains the offsets (in bytes) of the block to share.
+ * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
+ * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared. The others are not.
+ */
+static void *smpi_shared_malloc_global(size_t size, const char *file, int line, size_t *shared_block_offsets=NULL, int nb_shared_blocks=-1) {
+ size_t tmp_shared_block_offsets[2];
+ if(nb_shared_blocks == -1) {
+ nb_shared_blocks = 1;
+ shared_block_offsets = tmp_shared_block_offsets;
+ shared_block_offsets[0] = 0;
+ shared_block_offsets[1] = size;
+ }
+ return smpi_shared_malloc_global__(size, file, line, shared_block_offsets, nb_shared_blocks);
+}
+
+void *smpi_shared_malloc(size_t size, const char *file, int line) {
+ void *mem;
+ if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) {
+ mem = smpi_shared_malloc_local(size, file, line);
+ } else if (smpi_cfg_shared_malloc == shmalloc_global) {
+ mem = smpi_shared_malloc_global(size, file, line);
} else {
mem = xbt_malloc(size);
XBT_DEBUG("Classic malloc %zu in %p", size, mem);
}
-
return mem;
}
-int smpi_is_shared(void* ptr){
+int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
+ private_blocks.clear(); // being paranoid
if (allocs_metadata.empty())
return 0;
if ( smpi_cfg_shared_malloc == shmalloc_local || smpi_cfg_shared_malloc == shmalloc_global) {
auto low = allocs_metadata.lower_bound(ptr);
- if (low->first==ptr)
+ if (low->first==ptr) {
+ private_blocks = low->second.private_blocks;
+ *offset = 0;
return 1;
+ }
if (low == allocs_metadata.begin())
return 0;
low --;
- if (ptr < (char*)low->first + low->second.size)
+ if (ptr < (char*)low->first + low->second.size) {
+ xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
+ *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
+ private_blocks = low->second.private_blocks;
return 1;
+ }
return 0;
} else {
return 0;
}
}
+std::vector<std::pair<size_t, size_t>> shift_and_frame_private_blocks(const std::vector<std::pair<size_t, size_t>> vec, size_t offset, size_t buff_size) {
+ std::vector<std::pair<size_t, size_t>> result;
+ for(auto block: vec) {
+ auto new_block = std::make_pair(std::min(std::max((size_t)0, block.first-offset), buff_size),
+ std::min(std::max((size_t)0, block.second-offset), buff_size));
+ if(new_block.second > 0 && new_block.first < buff_size)
+ result.push_back(new_block);
+ }
+ return result;
+}
+
+std::vector<std::pair<size_t, size_t>> merge_private_blocks(std::vector<std::pair<size_t, size_t>> src, std::vector<std::pair<size_t, size_t>> dst) {
+ std::vector<std::pair<size_t, size_t>> result;
+ unsigned i_src=0, i_dst=0;
+ while(i_src < src.size() && i_dst < dst.size()) {
+ std::pair<size_t, size_t> block;
+ if(src[i_src].second <= dst[i_dst].first) {
+ i_src++;
+ }
+ else if(dst[i_dst].second <= src[i_src].first) {
+ i_dst++;
+ }
+ else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
+ block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
+ std::min(src[i_src].second, dst[i_dst].second));
+ result.push_back(block);
+ if(src[i_src].second < dst[i_dst].second)
+ i_src ++;
+ else
+ i_dst ++;
+ }
+ }
+ return result;
+}
+
void smpi_shared_free(void *ptr)
{
if (smpi_cfg_shared_malloc == shmalloc_local) {
xbt_free(meta->second.data);
}
- munmap(ptr, 0); // the POSIX says that I should not give 0 as a length, but it seems to work OK
+ munmap(ptr, meta->second.size);
} else {
XBT_DEBUG("Classic free of %p", ptr);
xbt_free(ptr);
add_executable (macro-shared macro-shared/macro-shared.c)
target_link_libraries(macro-shared simgrid)
set_target_properties(macro-shared PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/macro-shared)
+
+ add_executable (macro-partial-shared macro-partial-shared/macro-partial-shared.c)
+ target_link_libraries(macro-partial-shared simgrid)
+ set_target_properties(macro-partial-shared PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/macro-partial-shared)
+
+ add_executable (macro-partial-shared-communication macro-partial-shared-communication/macro-partial-shared-communication.c)
+ target_link_libraries(macro-partial-shared-communication simgrid)
+ set_target_properties(macro-partial-shared-communication PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/macro-partial-shared-communication)
endif()
endif()
set(teshsuite_src ${teshsuite_src} ${CMAKE_CURRENT_SOURCE_DIR}/macro-shared/macro-shared.c PARENT_SCOPE)
+set(teshsuite_src ${teshsuite_src} ${CMAKE_CURRENT_SOURCE_DIR}/macro-partial-shared/macro-partial-shared.c PARENT_SCOPE)
+set(teshsuite_src ${teshsuite_src} ${CMAKE_CURRENT_SOURCE_DIR}/macro-partial-shared-communication/macro-partial-shared-communication.c PARENT_SCOPE)
set(tesh_files ${tesh_files} ${CMAKE_CURRENT_SOURCE_DIR}/coll-allreduce/coll-allreduce-large.tesh
${CMAKE_CURRENT_SOURCE_DIR}/coll-allreduce/coll-allreduce-automatic.tesh
${CMAKE_CURRENT_SOURCE_DIR}/coll-alltoall/clusters.tesh
${CMAKE_CURRENT_SOURCE_DIR}/macro-shared/macro-shared.tesh
+ ${CMAKE_CURRENT_SOURCE_DIR}/macro-partial-shared/macro-partial-shared.tesh
+ ${CMAKE_CURRENT_SOURCE_DIR}/macro-partial-shared-communication/macro-partial-shared-communication.tesh
${CMAKE_CURRENT_SOURCE_DIR}/pt2pt-pingpong/broken_hostfiles.tesh
${CMAKE_CURRENT_SOURCE_DIR}/pt2pt-pingpong/TI_output.tesh PARENT_SCOPE)
set(bin_files ${bin_files} ${CMAKE_CURRENT_SOURCE_DIR}/hostfile
if(enable_smpi)
if(NOT WIN32)
ADD_TESH_FACTORIES(tesh-smpi-macro-shared "thread;ucontext;raw;boost" --setenv bindir=${CMAKE_BINARY_DIR}/teshsuite/smpi/macro-shared --cd ${CMAKE_HOME_DIRECTORY}/teshsuite/smpi/macro-shared macro-shared.tesh)
+ ADD_TESH_FACTORIES(tesh-smpi-macro-partial-shared "thread;ucontext;raw;boost" --setenv bindir=${CMAKE_BINARY_DIR}/teshsuite/smpi/macro-partial-shared --cd ${CMAKE_HOME_DIRECTORY}/teshsuite/smpi/macro-partial-shared macro-partial-shared.tesh)
+ ADD_TESH_FACTORIES(tesh-smpi-macro-partial-shared-communication "thread;ucontext;raw;boost" --setenv bindir=${CMAKE_BINARY_DIR}/teshsuite/smpi/macro-partial-shared-communication --cd ${CMAKE_HOME_DIRECTORY}/teshsuite/smpi/macro-partial-shared-communication macro-partial-shared-communication.tesh)
endif()
foreach(x coll-allgather coll-allgatherv coll-allreduce coll-alltoall coll-alltoallv coll-barrier coll-bcast
--- /dev/null
+/* Copyright (c) 2009-2015. The SimGrid Team.
+ * All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
+#include <stdio.h>
+#include <mpi.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <assert.h>
+
+// Set the elements between buf[start] and buf[stop-1] to (i+value)%256
+static void set(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ for(size_t i = start; i < stop; i++) {
+ buf[i] = (i+value)%256;
+ }
+}
+
+// Return the number of times that an element is equal to (i+value)%256 between buf[start] and buf[stop-1].
+static int count_all(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ size_t occ = 0;
+ for(size_t i = start ; i < stop ; i++) {
+ if(buf[i] == (i+value)%256) {
+ occ ++;
+ }
+ }
+ return occ;
+}
+
+// Return true iff the values from buf[start] to buf[stop-1] are all equal to (i+value)%256.
+static int check_all(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ size_t occ = count_all(buf, start, stop, value);
+ return occ == stop-start;
+}
+
+// Return true iff "enough" elements are equal to (i+value)%256 between buf[start] and buf[stop-1].
+static int check_enough(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ int page_size = 0x1000;
+ size_t size = stop-start;
+ if(size <= 2*page_size) // we are not sure to have a whole page that is shared
+ return 1;
+ size_t occ = count_all(buf, start, stop, value);
+ return occ >= size - 2*page_size;
+}
+
+int main(int argc, char *argv[])
+{
+ MPI_Init(&argc, &argv);
+ int rank;
+ int size;
+ size_t mem_size = 0x10000000;
+ size_t shared_blocks[] = {
+ 0, 0x1234567,
+ 0x1300000, 0x1300010,
+ 0x3456789, 0x3457890,
+ 0x4444444, 0x5555555,
+ 0x5555565, 0x5600000,
+ 0x8000000, 0x10000000
+ };
+ int nb_blocks = (sizeof(shared_blocks)/sizeof(size_t))/2;
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+ //Let's Allocate a shared memory buffer
+ assert(size%2 == 0);
+ uint8_t *buf;
+ buf = SMPI_PARTIAL_SHARED_MALLOC(mem_size, shared_blocks, nb_blocks);
+ memset(buf, rank, mem_size);
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ // Even processes write their rank in private blocks
+ if(rank%2 == 0) {
+ for(int i = 0; i < nb_blocks-1; i++) {
+ size_t start = shared_blocks[2*i+1];
+ size_t stop = shared_blocks[2*i+2];
+ set(buf, start, stop, rank);
+ }
+ }
+ // Then, even processes send their buffer to their successor
+ if(rank%2 == 0) {
+ MPI_Send(buf, mem_size, MPI_UINT8_T, rank+1, 0, MPI_COMM_WORLD);
+ }
+ else {
+ MPI_Recv(buf, mem_size, MPI_UINT8_T, rank-1, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+ }
+
+
+ // Odd processes verify that they successfully received the message
+ if(rank%2 == 1) {
+ for(int i = 0; i < nb_blocks-1; i++) {
+ size_t start = shared_blocks[2*i+1];
+ size_t stop = shared_blocks[2*i+2];
+ int comm = check_all(buf, start, stop, rank-1);
+ printf("[%d] The result of the (normal) communication check for block (0x%lx, 0x%lx) is: %d\n", rank, start, stop, comm);
+ }
+ memset(buf, rank, mem_size);
+ }
+
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ // Then, even processes send a sub-part of their buffer their successor
+ // Note that the last block should not be copied entirely
+ if(rank%2 == 0) {
+ MPI_Send(buf+0x10000, mem_size-0xa000000, MPI_UINT8_T, rank+1, 0, MPI_COMM_WORLD);
+ }
+ else {
+ MPI_Recv(buf+0x10000, mem_size-0xa000000, MPI_UINT8_T, rank-1, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+ }
+
+
+ // Odd processes verify that they successfully received the message
+ if(rank%2 == 1) {
+ for(int i = 0; i < nb_blocks-1; i++) {
+ size_t start = shared_blocks[2*i+1];
+ size_t stop = shared_blocks[2*i+2];
+ int comm = check_all(buf, start, stop, rank-1);
+ printf("[%d] The result of the (shifted) communication check for block (0x%lx, 0x%lx) is: %d\n", rank, start, stop, comm);
+ }
+ }
+
+ SMPI_SHARED_FREE(buf);
+
+ MPI_Finalize();
+ return 0;
+}
--- /dev/null
+p Test compute
+! setenv LD_LIBRARY_PATH=../../lib
+! output sort
+! timeout 5
+$ ${bindir:=.}/../../../smpi_script/bin/smpirun -hostfile ../hostfile -platform ../../../examples/platforms/small_platform.xml -np 4 ${bindir:=.}/macro-partial-shared-communication --log=smpi_kernel.thres:warning --log=xbt_cfg.thres:warning
+> [3] The result of the (normal) communication check for block (0x1234567, 0x1300000) is: 1
+> [3] The result of the (normal) communication check for block (0x1300010, 0x3456789) is: 1
+> [3] The result of the (normal) communication check for block (0x3457890, 0x4444444) is: 1
+> [3] The result of the (normal) communication check for block (0x5555555, 0x5555565) is: 1
+> [3] The result of the (normal) communication check for block (0x5600000, 0x8000000) is: 1
+> [1] The result of the (normal) communication check for block (0x1234567, 0x1300000) is: 1
+> [1] The result of the (normal) communication check for block (0x1300010, 0x3456789) is: 1
+> [1] The result of the (normal) communication check for block (0x3457890, 0x4444444) is: 1
+> [1] The result of the (normal) communication check for block (0x5555555, 0x5555565) is: 1
+> [1] The result of the (normal) communication check for block (0x5600000, 0x8000000) is: 1
+> [3] The result of the (shifted) communication check for block (0x1234567, 0x1300000) is: 1
+> [3] The result of the (shifted) communication check for block (0x1300010, 0x3456789) is: 1
+> [3] The result of the (shifted) communication check for block (0x3457890, 0x4444444) is: 1
+> [3] The result of the (shifted) communication check for block (0x5555555, 0x5555565) is: 1
+> [3] The result of the (shifted) communication check for block (0x5600000, 0x8000000) is: 0
+> [1] The result of the (shifted) communication check for block (0x1234567, 0x1300000) is: 1
+> [1] The result of the (shifted) communication check for block (0x1300010, 0x3456789) is: 1
+> [1] The result of the (shifted) communication check for block (0x3457890, 0x4444444) is: 1
+> [1] The result of the (shifted) communication check for block (0x5555555, 0x5555565) is: 1
+> [1] The result of the (shifted) communication check for block (0x5600000, 0x8000000) is: 0
--- /dev/null
+/* Copyright (c) 2009-2015. The SimGrid Team.
+ * All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
+#include <stdio.h>
+#include <mpi.h>
+#include <stdint.h>
+#include <inttypes.h>
+
+// Set the elements between buf[start] and buf[stop-1] to (i+value)%256
+static void set(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ for(size_t i = start; i < stop; i++) {
+ buf[i] = (i+value)%256;
+ }
+}
+
+// Return the number of times that an element is equal to (i+value)%256 between buf[start] and buf[stop-1].
+static int count_all(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ size_t occ = 0;
+ for(size_t i = start ; i < stop ; i++) {
+ if(buf[i] == (i+value)%256) {
+ occ ++;
+ }
+ }
+ return occ;
+}
+
+// Return true iff the values from buf[start] to buf[stop-1] are all equal to (i+value)%256.
+static int check_all(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ size_t occ = count_all(buf, start, stop, value);
+ return occ == stop-start;
+}
+
+// Return true iff "enough" elements are equal to (i+value)%256 between buf[start] and buf[stop-1].
+static int check_enough(uint8_t *buf, size_t start, size_t stop, uint8_t value) {
+ int page_size = 0x1000;
+ size_t size = stop-start;
+ if(size <= 2*page_size) // we are not sure to have a whole page that is shared
+ return 1;
+ size_t occ = count_all(buf, start, stop, value);
+ return occ >= size - 2*page_size;
+}
+
+int main(int argc, char *argv[])
+{
+ MPI_Init(&argc, &argv);
+ int rank;
+ int size;
+ size_t mem_size = 0x10000000;
+ size_t shared_blocks[] = {
+ 0, 0x1234567,
+ 0x1300000, 0x1300010,
+ 0x3456789, 0x3457890,
+ 0x4444444, 0x5555555,
+ 0x5555565, 0x5600000,
+ 0x8000000, 0x10000000
+ };
+ int nb_blocks = (sizeof(shared_blocks)/sizeof(size_t))/2;
+ MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+ MPI_Comm_size(MPI_COMM_WORLD, &size);
+ //Let's Allocate a shared memory buffer
+ uint8_t *buf;
+ buf = SMPI_PARTIAL_SHARED_MALLOC(mem_size, shared_blocks, nb_blocks);
+ set(buf, 0, mem_size, 0);
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ // Process 0 write in shared blocks
+ if(rank == 0) {
+ for(int i = 0; i < nb_blocks; i++) {
+ size_t start = shared_blocks[2*i];
+ size_t stop = shared_blocks[2*i+1];
+ set(buf, start, stop, 42);
+ }
+ }
+ MPI_Barrier(MPI_COMM_WORLD);
+ // All processes check that their shared blocks have been written (at least partially)
+ for(int i = 0; i < nb_blocks; i++) {
+ size_t start = shared_blocks[2*i];
+ size_t stop = shared_blocks[2*i+1];
+ int is_shared = check_enough(buf, start, stop, 42);
+ printf("[%d] The result of the shared check for block (0x%lx, 0x%lx) is: %d\n", rank, start, stop, is_shared);
+ }
+
+
+ // Check the private blocks
+ MPI_Barrier(MPI_COMM_WORLD);
+ for(int i = 0; i < nb_blocks-1; i++) {
+ size_t start = shared_blocks[2*i+1];
+ size_t stop = shared_blocks[2*i+2];
+ int is_private = check_all(buf, start, stop, 0);
+ printf("[%d] The result of the private check for block (0x%lx, 0x%lx) is: %d\n", rank, start, stop, is_private);
+ }
+
+ SMPI_SHARED_FREE(buf);
+
+ MPI_Finalize();
+ return 0;
+}
--- /dev/null
+p Test compute
+! setenv LD_LIBRARY_PATH=../../lib
+! output sort
+! timeout 5
+$ ${bindir:=.}/../../../smpi_script/bin/smpirun -hostfile ../hostfile -platform ../../../examples/platforms/small_platform.xml -np 4 ${bindir:=.}/macro-partial-shared --log=smpi_kernel.thres:warning --log=xbt_cfg.thres:warning
+> [0] The result of the shared check for block (0x0, 0x1234567) is: 1
+> [0] The result of the shared check for block (0x1300000, 0x1300010) is: 1
+> [0] The result of the shared check for block (0x3456789, 0x3457890) is: 1
+> [0] The result of the shared check for block (0x4444444, 0x5555555) is: 1
+> [0] The result of the shared check for block (0x5555565, 0x5600000) is: 1
+> [0] The result of the shared check for block (0x8000000, 0x10000000) is: 1
+> [3] The result of the shared check for block (0x0, 0x1234567) is: 1
+> [3] The result of the shared check for block (0x1300000, 0x1300010) is: 1
+> [3] The result of the shared check for block (0x3456789, 0x3457890) is: 1
+> [3] The result of the shared check for block (0x4444444, 0x5555555) is: 1
+> [3] The result of the shared check for block (0x5555565, 0x5600000) is: 1
+> [3] The result of the shared check for block (0x8000000, 0x10000000) is: 1
+> [1] The result of the shared check for block (0x0, 0x1234567) is: 1
+> [1] The result of the shared check for block (0x1300000, 0x1300010) is: 1
+> [1] The result of the shared check for block (0x3456789, 0x3457890) is: 1
+> [1] The result of the shared check for block (0x4444444, 0x5555555) is: 1
+> [1] The result of the shared check for block (0x5555565, 0x5600000) is: 1
+> [1] The result of the shared check for block (0x8000000, 0x10000000) is: 1
+> [2] The result of the shared check for block (0x0, 0x1234567) is: 1
+> [2] The result of the shared check for block (0x1300000, 0x1300010) is: 1
+> [2] The result of the shared check for block (0x3456789, 0x3457890) is: 1
+> [2] The result of the shared check for block (0x4444444, 0x5555555) is: 1
+> [2] The result of the shared check for block (0x5555565, 0x5600000) is: 1
+> [2] The result of the shared check for block (0x8000000, 0x10000000) is: 1
+> [0] The result of the private check for block (0x1234567, 0x1300000) is: 1
+> [0] The result of the private check for block (0x1300010, 0x3456789) is: 1
+> [0] The result of the private check for block (0x3457890, 0x4444444) is: 1
+> [0] The result of the private check for block (0x5555555, 0x5555565) is: 1
+> [0] The result of the private check for block (0x5600000, 0x8000000) is: 1
+> [3] The result of the private check for block (0x1234567, 0x1300000) is: 1
+> [3] The result of the private check for block (0x1300010, 0x3456789) is: 1
+> [3] The result of the private check for block (0x3457890, 0x4444444) is: 1
+> [3] The result of the private check for block (0x5555555, 0x5555565) is: 1
+> [3] The result of the private check for block (0x5600000, 0x8000000) is: 1
+> [1] The result of the private check for block (0x1234567, 0x1300000) is: 1
+> [1] The result of the private check for block (0x1300010, 0x3456789) is: 1
+> [1] The result of the private check for block (0x3457890, 0x4444444) is: 1
+> [1] The result of the private check for block (0x5555555, 0x5555565) is: 1
+> [1] The result of the private check for block (0x5600000, 0x8000000) is: 1
+> [2] The result of the private check for block (0x1234567, 0x1300000) is: 1
+> [2] The result of the private check for block (0x1300010, 0x3456789) is: 1
+> [2] The result of the private check for block (0x3457890, 0x4444444) is: 1
+> [2] The result of the private check for block (0x5555555, 0x5555565) is: 1
+> [2] The result of the private check for block (0x5600000, 0x8000000) is: 1
