Skip to content

Commit

Permalink
ch4/ofi: Remove overly complex noncontiguous put/get
Browse files Browse the repository at this point in the history 
This implementation complicates datatype optimization, and also has
correctness issues. See pmodels#4468.
  • Loading branch information
@raffenet @pavanbalaji
raffenet authored and pavanbalaji committed May 6, 2020
1 parent 1fec9f0 commit 1bdb0f2
Show file tree
Hide file tree
Showing 2 changed files with 21 additions and 195 deletions.
211 changes: 20 additions & 191 deletions src/mpid/ch4/netmod/ofi/ofi_rma.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -176,63 +176,6 @@ MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_sigreq_complete(MPIR_Request ** sigreq)
}
}

/* _count: count of data elements of certain datatype
* _datatype: the datatype
* _bytes: total byte size
* Note: _bytes is calculated based on _count & _datatype and passed in here for reusing. */
MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_allocate_win_request_put_get(MPIR_Win * win,
int origin_count,
int target_count,
int target_rank,
MPI_Datatype origin_datatype,
MPI_Datatype target_datatype,
size_t origin_bytes,
size_t target_bytes,
size_t max_pipe,
MPIDI_OFI_win_request_t **
winreq, uint64_t * flags,
struct fid_ep **ep,
MPIR_Request ** sigreq)
{
int mpi_errno = MPI_SUCCESS;
size_t o_size, t_size, alloc_iovs, alloc_iov_size;
MPIDI_OFI_win_request_t *req = NULL;

MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_OFI_ALLOCATE_WIN_REQUEST_PUT_GET);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_OFI_ALLOCATE_WIN_REQUEST_PUT_GET);

o_size = sizeof(struct iovec);
t_size = sizeof(struct fi_rma_iov);
/* An upper bound on the iov limit */
MPIDI_OFI_count_iovecs(origin_count, target_count, 0, origin_datatype,
target_datatype, MPI_DATATYPE_NULL, origin_bytes, target_bytes, 0,
max_pipe, &alloc_iovs);

alloc_iov_size = alloc_iovs * o_size + alloc_iovs * t_size;

req = MPIDI_OFI_win_request_create();
MPIR_ERR_CHKANDSTMT((req) == NULL, mpi_errno, MPIX_ERR_NOREQ, goto fn_fail, "**nomemreq");
req->noncontig.iov_store = (char *) MPL_malloc(alloc_iov_size, MPL_MEM_BUFFER);
MPIR_ERR_CHKANDSTMT((req->noncontig.iov_store) == NULL, mpi_errno, MPI_ERR_NO_MEM, goto fn_fail,
"**nomem");
*winreq = req;

req->noncontig.iov.put_get.originv = (struct iovec *) req->noncontig.iov_store;
req->noncontig.iov.put_get.targetv =
(struct fi_rma_iov *) (req->noncontig.iov_store + o_size * alloc_iovs);
MPIDI_OFI_INIT_SIGNAL_REQUEST(win, sigreq, flags);
*ep = MPIDI_OFI_WIN(win).ep;
req->target_rank = target_rank;

fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_OFI_ALLOCATE_WIN_REQUEST_PUT_GET);
return mpi_errno;
fn_fail:
MPL_free(req);
req = NULL;
goto fn_exit;
}

/* _count: count of data elements of certain datatype
* _datatype: the datatype
* _bytes: total byte size
Expand Down Expand Up @@ -364,16 +307,10 @@ static inline int MPIDI_OFI_do_put(const void *origin_addr,
MPI_Datatype target_datatype,
MPIR_Win * win, MPIDI_av_entry_t * addr, MPIR_Request ** sigreq)
{
int rc, mpi_errno = MPI_SUCCESS;
MPIDI_OFI_win_request_t *req = NULL;
size_t offset, omax, tmax, tout, oout;
int mpi_errno = MPI_SUCCESS;
size_t offset;
uint64_t flags;
struct fid_ep *ep = NULL;
struct fi_msg_rma msg;
unsigned i;
struct iovec *originv;
struct fi_rma_iov *targetv;
MPIDI_OFI_seg_state_t p;
int target_contig, origin_contig;
size_t target_bytes, origin_bytes;
MPI_Aint origin_true_lb, target_true_lb;
Expand Down Expand Up @@ -447,64 +384,14 @@ static inline int MPIDI_OFI_do_put(const void *origin_addr,
}

/* noncontiguous messages */
mpi_errno =
MPIDI_OFI_allocate_win_request_put_get(win, origin_count, target_count, target_rank,
origin_datatype, target_datatype, origin_bytes,
target_bytes, MPIDI_OFI_global.max_msg_size, &req,
&flags, &ep, sigreq);
MPIR_ERR_CHECK(mpi_errno);

offset = target_disp * MPIDI_OFI_winfo_disp_unit(win, target_rank);

req->event_id = MPIDI_OFI_EVENT_ABORT;
msg.desc = NULL;
msg.addr = MPIDI_OFI_av_to_phys(addr);
msg.context = NULL;
msg.data = 0;
req->next = MPIDI_OFI_WIN(win).syncQ;
MPIDI_OFI_WIN(win).syncQ = req;
MPIDI_OFI_init_seg_state(&p,
origin_addr,
MPIDI_OFI_winfo_base(win, req->target_rank) + offset,
origin_count,
target_count,
origin_bytes,
target_bytes,
MPIDI_OFI_global.max_msg_size, origin_datatype, target_datatype);
rc = MPIDI_OFI_SEG_EAGAIN;

size_t cur_o = 0, cur_t = 0;
while (rc == MPIDI_OFI_SEG_EAGAIN) {
originv = &req->noncontig.iov.put_get.originv[cur_o];
targetv = &req->noncontig.iov.put_get.targetv[cur_t];
omax = MPIDI_OFI_global.rma_iov_limit;
tmax = MPIDI_OFI_global.rma_iov_limit;
rc = MPIDI_OFI_merge_segment(&p, originv, omax, targetv, tmax, &oout, &tout);

if (rc == MPIDI_OFI_SEG_DONE)
break;

cur_o += oout;
cur_t += tout;
for (i = 0; i < tout; i++)
targetv[i].key = MPIDI_OFI_winfo_mr_key(win, target_rank);
MPIR_Assert(rc != MPIDI_OFI_SEG_ERROR);
msg.msg_iov = originv;
msg.iov_count = oout;
msg.rma_iov = targetv;
msg.rma_iov_count = tout;
MPIDI_OFI_INIT_CHUNK_CONTEXT(win, sigreq);
MPIDI_OFI_CALL_RETRY(fi_writemsg(ep, &msg, flags), rdma_write, FALSE);

/* By default, progress is called only during fence, which significantly
* slows down the RMA operations for large non-contiguous data. Adding manual
* progress helps improve the performance. */
MPIDI_OFI_win_trigger_rma_progress(win);
}

MPIDI_OFI_finalize_seg_state(p);
/* Complete signal request to inform completion to user. */
MPIDI_OFI_sigreq_complete(sigreq);
if (sigreq)
mpi_errno =
MPIDIG_mpi_rput(origin_addr, origin_count, origin_datatype, target_rank, target_disp,
target_count, target_datatype, win, sigreq);
else
mpi_errno =
MPIDIG_mpi_put(origin_addr, origin_count, origin_datatype, target_rank, target_disp,
target_count, target_datatype, win);

fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_OFI_DO_PUT);
Expand Down Expand Up @@ -557,16 +444,10 @@ static inline int MPIDI_OFI_do_get(void *origin_addr,
MPI_Datatype target_datatype,
MPIR_Win * win, MPIDI_av_entry_t * addr, MPIR_Request ** sigreq)
{
int rc, mpi_errno = MPI_SUCCESS;
MPIDI_OFI_win_request_t *req = NULL;
size_t offset, omax, tmax, tout, oout;
int mpi_errno = MPI_SUCCESS;
size_t offset;
uint64_t flags;
struct fid_ep *ep = NULL;
struct fi_msg_rma msg;
struct iovec *originv;
struct fi_rma_iov *targetv;
unsigned i;
MPIDI_OFI_seg_state_t p;
int origin_contig, target_contig;
MPI_Aint origin_true_lb, target_true_lb;
size_t origin_bytes, target_bytes;
Expand Down Expand Up @@ -628,66 +509,14 @@ static inline int MPIDI_OFI_do_get(void *origin_addr,
}

/* noncontiguous messages */
mpi_errno =
MPIDI_OFI_allocate_win_request_put_get(win, origin_count, target_count, target_rank,
origin_datatype, target_datatype, origin_bytes,
target_bytes, MPIDI_OFI_global.max_msg_size, &req,
&flags, &ep, sigreq);
MPIR_ERR_CHECK(mpi_errno);

offset = target_disp * MPIDI_OFI_winfo_disp_unit(win, target_rank);
req->event_id = MPIDI_OFI_EVENT_ABORT;
msg.desc = NULL;
msg.addr = MPIDI_OFI_av_to_phys(addr);
msg.context = NULL;
msg.data = 0;
req->next = MPIDI_OFI_WIN(win).syncQ;
MPIDI_OFI_WIN(win).syncQ = req;
MPIDI_OFI_init_seg_state(&p,
origin_addr,
MPIDI_OFI_winfo_base(win, req->target_rank) + offset,
origin_count,
target_count,
origin_bytes,
target_bytes,
MPIDI_OFI_global.max_msg_size, origin_datatype, target_datatype);
rc = MPIDI_OFI_SEG_EAGAIN;

size_t cur_o = 0, cur_t = 0;
while (rc == MPIDI_OFI_SEG_EAGAIN) {
originv = &req->noncontig.iov.put_get.originv[cur_o];
targetv = &req->noncontig.iov.put_get.targetv[cur_t];
omax = MPIDI_OFI_global.rma_iov_limit;
tmax = MPIDI_OFI_global.rma_iov_limit;

rc = MPIDI_OFI_merge_segment(&p, originv, omax, targetv, tmax, &oout, &tout);

if (rc == MPIDI_OFI_SEG_DONE)
break;

cur_o += oout;
cur_t += tout;
MPIR_Assert(rc != MPIDI_OFI_SEG_ERROR);

for (i = 0; i < tout; i++)
targetv[i].key = MPIDI_OFI_winfo_mr_key(win, target_rank);

msg.msg_iov = originv;
msg.iov_count = oout;
msg.rma_iov = targetv;
msg.rma_iov_count = tout;
MPIDI_OFI_INIT_CHUNK_CONTEXT(win, sigreq);
MPIDI_OFI_CALL_RETRY(fi_readmsg(ep, &msg, flags), rdma_write, FALSE);

/* By default, progress is called only during fence, which significantly
* slows down the RMA operations for large non-contiguous data. Adding manual
* progress helps improve the performance. */
MPIDI_OFI_win_trigger_rma_progress(win);
}

MPIDI_OFI_finalize_seg_state(p);
/* Complete signal request to inform completion to user. */
MPIDI_OFI_sigreq_complete(sigreq);
if (sigreq)
mpi_errno =
MPIDIG_mpi_rget(origin_addr, origin_count, origin_datatype, target_rank, target_disp,
target_count, target_datatype, win, sigreq);
else
mpi_errno =
MPIDIG_mpi_get(origin_addr, origin_count, origin_datatype, target_rank, target_disp,
target_count, target_datatype, win);

fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_OFI_DO_GET);
Expand Down
5 changes: 1 addition & 4 deletions src/mpid/ch4/netmod/ofi/ofi_types.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -459,15 +459,12 @@ typedef struct MPIDI_OFI_win_request {
int event_id; /* fixed field, do not move */
struct MPIDI_OFI_win_request *next;
int target_rank;
char *pack_buffer;
struct {
char pad[MPIDI_REQUEST_HDR_SIZE];
struct fi_context context[MPIDI_OFI_CONTEXT_STRUCTS]; /* fixed field, do not move */
int event_id; /* fixed field, do not move */
union {
struct {
struct iovec *originv;
struct fi_rma_iov *targetv;
} put_get;
struct {
struct fi_ioc *originv;
struct fi_rma_ioc *targetv;
Expand Down

0 comments on commit 1bdb0f2

Please sign in to comment.