6b22641669
I know it does not make much sense but one can play around with the performance. Numbers are available at http://www.unixer.de/research/nbcoll/perf/. This is the first step towards collv2. Next step includes the addition of non-blocking functions to the MPI-Layer and the collv1 interface. It implements all MPI-1 collective algorithms in a non-blocking manner. However, the collv1 interface does not allow non-blocking collectives so that all collectives are used blocking by the ompi-glue layer. I wanted to add LibNBC as a separate subdirectory, but I could not convince the buildsystem (and had not the time). So the component looks pretty messy. It would be great if somebody could explain me how to move all nbc*{c,h}, and {hb,dict}*{c,h} to a seperate subdirectory. It's .ompi_ignored because I did not test it exhaustively yet. This commit was SVN r11401.
278 строки
11 KiB
C
278 строки
11 KiB
C
#include "nbc.h"
|
|
static __inline__ int red_sched_binomial(int rank, int p, int root, void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, char *redbuf, NBC_Schedule *schedule, NBC_Handle *handle);
|
|
static __inline__ int red_sched_chain(int rank, int p, int root, void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int ext, int size, NBC_Schedule *schedule, NBC_Handle *handle, int fragsize);
|
|
|
|
#ifdef NBC_CACHE_SCHEDULE
|
|
/* tree comparison function for schedule cache */
|
|
int NBC_Reduce_args_compare(NBC_Reduce_args *a, NBC_Reduce_args *b, void *param) {
|
|
|
|
if( (a->sendbuf == b->sendbuf) &&
|
|
(a->recvbuf == b->recvbuf) &&
|
|
(a->count == b->count) &&
|
|
(a->datatype == b->datatype) &&
|
|
(a->op == b->op) &&
|
|
(a->root == b->root) ) {
|
|
return 0;
|
|
}
|
|
if( a->sendbuf < b->sendbuf ) {
|
|
return -1;
|
|
}
|
|
return +1;
|
|
}
|
|
#endif
|
|
|
|
/* the non-blocking reduce */
|
|
int NBC_Ireduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm, NBC_Handle* handle) {
|
|
int rank, p, res, segsize, size;
|
|
MPI_Aint ext;
|
|
NBC_Schedule *schedule;
|
|
char *redbuf=NULL, inplace;
|
|
#ifdef NBC_CACHE_SCHEDULE
|
|
NBC_Reduce_args *args, *found, search;
|
|
#endif
|
|
enum { NBC_RED_BINOMIAL, NBC_RED_CHAIN } alg;
|
|
|
|
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
|
|
|
|
res = NBC_Init_handle(handle, comm);
|
|
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
|
|
res = MPI_Comm_rank(comm, &rank);
|
|
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
|
|
res = MPI_Comm_size(comm, &p);
|
|
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
|
|
res = MPI_Type_extent(datatype, &ext);
|
|
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
|
|
res = MPI_Type_size(datatype, &size);
|
|
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
|
|
|
|
/* only one node -> copy data */
|
|
if((p == 1) && !inplace) {
|
|
res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
|
|
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
|
|
}
|
|
|
|
/* algorithm selection */
|
|
if(p < 4 || size*count < 65536) {
|
|
alg = NBC_RED_BINOMIAL;
|
|
if(rank == root) {
|
|
/* root reduces in receivebuffer */
|
|
handle->tmpbuf = malloc(ext*count);
|
|
} else {
|
|
/* recvbuf may not be valid on non-root nodes */
|
|
handle->tmpbuf = malloc(ext*count*2);
|
|
redbuf = ((char*)handle->tmpbuf)+(ext*count);
|
|
}
|
|
} else {
|
|
handle->tmpbuf = malloc(ext*count);
|
|
alg = NBC_RED_CHAIN;
|
|
segsize = 16384/2;
|
|
}
|
|
if (NULL == handle->tmpbuf) { printf("Error in malloc() (%i)\n", res); return res; }
|
|
|
|
#ifdef NBC_CACHE_SCHEDULE
|
|
/* search schedule in communicator specific tree */
|
|
search.sendbuf=sendbuf;
|
|
search.recvbuf=recvbuf;
|
|
search.count=count;
|
|
search.datatype=datatype;
|
|
search.op=op;
|
|
search.root=root;
|
|
found = hb_tree_search(handle->comminfo->NBC_Dict[NBC_REDUCE], &search);
|
|
if(found == NULL) {
|
|
#endif
|
|
schedule = malloc(sizeof(NBC_Schedule));
|
|
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
|
|
|
|
res = NBC_Sched_create(schedule);
|
|
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
|
|
|
|
switch(alg) {
|
|
case NBC_RED_BINOMIAL:
|
|
res = red_sched_binomial(rank, p, root, sendbuf, recvbuf, count, datatype, op, redbuf, schedule, handle);
|
|
break;
|
|
case NBC_RED_CHAIN:
|
|
res = red_sched_chain(rank, p, root, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, handle, segsize);
|
|
break;
|
|
}
|
|
if (NBC_OK != res) { printf("Error in Schedule creation() (%i)\n", res); return res; }
|
|
|
|
res = NBC_Sched_commit(schedule);
|
|
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
|
|
#ifdef NBC_CACHE_SCHEDULE
|
|
/* save schedule to tree */
|
|
args = malloc(sizeof(NBC_Alltoall_args));
|
|
args->sendbuf=sendbuf;
|
|
args->recvbuf=recvbuf;
|
|
args->count=count;
|
|
args->datatype=datatype;
|
|
args->op=op;
|
|
args->root=root;
|
|
args->schedule=schedule;
|
|
res = hb_tree_insert (handle->comminfo->NBC_Dict[NBC_REDUCE], args, args, 0);
|
|
if(res != 0) printf("error in dict_insert() (%i)\n", res);
|
|
/* increase number of elements for Reduce */
|
|
if(++handle->comminfo->NBC_Dict_size[NBC_REDUCE] > SCHED_DICT_UPPER) {
|
|
NBC_SchedCache_dictwipe(handle->comminfo->NBC_Dict[NBC_REDUCE], &handle->comminfo->NBC_Dict_size[NBC_REDUCE]);
|
|
}
|
|
} else {
|
|
/* found schedule */
|
|
schedule=found->schedule;
|
|
}
|
|
#endif
|
|
|
|
res = NBC_Start(handle, schedule);
|
|
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
|
|
|
|
/* tmpbuf is freed with the handle */
|
|
return NBC_OK;
|
|
}
|
|
|
|
|
|
/* binomial reduce
|
|
* working principle:
|
|
* - each node gets a virtual rank vrank
|
|
* - the 'root' node get vrank 0
|
|
* - node 0 gets the vrank of the 'root'
|
|
* - all other ranks stay identical (they do not matter)
|
|
*
|
|
* Algorithm:
|
|
* pairwise exchange
|
|
* round r:
|
|
* grp = rank % 2^r
|
|
* if grp == 0: receive from rank + 2^(r-1) if it exists and reduce value
|
|
* if grp == 1: send to rank - 2^(r-1) and exit function
|
|
*
|
|
* do this for R=log_2(p) rounds
|
|
*
|
|
*/
|
|
#define RANK2VRANK(rank, vrank, root) \
|
|
{ \
|
|
vrank = rank; \
|
|
if (rank == 0) vrank = root; \
|
|
if (rank == root) vrank = 0; \
|
|
}
|
|
#define VRANK2RANK(rank, vrank, root) \
|
|
{ \
|
|
rank = vrank; \
|
|
if (vrank == 0) rank = root; \
|
|
if (vrank == root) rank = 0; \
|
|
}
|
|
static __inline__ int red_sched_binomial(int rank, int p, int root, void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, char *redbuf, NBC_Schedule *schedule, NBC_Handle *handle) {
|
|
int firstred, vrank, vpeer, peer, res, maxr, r;
|
|
|
|
RANK2VRANK(rank, vrank, root);
|
|
maxr = (int)ceil((log(p)/LOG2));
|
|
|
|
firstred = 1;
|
|
for(r=1; r<=maxr; r++) {
|
|
if((vrank % (1<<r)) == 0) {
|
|
/* we have to receive this round */
|
|
vpeer = vrank + (1<<(r-1));
|
|
VRANK2RANK(peer, vpeer, root)
|
|
if(peer<p) {
|
|
res = NBC_Sched_recv(0, true, count, datatype, peer, schedule);
|
|
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
|
|
/* we have to wait until we have the data */
|
|
res = NBC_Sched_barrier(schedule);
|
|
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
|
|
/* perform the reduce in my local buffer */
|
|
if(firstred) {
|
|
if(rank == root) {
|
|
/* root is the only one who reduces in the receivebuffer
|
|
* take data from sendbuf in first round - save copy */
|
|
res = NBC_Sched_op(recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule);
|
|
} else {
|
|
/* all others may not have a receive buffer
|
|
* take data from sendbuf in first round - save copy */
|
|
res = NBC_Sched_op(redbuf-(unsigned long)handle->tmpbuf, true, sendbuf, false, 0, true, count, datatype, op, schedule);
|
|
}
|
|
firstred = 0;
|
|
} else {
|
|
if(rank == root) {
|
|
/* root is the only one who reduces in the receivebuffer */
|
|
res = NBC_Sched_op(recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule);
|
|
} else {
|
|
/* all others may not have a receive buffer */
|
|
res = NBC_Sched_op(redbuf-(unsigned long)handle->tmpbuf, true, redbuf-(unsigned long)handle->tmpbuf, true, 0, true, count, datatype, op, schedule);
|
|
}
|
|
}
|
|
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_op() (%i)\n", res); return res; }
|
|
/* this cannot be done until handle->tmpbuf is unused :-( */
|
|
res = NBC_Sched_barrier(schedule);
|
|
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
|
|
}
|
|
} else {
|
|
/* we have to send this round */
|
|
vpeer = vrank - (1<<(r-1));
|
|
VRANK2RANK(peer, vpeer, root)
|
|
if(firstred) {
|
|
/* we did not reduce anything */
|
|
res = NBC_Sched_send(sendbuf, false, count, datatype, peer, schedule);
|
|
} else {
|
|
/* we have to use the redbuf the root (which works in receivebuf) is never sending .. */
|
|
res = NBC_Sched_send(redbuf-(unsigned long)handle->tmpbuf, true, count, datatype, peer, schedule);
|
|
}
|
|
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
|
|
/* leave the game */
|
|
break;
|
|
}
|
|
}
|
|
|
|
return NBC_OK;
|
|
}
|
|
|
|
/* chain send ... */
|
|
static __inline__ int red_sched_chain(int rank, int p, int root, void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int ext, int size, NBC_Schedule *schedule, NBC_Handle *handle, int fragsize) {
|
|
int res, vrank, rpeer, speer, numfrag, fragnum, fragcount, thiscount;
|
|
long offset;
|
|
|
|
RANK2VRANK(rank, vrank, root);
|
|
VRANK2RANK(rpeer, vrank+1, root);
|
|
VRANK2RANK(speer, vrank-1, root);
|
|
|
|
if(count == 0) return NBC_OK;
|
|
|
|
numfrag = count*size/fragsize;
|
|
if((count*size)%fragsize != 0) numfrag++;
|
|
fragcount = count/numfrag;
|
|
/*printf("numfrag: %i, count: %i, size: %i, fragcount: %i\n", numfrag, count, size, fragcount);*/
|
|
|
|
for(fragnum = 0; fragnum < numfrag; fragnum++) {
|
|
offset = fragnum*fragcount*ext;
|
|
thiscount = fragcount;
|
|
if(fragnum == numfrag-1) {
|
|
/* last fragment may not be full */
|
|
thiscount = count-fragcount*fragnum;
|
|
}
|
|
|
|
/* last node does not recv */
|
|
if(vrank != p-1) {
|
|
res = NBC_Sched_recv((char*)offset, true, thiscount, datatype, rpeer, schedule);
|
|
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
|
|
res = NBC_Sched_barrier(schedule);
|
|
/* root reduces into receivebuf */
|
|
if(vrank == 0) {
|
|
res = NBC_Sched_op((char*)recvbuf+offset, false, (char*)sendbuf+offset, false, (char*)offset, true, thiscount, datatype, op, schedule);
|
|
} else {
|
|
res = NBC_Sched_op((char*)offset, true, (char*)sendbuf+offset, false, (char*)offset, true, thiscount, datatype, op, schedule);
|
|
}
|
|
res = NBC_Sched_barrier(schedule);
|
|
}
|
|
|
|
/* root does not send */
|
|
if(vrank != 0) {
|
|
/* rank p-1 has to send out of sendbuffer :) */
|
|
if(vrank == p-1) {
|
|
res = NBC_Sched_send((char*)sendbuf+offset, false, thiscount, datatype, speer, schedule);
|
|
} else {
|
|
res = NBC_Sched_send((char*)offset, true, thiscount, datatype, speer, schedule);
|
|
}
|
|
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
|
|
/* this barrier here seems awkward but isn't!!!! */
|
|
res = NBC_Sched_barrier(schedule);
|
|
}
|
|
}
|
|
|
|
return NBC_OK;
|
|
}
|