Skip to content

Commit 22fa5a8

Browse files
mkurnosovhjelmn
mkurnosov
authored and
hjelmn
committed
coll/base/scatter: replaces right skewed binomial tree (in order) with left skewed binomial tree
Current implementation of `coll/base/MPI_Scatter` is based on in-order binomial tree. This tree is right skewed and it provides good performance for a MPI_Gather operation. But for a MPI_Scatter operation left skewed binomial tree is effective. Signed-off-by: Mikhail Kurnosov <mkurnosov@gmail.com>
1 parent 77057ad commit 22fa5a8

File tree

1 file changed

+62
-43
lines changed

1 file changed

+62
-43
lines changed

‎ompi/mca/coll/base/coll_base_scatter.c

Lines changed: 62 additions & 43 deletions
Original file line numberDiff line numberDiff line change
@@ -34,47 +34,66 @@
3434
#include "coll_base_topo.h"
3535
#include "coll_base_util.h"
3636

37+
/*
38+
* ompi_coll_base_scatter_intra_binomial
39+
*
40+
* Function: Binomial tree algorithm for scatter
41+
* Accepts: Same as MPI_Scatter
42+
* Returns: MPI_SUCCESS or error code
43+
*
44+
* Time complexity: \alpha\log(p) + \beta*m((p-1)/p),
45+
* where m = scount * comm_size, p = comm_size
46+
*
47+
* Memory requirements (per process):
48+
* root process (root > 0): scount * comm_size * sdtype_size
49+
* non-root, non-leaf process: rcount * comm_size * rdtype_size
50+
*
51+
* Examples:
52+
* comm_size=8 comm_size=10 comm_size=12
53+
* 0 0 0
54+
* / | \ / / | \ / / \ \
55+
* 4 2 1 8 4 2 1 8 4 2 1
56+
* / | | / / | | / | / | |
57+
* 6 5 3 9 6 5 3 10 9 6 5 3
58+
* | | | |
59+
* 7 7 11 7
60+
*/
3761
int
38-
ompi_coll_base_scatter_intra_binomial( const void *sbuf, int scount,
39-
struct ompi_datatype_t *sdtype,
40-
void *rbuf, int rcount,
41-
struct ompi_datatype_t *rdtype,
42-
int root,
43-
struct ompi_communicator_t *comm,
44-
mca_coll_base_module_t *module)
62+
ompi_coll_base_scatter_intra_binomial(
63+
const void *sbuf, int scount, struct ompi_datatype_t *sdtype,
64+
void *rbuf, int rcount, struct ompi_datatype_t *rdtype,
65+
int root, struct ompi_communicator_t *comm,
66+
mca_coll_base_module_t *module)
4567
{
46-
int line = -1, i, rank, vrank, size, total_send = 0, err;
68+
int line = -1, rank, vrank, size, err;
4769
char *ptmp, *tempbuf = NULL;
48-
ompi_coll_tree_t* bmtree;
4970
MPI_Status status;
50-
mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module;
71+
mca_coll_base_module_t *base_module = (mca_coll_base_module_t*)module;
5172
mca_coll_base_comm_t *data = base_module->base_data;
5273
ptrdiff_t sextent, rextent, ssize, rsize, sgap = 0, rgap = 0;
5374

54-
5575
size = ompi_comm_size(comm);
5676
rank = ompi_comm_rank(comm);
5777

5878
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
59-
"ompi_coll_base_scatter_intra_binomial rank %d", rank));
60-
61-
/* create the binomial tree */
62-
COLL_BASE_UPDATE_IN_ORDER_BMTREE( comm, base_module, root );
63-
bmtree = data->cached_in_order_bmtree;
79+
"coll:base:scatter_intra_binomial rank %d/%d", rank, size));
6480

65-
ompi_datatype_type_extent(rdtype, &rextent);
66-
67-
rsize = opal_datatype_span(&rdtype->super, (int64_t)rcount * size, &rgap);
81+
/* Create the binomial tree */
82+
COLL_BASE_UPDATE_IN_ORDER_BMTREE(comm, base_module, root);
83+
if (NULL == data->cached_in_order_bmtree) {
84+
err = OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto err_hndl;
85+
}
86+
ompi_coll_tree_t *bmtree = data->cached_in_order_bmtree;
6887

6988
vrank = (rank - root + size) % size;
70-
ptmp = (char *) rbuf; /* by default suppose leaf nodes, just use rbuf */
89+
ptmp = (char *)rbuf; /* by default suppose leaf nodes, just use rbuf */
7190

7291
if (rank == root) {
7392
ompi_datatype_type_extent(sdtype, &sextent);
7493
ssize = opal_datatype_span(&sdtype->super, (int64_t)scount * size, &sgap);
7594
if (0 == root) {
7695
/* root on 0, just use the send buffer */
77-
ptmp = (char *) sbuf;
96+
ptmp = (char *)sbuf;
7897
if (rbuf != MPI_IN_PLACE) {
7998
/* local copy to rbuf */
8099
err = ompi_datatype_sndrcv(sbuf, scount, sdtype,
@@ -83,7 +102,7 @@ ompi_coll_base_scatter_intra_binomial( const void *sbuf, int scount,
83102
}
84103
} else {
85104
/* root is not on 0, allocate temp buffer for send */
86-
tempbuf = (char *) malloc(ssize);
105+
tempbuf = (char *)malloc(ssize);
87106
if (NULL == tempbuf) {
88107
err = OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto err_hndl;
89108
}
@@ -94,7 +113,6 @@ ompi_coll_base_scatter_intra_binomial( const void *sbuf, int scount,
94113
ptmp, (char *) sbuf + sextent * (ptrdiff_t)root * (ptrdiff_t)scount);
95114
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
96115

97-
98116
err = ompi_datatype_copy_content_same_ddt(sdtype, (ptrdiff_t)scount * (ptrdiff_t)root,
99117
ptmp + sextent * (ptrdiff_t)scount * (ptrdiff_t)(size - root), (char *)sbuf);
100118
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
@@ -106,53 +124,54 @@ ompi_coll_base_scatter_intra_binomial( const void *sbuf, int scount,
106124
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
107125
}
108126
}
109-
total_send = scount;
110127
} else if (!(vrank % 2)) {
111-
/* non-root, non-leaf nodes, allocte temp buffer for recv
128+
/* non-root, non-leaf nodes, allocate temp buffer for recv
112129
* the most we need is rcount*size/2 */
113-
tempbuf = (char *) malloc(rsize);
130+
ompi_datatype_type_extent(rdtype, &rextent);
131+
rsize = opal_datatype_span(&rdtype->super, (int64_t)rcount * size, &rgap);
132+
tempbuf = (char *)malloc(rsize / 2);
114133
if (NULL == tempbuf) {
115-
err= OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto err_hndl;
134+
err = OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto err_hndl;
116135
}
117136
ptmp = tempbuf - rgap;
118-
119137
sdtype = rdtype;
120138
scount = rcount;
121139
sextent = rextent;
122-
total_send = scount;
123140
}
124141

142+
int curr_count = (rank == root) ? scount * size : 0;
125143
if (!(vrank % 2)) {
126144
if (rank != root) {
127145
/* recv from parent on non-root */
128146
err = MCA_PML_CALL(recv(ptmp, (ptrdiff_t)rcount * (ptrdiff_t)size, rdtype, bmtree->tree_prev,
129147
MCA_COLL_BASE_TAG_SCATTER, comm, &status));
130148
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
149+
150+
/* Get received count */
151+
size_t rdtype_size;
152+
ompi_datatype_type_size(rdtype, &rdtype_size);
153+
curr_count = (int)(status._ucount / rdtype_size);
154+
131155
/* local copy to rbuf */
132156
err = ompi_datatype_sndrcv(ptmp, scount, sdtype,
133157
rbuf, rcount, rdtype);
134158
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
135159
}
136160
/* send to children on all non-leaf */
137-
for (i = 0; i < bmtree->tree_nextsize; i++) {
138-
size_t mycount = 0;
139-
int vkid;
161+
for (int i = bmtree->tree_nextsize - 1; i >= 0; i--) {
140162
/* figure out how much data I have to send to this child */
141-
vkid = (bmtree->tree_next[i] - root + size) % size;
142-
mycount = vkid - vrank;
143-
if( (int)mycount > (size - vkid) )
144-
mycount = size - vkid;
145-
mycount *= scount;
146-
147-
err = MCA_PML_CALL(send(ptmp + (ptrdiff_t)total_send * sextent, mycount, sdtype,
148-
bmtree->tree_next[i],
163+
int vchild = (bmtree->tree_next[i] - root + size) % size;
164+
int send_count = vchild - vrank;
165+
if (send_count > size - vchild)
166+
send_count = size - vchild;
167+
send_count *= scount;
168+
err = MCA_PML_CALL(send(ptmp + (ptrdiff_t)(curr_count - send_count) * sextent,
169+
send_count, sdtype, bmtree->tree_next[i],
149170
MCA_COLL_BASE_TAG_SCATTER,
150171
MCA_PML_BASE_SEND_STANDARD, comm));
151172
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
152-
153-
total_send += mycount;
173+
curr_count -= send_count;
154174
}
155-
156175
if (NULL != tempbuf)
157176
free(tempbuf);
158177
} else {

0 commit comments

Comments
 (0)