ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

OSHMEM: mxm: optimize mxm_peer layout.

Просмотр исходного кода 
Signed-off-by: Alex Mikheev <alexm@mellanox.com>
Этот коммит содержится в:
Alex Mikheev 2016-10-27 14:00:46 +03:00
родитель df74d549dc
Коммит 23c3dc8345
5 изменённых файлов: 71 добавлений и 130 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

8
oshmem/mca/atomic/mxm/atomic_mxm_cswap.c
Просмотреть файл

@ -34,13 +34,11 @@ int mca_atomic_mxm_cswap(void *target,
unsigned my_pe; unsigned my_pe;
uint8_t nlong_order; uint8_t nlong_order;
void *remote_addr; void *remote_addr;
int ptl_id;
mxm_send_req_t sreq; mxm_send_req_t sreq;
mxm_error_t mxm_err; mxm_error_t mxm_err;
sshmem_mkey_t *r_mkey; sshmem_mkey_t *r_mkey;
my_pe = oshmem_my_proc_id(); my_pe = oshmem_my_proc_id();
ptl_id = -1;
mxm_err = MXM_OK; mxm_err = MXM_OK;
switch (nlong) { switch (nlong) {
@ -62,11 +60,7 @@ int mca_atomic_mxm_cswap(void *target,
return OSHMEM_ERR_BAD_PARAM; return OSHMEM_ERR_BAD_PARAM;
} }
ptl_id = OSHMEM_PROC_DATA(oshmem_proc_group_all(pe))->transport_ids[0]; r_mkey = mca_memheap_base_get_cached_mkey(pe, target, MXM_PTL_RDMA, &remote_addr);
if (MXM_PTL_SHM == ptl_id) {
ptl_id = MXM_PTL_RDMA;
}
r_mkey = mca_memheap_base_get_cached_mkey(pe, target, ptl_id, &remote_addr);
if (!r_mkey) { if (!r_mkey) {
ATOMIC_ERROR("[#%d] %p is not address of symmetric variable", ATOMIC_ERROR("[#%d] %p is not address of symmetric variable",
my_pe, target); my_pe, target);

8
oshmem/mca/atomic/mxm/atomic_mxm_fadd.c
Просмотреть файл

@ -35,14 +35,12 @@ int mca_atomic_mxm_fadd(void *target,
unsigned my_pe; unsigned my_pe;
uint8_t nlong_order; uint8_t nlong_order;
void *remote_addr; void *remote_addr;
int ptl_id;
mxm_send_req_t sreq; mxm_send_req_t sreq;
mxm_error_t mxm_err; mxm_error_t mxm_err;
sshmem_mkey_t *r_mkey; sshmem_mkey_t *r_mkey;
static char dummy_buf[8]; static char dummy_buf[8];
my_pe = oshmem_my_proc_id(); my_pe = oshmem_my_proc_id();
ptl_id = -1;
mxm_err = MXM_OK; mxm_err = MXM_OK;
switch (nlong) { switch (nlong) {
@ -64,11 +62,7 @@ int mca_atomic_mxm_fadd(void *target,
return OSHMEM_ERR_BAD_PARAM; return OSHMEM_ERR_BAD_PARAM;
} }
ptl_id = OSHMEM_PROC_DATA(oshmem_proc_group_all(pe))->transport_ids[0]; r_mkey = mca_memheap_base_get_cached_mkey(pe, target, MXM_PTL_RDMA, &remote_addr);
if (MXM_PTL_SHM == ptl_id) {
ptl_id = MXM_PTL_RDMA;
}
r_mkey = mca_memheap_base_get_cached_mkey(pe, target, ptl_id, &remote_addr);
if (!r_mkey) { if (!r_mkey) {
ATOMIC_ERROR("[#%d] %p is not address of symmetric variable", ATOMIC_ERROR("[#%d] %p is not address of symmetric variable",
my_pe, target); my_pe, target);

2
oshmem/mca/spml/base/base.h
Просмотреть файл

@ -104,6 +104,8 @@ OSHMEM_DECLSPEC extern mca_base_framework_t oshmem_spml_base_framework;
#define SPML_VERBOSE(level, ...) #define SPML_VERBOSE(level, ...)
#endif #endif
#define SPML_VERBOSE_FASTPATH(level, ...)
#define SPML_ERROR(...) \ #define SPML_ERROR(...) \
oshmem_output(oshmem_spml_base_framework.framework_output, \ oshmem_output(oshmem_spml_base_framework.framework_output, \
"Error %s:%d - %s()", __SPML_FILE__, __LINE__, __func__, __VA_ARGS__) "Error %s:%d - %s()", __SPML_FILE__, __LINE__, __func__, __VA_ARGS__)

156
oshmem/mca/spml/ikrit/spml_ikrit.c
Просмотреть файл

@ -55,6 +55,11 @@ do { \
} \ } \
} while(0) } while(0)
static int mca_spml_ikrit_get_async(void *src_addr,
size_t size,
void *dst_addr,
int src);
typedef struct spml_ikrit_am_hdr { typedef struct spml_ikrit_am_hdr {
uint64_t va; uint64_t va;
} spml_ikrit_am_hdr_t; } spml_ikrit_am_hdr_t;
@ -279,47 +284,20 @@ int mca_spml_ikrit_enable(bool enable)
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
} }
static int create_ptl_idx(int dst_pe)
{
ompi_proc_t *proc;
proc = oshmem_proc_group_find(oshmem_group_all, dst_pe);
OSHMEM_PROC_DATA(proc)->transport_ids = (char *) malloc(MXM_PTL_LAST * sizeof(char));
if (NULL == OSHMEM_PROC_DATA(proc)->transport_ids)
return OSHMEM_ERROR;
OSHMEM_PROC_DATA(proc)->num_transports = 1;
OSHMEM_PROC_DATA(proc)->transport_ids[0] = MXM_PTL_RDMA;
return OSHMEM_SUCCESS;
}
static void destroy_ptl_idx(int dst_pe)
{
ompi_proc_t *proc;
proc = oshmem_proc_group_find(oshmem_group_all, dst_pe);
if (NULL != OSHMEM_PROC_DATA(proc)->transport_ids)
free(OSHMEM_PROC_DATA(proc)->transport_ids);
}
static void mxm_peer_construct(mxm_peer_t *p) static void mxm_peer_construct(mxm_peer_t *p)
{ {
p->pe = -1; p->pe = -1;
p->n_active_puts = 0; p->n_active_puts = 0;
p->need_fence = 0; p->need_fence = 0;
p->ptl_id = MXM_PTL_RDMA;
OBJ_CONSTRUCT(&p->link, opal_list_item_t);
} }
static void mxm_peer_destruct(mxm_peer_t *p) static void mxm_peer_destruct(mxm_peer_t *p)
{ {
/* may be we need to remov item from list */ OBJ_DESTRUCT(&p->link);
} }
OBJ_CLASS_INSTANCE( mxm_peer_t,
opal_list_item_t,
mxm_peer_construct,
mxm_peer_destruct);
int mca_spml_ikrit_del_procs(ompi_proc_t** procs, size_t nprocs) int mca_spml_ikrit_del_procs(ompi_proc_t** procs, size_t nprocs)
{ {
size_t i, n; size_t i, n;
@ -341,8 +319,7 @@ int mca_spml_ikrit_del_procs(ompi_proc_t** procs, size_t nprocs)
assert(mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn != mca_spml_ikrit.mxm_peers[i].mxm_conn); assert(mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn != mca_spml_ikrit.mxm_peers[i].mxm_conn);
mxm_ep_disconnect(mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn); mxm_ep_disconnect(mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn);
} }
destroy_ptl_idx(i); mxm_peer_destruct(&mca_spml_ikrit.mxm_peers[i]);
OBJ_DESTRUCT(&mca_spml_ikrit.mxm_peers[i]);
} }
free(mca_spml_ikrit.mxm_peers); free(mca_spml_ikrit.mxm_peers);
@ -377,7 +354,7 @@ int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
} }
} }
mca_spml_ikrit.mxm_peers = (mxm_peer_t *) malloc(nprocs * sizeof(mxm_peer_t)); mca_spml_ikrit.mxm_peers = (mxm_peer_t *) calloc(nprocs , sizeof(mxm_peer_t));
if (NULL == mca_spml_ikrit.mxm_peers) { if (NULL == mca_spml_ikrit.mxm_peers) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE; rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail; goto bail;
@ -413,8 +390,7 @@ int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
/* mxm 2.0 keeps its connections on a list. Make sure /* mxm 2.0 keeps its connections on a list. Make sure
* that list have different order on every rank */ * that list have different order on every rank */
i = (my_rank + n) % nprocs; i = (my_rank + n) % nprocs;
OBJ_CONSTRUCT(&mca_spml_ikrit.mxm_peers[i], mxm_peer_t); mxm_peer_construct(&mca_spml_ikrit.mxm_peers[i]);
mca_spml_ikrit.mxm_peers[i].pe = i; mca_spml_ikrit.mxm_peers[i].pe = i;
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, ep_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_conn); err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, ep_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_conn);
@ -422,8 +398,6 @@ int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err)); SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail; goto bail;
} }
if (OSHMEM_SUCCESS != create_ptl_idx(i))
goto bail;
mxm_conn_ctx_set(mca_spml_ikrit.mxm_peers[i].mxm_conn, &mca_spml_ikrit.mxm_peers[i]); mxm_conn_ctx_set(mca_spml_ikrit.mxm_peers[i].mxm_conn, &mca_spml_ikrit.mxm_peers[i]);
if (mca_spml_ikrit.hw_rdma_channel) { if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_connect(mca_spml_ikrit.mxm_hw_rdma_ep, ep_hw_rdma_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn); err = mxm_ep_connect(mca_spml_ikrit.mxm_hw_rdma_ep, ep_hw_rdma_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn);
@ -457,10 +431,8 @@ int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
if (procs[i] == proc_self) if (procs[i] == proc_self)
continue; continue;
/* use zcopy for put/get via sysv shared memory */ /* use zcopy for put/get via sysv shared memory with fallback to RDMA */
OSHMEM_PROC_DATA(procs[i])->transport_ids[0] = MXM_PTL_SHM; mca_spml_ikrit.mxm_peers[i].ptl_id = MXM_PTL_SHM;
OSHMEM_PROC_DATA(procs[i])->transport_ids[1] = MXM_PTL_RDMA;
OSHMEM_PROC_DATA(procs[i])->num_transports = 2;
} }
SPML_VERBOSE(50, "*** ADDED PROCS ***"); SPML_VERBOSE(50, "*** ADDED PROCS ***");
@ -591,16 +563,7 @@ int mca_spml_ikrit_deregister(sshmem_mkey_t *mkeys)
static inline int get_ptl_id(int dst) static inline int get_ptl_id(int dst)
{ {
ompi_proc_t *proc; return mca_spml_ikrit.mxm_peers[dst].ptl_id;
/* get endpoint and btl */
proc = oshmem_proc_group_all(dst);
if (!proc) {
SPML_ERROR("Can not find destination proc for pe=%d", dst);
oshmem_shmem_abort(-1);
return -1;
}
return OSHMEM_PROC_DATA(proc)->transport_ids[0];
} }
int mca_spml_ikrit_oob_get_mkeys(int pe, uint32_t seg, sshmem_mkey_t *mkeys) int mca_spml_ikrit_oob_get_mkeys(int pe, uint32_t seg, sshmem_mkey_t *mkeys)
@ -625,27 +588,22 @@ int mca_spml_ikrit_oob_get_mkeys(int pe, uint32_t seg, sshmem_mkey_t *mkeys)
return OSHMEM_ERROR; return OSHMEM_ERROR;
} }
static int mca_spml_ikrit_get_helper(mxm_send_req_t *sreq, static inline int mca_spml_ikrit_get_helper(mxm_send_req_t *sreq,
void *src_addr, void *src_addr,
size_t size, size_t size,
void *dst_addr, void *dst_addr,
int src) int src)
{ {
/* shmem spec states that get() operations are blocking. So it is enough /* shmem spec states that get() operations are blocking. So it is enough
to have single mxm request. Also we count on mxm doing copy */ to have single mxm request. Also we count on mxm doing copy */
void *rva; void *rva;
sshmem_mkey_t *r_mkey; sshmem_mkey_t *r_mkey;
int ptl_id;
ptl_id = get_ptl_id(src);
/* already tried to send via shm and failed. go via rdma */ /* already tried to send via shm and failed. go via rdma */
if (ptl_id == MXM_PTL_SHM)
ptl_id = MXM_PTL_RDMA;
/** /**
* Get the address to the remote rkey. * Get the address to the remote rkey.
**/ **/
r_mkey = mca_memheap_base_get_cached_mkey(src, src_addr, ptl_id, &rva); r_mkey = mca_memheap_base_get_cached_mkey(src, src_addr, MXM_PTL_RDMA, &rva);
if (!r_mkey) { if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable", SPML_ERROR("pe=%d: %p is not address of shared variable",
src, src_addr); src, src_addr);
@ -653,9 +611,10 @@ static int mca_spml_ikrit_get_helper(mxm_send_req_t *sreq,
return OSHMEM_ERROR; return OSHMEM_ERROR;
} }
SPML_VERBOSE(100, SPML_VERBOSE_FASTPATH(100,
"get: pe:%d ptl=%d src=%p -> dst: %p sz=%d. src_rva=%p, %s", "get: pe:%d ptl=%d src=%p -> dst: %p sz=%d. src_rva=%p, %s",
src, ptl_id, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey)); src, MXM_PTL_RDMA, src_addr, dst_addr, (int)size, (void *)rva,
mca_spml_base_mkey2str(r_mkey));
/* mxm does not really cares for get lkey */ /* mxm does not really cares for get lkey */
sreq->base.mq = mca_spml_ikrit.mxm_mq; sreq->base.mq = mca_spml_ikrit.mxm_mq;
@ -698,9 +657,10 @@ static inline int mca_spml_ikrit_get_shm(void *src_addr,
if (!mca_memheap_base_can_local_copy(r_mkey, src_addr)) if (!mca_memheap_base_can_local_copy(r_mkey, src_addr))
return OSHMEM_ERROR; return OSHMEM_ERROR;
SPML_VERBOSE(100, SPML_VERBOSE_FASTPATH(100,
"shm get: pe:%d src=%p -> dst: %p sz=%d. src_rva=%p, %s", "shm get: pe:%d src=%p -> dst: %p sz=%d. src_rva=%p, %s",
src, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey)); src, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
memcpy(dst_addr, (void *) (unsigned long) rva, size); memcpy(dst_addr, (void *) (unsigned long) rva, size);
opal_progress(); opal_progress();
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
@ -763,11 +723,10 @@ static inline void get_completion_cb(void *ctx)
oshmem_request_free((oshmem_request_t**) &get_req); oshmem_request_free((oshmem_request_t**) &get_req);
} }
/* extension. used 4 fence implementation b4 fence was added to mxm */ static inline int mca_spml_ikrit_get_async(void *src_addr,
int mca_spml_ikrit_get_async(void *src_addr, size_t size,
size_t size, void *dst_addr,
void *dst_addr, int src)
int src)
{ {
mca_spml_ikrit_get_request_t *get_req; mca_spml_ikrit_get_request_t *get_req;
@ -868,7 +827,7 @@ static inline void put_completion_cb(void *ctx)
peer->n_active_puts--; peer->n_active_puts--;
if (0 == peer->n_active_puts && if (0 == peer->n_active_puts &&
(put_req->mxm_req.opcode == MXM_REQ_OP_PUT_SYNC)) { (put_req->mxm_req.opcode == MXM_REQ_OP_PUT_SYNC)) {
opal_list_remove_item(&mca_spml_ikrit.active_peers, &peer->super); opal_list_remove_item(&mca_spml_ikrit.active_peers, &peer->link);
peer->need_fence = 0; peer->need_fence = 0;
} }
} }
@ -911,13 +870,11 @@ static inline int mca_spml_ikrit_put_internal(void* dst_addr,
return OSHMEM_ERROR; return OSHMEM_ERROR;
} }
#if SPML_IKRIT_PUT_DEBUG == 1 SPML_VERBOSE_FASTPATH(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva,
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s", mca_spml_base_mkey2str(r_mkey));
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
if (ptl_id == MXM_PTL_SHM) {
if (OPAL_UNLIKELY(MXM_PTL_SHM == ptl_id)) {
if (mca_memheap_base_can_local_copy(r_mkey, dst_addr)) { if (mca_memheap_base_can_local_copy(r_mkey, dst_addr)) {
memcpy((void *) (unsigned long) rva, src_addr, size); memcpy((void *) (unsigned long) rva, src_addr, size);
/* call progress as often as we would have with regular put */ /* call progress as often as we would have with regular put */
@ -926,8 +883,7 @@ static inline int mca_spml_ikrit_put_internal(void* dst_addr,
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
} }
/* segment not mapped - fallback to rmda */ /* segment not mapped - fallback to rmda */
ptl_id = MXM_PTL_RDMA; r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, MXM_PTL_RDMA, &rva);
r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, ptl_id, &rva);
if (!r_mkey) { if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable", SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr); dst, dst_addr);
@ -936,10 +892,9 @@ static inline int mca_spml_ikrit_put_internal(void* dst_addr,
} }
} }
#if SPML_IKRIT_PUT_DEBUG == 1 SPML_VERBOSE_FASTPATH(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s", dst, MXM_PTL_RDMA, dst_addr, src_addr, (int)size, (void *)rva,
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey)); mca_spml_base_mkey2str(r_mkey));
#endif
put_req = alloc_put_req(); put_req = alloc_put_req();
if (NULL == put_req) { if (NULL == put_req) {
@ -996,7 +951,7 @@ static inline int mca_spml_ikrit_put_internal(void* dst_addr,
OPAL_THREAD_ADD32(&mca_spml_ikrit.n_active_puts, 1); OPAL_THREAD_ADD32(&mca_spml_ikrit.n_active_puts, 1);
if (mca_spml_ikrit.mxm_peers[dst].need_fence == 0) { if (mca_spml_ikrit.mxm_peers[dst].need_fence == 0) {
opal_list_append(&mca_spml_ikrit.active_peers, opal_list_append(&mca_spml_ikrit.active_peers,
&mca_spml_ikrit.mxm_peers[dst].super); &mca_spml_ikrit.mxm_peers[dst].link);
mca_spml_ikrit.mxm_peers[dst].need_fence = 1; mca_spml_ikrit.mxm_peers[dst].need_fence = 1;
} }
@ -1038,12 +993,11 @@ int mca_spml_ikrit_put_simple(void* dst_addr,
return OSHMEM_ERROR; return OSHMEM_ERROR;
} }
#if SPML_IKRIT_PUT_DEBUG == 1 SPML_VERBOSE_FASTPATH(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s", dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva,
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey)); mca_spml_base_mkey2str(r_mkey));
#endif
if (ptl_id == MXM_PTL_SHM) {
if (MXM_PTL_SHM == ptl_id) {
if (mca_memheap_base_can_local_copy(r_mkey, dst_addr)) { if (mca_memheap_base_can_local_copy(r_mkey, dst_addr)) {
memcpy((void *) (unsigned long) rva, src_addr, size); memcpy((void *) (unsigned long) rva, src_addr, size);
/* call progress as often as we would have with regular put */ /* call progress as often as we would have with regular put */
@ -1052,11 +1006,10 @@ int mca_spml_ikrit_put_simple(void* dst_addr,
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
} }
/* segment not mapped - fallback to rmda */ /* segment not mapped - fallback to rmda */
ptl_id = MXM_PTL_RDMA;
r_mkey = mca_memheap_base_get_cached_mkey(dst, r_mkey = mca_memheap_base_get_cached_mkey(dst,
//(unsigned long) dst_addr, //(unsigned long) dst_addr,
dst_addr, dst_addr,
ptl_id, MXM_PTL_RDMA,
&rva); &rva);
if (!r_mkey) { if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable", SPML_ERROR("pe=%d: %p is not address of shared variable",
@ -1066,10 +1019,9 @@ int mca_spml_ikrit_put_simple(void* dst_addr,
} }
} }
#if SPML_IKRIT_PUT_DEBUG == 1 SPML_VERBOSE_FASTPATH(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s", dst, MXM_PTL_RDMA, dst_addr, src_addr, (int)size, (void *)rva,
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey)); mca_spml_base_mkey2str(r_mkey));
#endif
/* fill out request */ /* fill out request */
mxm_req.base.mq = mca_spml_ikrit.mxm_mq; mxm_req.base.mq = mca_spml_ikrit.mxm_mq;
@ -1089,7 +1041,7 @@ int mca_spml_ikrit_put_simple(void* dst_addr,
if (mca_spml_ikrit.mxm_peers[dst].need_fence == 0) { if (mca_spml_ikrit.mxm_peers[dst].need_fence == 0) {
opal_list_append(&mca_spml_ikrit.active_peers, opal_list_append(&mca_spml_ikrit.active_peers,
&mca_spml_ikrit.mxm_peers[dst].super); &mca_spml_ikrit.mxm_peers[dst].link);
mca_spml_ikrit.mxm_peers[dst].need_fence = 1; mca_spml_ikrit.mxm_peers[dst].need_fence = 1;
} }
@ -1163,7 +1115,7 @@ int mca_spml_ikrit_fence(void)
/* puts(unless are send sync) are completed by remote side lazily. That is either when remote decides to /* puts(unless are send sync) are completed by remote side lazily. That is either when remote decides to
* ack window which can take hundreds of ms. So speed things up by doing fence */ * ack window which can take hundreds of ms. So speed things up by doing fence */
while (NULL != (item = opal_list_remove_first(&mca_spml_ikrit.active_peers))) { while (NULL != (item = opal_list_remove_first(&mca_spml_ikrit.active_peers))) {
peer = (mxm_peer_t *) item; peer = spml_ikrit_container_of(item, mxm_peer_t, link);
peer->n_active_puts = 0; peer->n_active_puts = 0;
peer->need_fence = 0; peer->need_fence = 0;
mca_spml_ikrit_mxm_fence(peer->pe); mca_spml_ikrit_mxm_fence(peer->pe);

27
oshmem/mca/spml/ikrit/spml_ikrit.h
Просмотреть файл

@ -47,22 +47,31 @@
/* request explicit ack (SYNC) per every X put requests per connection */ /* request explicit ack (SYNC) per every X put requests per connection */
#define SPML_IKRIT_PACKETS_PER_SYNC 64 #define SPML_IKRIT_PACKETS_PER_SYNC 64
#define spml_ikrit_container_of(ptr, type, member) ( \
(type *)( ((char *)(ptr)) - offsetof(type,member) ))
#define MXM_MAX_ADDR_LEN 512
#define MXM_PTL_RDMA 0
#define MXM_PTL_SHM 1
#define MXM_PTL_LAST 2
BEGIN_C_DECLS BEGIN_C_DECLS
/** /**
* UD MXM SPML module * UD MXM SPML module
*/ */
struct mxm_peer { struct mxm_peer {
opal_list_item_t super;
mxm_conn_h mxm_conn; mxm_conn_h mxm_conn;
mxm_conn_h mxm_hw_rdma_conn; mxm_conn_h mxm_hw_rdma_conn;
int pe; uint8_t ptl_id;
opal_list_item_t link;
int32_t n_active_puts; int32_t n_active_puts;
int need_fence; uint32_t pe;
uint8_t need_fence;
}; };
typedef struct mxm_peer mxm_peer_t; typedef struct mxm_peer mxm_peer_t;
OBJ_CLASS_DECLARATION(mxm_peer_t);
struct mca_spml_ikrit_t { struct mca_spml_ikrit_t {
mca_spml_base_module_t super; mca_spml_base_module_t super;
@ -105,11 +114,6 @@ struct mca_spml_ikrit_t {
typedef struct mca_spml_ikrit_t mca_spml_ikrit_t; typedef struct mca_spml_ikrit_t mca_spml_ikrit_t;
#define MXM_MAX_ADDR_LEN 512
#define MXM_PTL_RDMA 0
#define MXM_PTL_SHM 1
#define MXM_PTL_LAST 2
typedef struct spml_ikrit_mxm_ep_conn_info_t { typedef struct spml_ikrit_mxm_ep_conn_info_t {
union { union {
@ -130,11 +134,6 @@ extern int mca_spml_ikrit_get_nb(void* src_addr,
void* dst_addr, void* dst_addr,
int src, int src,
void **handle); void **handle);
/* extension. used 4 fence implementation b4 fence was added to mxm */
extern int mca_spml_ikrit_get_async(void *src_addr,
size_t size,
void *dst_addr,
int src);
extern int mca_spml_ikrit_put(void* dst_addr, extern int mca_spml_ikrit_put(void* dst_addr,
size_t size, size_t size,