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.
333 строки
13 KiB
C
333 строки
13 KiB
C
#include "nbc.h"
|
|
static __inline__ int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, void *sendbuf, void *recvbuf, MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle);
|
|
static __inline__ int allred_sched_chain(int rank, int p, int count, MPI_Datatype datatype, void *sendbuf, void *recvbuf, MPI_Op op, int size, int ext, NBC_Schedule *schedule, NBC_Handle *handle, int fragsize);
|
|
|
|
#ifdef NBC_CACHE_SCHEDULE
|
|
/* tree comparison function for schedule cache */
|
|
int NBC_Allreduce_args_compare(NBC_Allreduce_args *a, NBC_Allreduce_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) ) {
|
|
return 0;
|
|
}
|
|
if( a->sendbuf < b->sendbuf ) {
|
|
return -1;
|
|
}
|
|
return +1;
|
|
}
|
|
#endif
|
|
|
|
int NBC_Iallreduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, NBC_Handle* handle) {
|
|
int rank, p, res, size, segsize;
|
|
MPI_Aint ext;
|
|
NBC_Schedule *schedule;
|
|
#ifdef NBC_CACHE_SCHEDULE
|
|
NBC_Allreduce_args *args, *found, search;
|
|
#endif
|
|
enum { NBC_ARED_BINOMIAL, NBC_ARED_CHAIN } alg;
|
|
char inplace;
|
|
|
|
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; }
|
|
|
|
handle->tmpbuf = malloc(ext*count);
|
|
if(handle->tmpbuf == NULL) { printf("Error in malloc() (%i)\n", res); return NBC_OOR; }
|
|
|
|
if((p == 1) && !inplace) {
|
|
/* for a single node - copy data to receivebuf */
|
|
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_ARED_BINOMIAL;
|
|
} else if(size*count < 262144){
|
|
alg = NBC_ARED_CHAIN;
|
|
segsize = 16384/2;
|
|
} else {
|
|
alg = NBC_ARED_CHAIN;
|
|
segsize = 65536/2;
|
|
}
|
|
|
|
#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;
|
|
found = hb_tree_search(handle->comminfo->NBC_Dict[NBC_ALLREDUCE], &search);
|
|
if(found == NULL) {
|
|
#endif
|
|
schedule = malloc(sizeof(NBC_Schedule));
|
|
if (NULL == schedule) { printf("Error in malloc()\n"); 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_ARED_BINOMIAL:
|
|
res = allred_sched_diss(rank, p, count, datatype, sendbuf, recvbuf, op, schedule, handle);
|
|
break;
|
|
case NBC_ARED_CHAIN:
|
|
res = allred_sched_chain(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, handle, segsize);
|
|
break;
|
|
}
|
|
if (NBC_OK != res) { printf("Error in Schedule creation() (%i)\n", res); return res; }
|
|
|
|
res = NBC_Sched_commit(schedule);
|
|
if(res != NBC_OK) { 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_Allreduce_args));
|
|
args->sendbuf=sendbuf;
|
|
args->recvbuf=recvbuf;
|
|
args->count=count;
|
|
args->datatype=datatype;
|
|
args->op=op;
|
|
args->schedule=schedule;
|
|
res = hb_tree_insert (handle->comminfo->NBC_Dict[NBC_ALLREDUCE], args, args, 0);
|
|
if(res != 0) printf("error in dict_insert() (%i)\n", res);
|
|
/* increase number of elements for A2A */
|
|
if(++handle->comminfo->NBC_Dict_size[NBC_ALLREDUCE] > SCHED_DICT_UPPER) {
|
|
NBC_SchedCache_dictwipe(handle->comminfo->NBC_Dict[NBC_ALLREDUCE], &handle->comminfo->NBC_Dict_size[NBC_ALLREDUCE]);
|
|
}
|
|
} else {
|
|
/* found schedule */
|
|
schedule=found->schedule;
|
|
}
|
|
#endif
|
|
|
|
res = NBC_Start(handle, schedule);
|
|
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
|
|
|
|
/* tmpbuf is freed with the handle */
|
|
return NBC_OK;
|
|
}
|
|
|
|
|
|
/* binomial allreduce (binomial tree up and binomial bcast down)
|
|
* 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
|
|
* followed by a Bcast:
|
|
* Algorithm:
|
|
* - each node with vrank > 2^r and vrank < 2^r+1 receives from node
|
|
* vrank - 2^r (vrank=1 receives from 0, vrank 0 receives never)
|
|
* - each node sends each round r to node vrank + 2^r
|
|
* - a node stops to send if 2^r > commsize
|
|
*
|
|
*/
|
|
#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 allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, void *sendbuf, void *recvbuf, MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle) {
|
|
int root, vrank, r, maxr, firstred, vpeer, peer, res;
|
|
|
|
root = 0; /* this makes the code for ireduce and iallreduce nearly identical - could be changed to improve performance */
|
|
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(res != NBC_OK) { 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(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
|
|
if(firstred) {
|
|
/* perform the reduce with the senbuf */
|
|
res = NBC_Sched_op(recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule);
|
|
firstred = 0;
|
|
} else {
|
|
/* perform the reduce in my local buffer */
|
|
res = NBC_Sched_op(recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule);
|
|
}
|
|
if(res != NBC_OK) { 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(res != NBC_OK) { 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 have to use the sendbuf in the first round .. */
|
|
res = NBC_Sched_send(sendbuf, false, count, datatype, peer, schedule);
|
|
} else {
|
|
/* and the recvbuf in all remeining rounds */
|
|
res = NBC_Sched_send(recvbuf, false, count, datatype, peer, schedule);
|
|
}
|
|
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
|
|
/* leave the game */
|
|
break;
|
|
}
|
|
}
|
|
|
|
/* this is the Bcast part - copied with minor changes from nbc_ibcast.c
|
|
* changed: buffer -> recvbuf */
|
|
RANK2VRANK(rank, vrank, root);
|
|
|
|
/* receive from the right hosts */
|
|
if(vrank != 0) {
|
|
for(r=0; r<maxr; r++) {
|
|
if((vrank >= (1<<r)) && (vrank < (1<<(r+1)))) {
|
|
VRANK2RANK(peer, vrank-(1<<r), root);
|
|
res = NBC_Sched_recv(recvbuf, false, count, datatype, peer, schedule);
|
|
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
|
|
}
|
|
}
|
|
res = NBC_Sched_barrier(schedule);
|
|
if(NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
|
|
}
|
|
|
|
/* now send to the right hosts */
|
|
for(r=0; r<maxr; r++) {
|
|
if(((vrank + (1<<r) < p) && (vrank < (1<<r))) || (vrank == 0)) {
|
|
VRANK2RANK(peer, vrank+(1<<r), root);
|
|
res = NBC_Sched_send(recvbuf, false, count, datatype, peer, schedule);
|
|
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
|
|
}
|
|
}
|
|
/* end of the bcast */
|
|
|
|
return NBC_OK;
|
|
}
|
|
|
|
static __inline__ int allred_sched_chain(int rank, int p, int count, MPI_Datatype datatype, void *sendbuf, void *recvbuf, MPI_Op op, int size, int ext, NBC_Schedule *schedule, NBC_Handle *handle, int fragsize) {
|
|
int res, rrpeer, rbpeer, srpeer, sbpeer, numfrag, fragnum, fragcount, thiscount, bstart, bend;
|
|
long roffset, boffset;
|
|
|
|
/* reduce peers */
|
|
rrpeer = rank+1;
|
|
srpeer = rank-1;
|
|
/* bcast peers */
|
|
rbpeer = rank-1;
|
|
sbpeer = rank+1;
|
|
|
|
if(count == 0) return NBC_OK;
|
|
|
|
numfrag = count*size/fragsize;
|
|
if((count*size)%fragsize != 0) numfrag++;
|
|
fragcount = count/numfrag;
|
|
|
|
/* determine the starting round of bcast ... the first reduced packet
|
|
* is after p-1 rounds at rank 0 and will be sent back ... */
|
|
bstart = p-1+rank;
|
|
/* determine the ending round of bcast ... after arrival of the first
|
|
* packet, each rank has to forward numfrag packets */
|
|
bend = bstart+numfrag;
|
|
//printf("[%i] numfrag: %i, count: %i, size: %i, fragcount: %i, bstart: %i, bend: %i\n", rank, numfrag, count, size, fragcount, bstart, bend);
|
|
|
|
/* this are two loops in one - this is a little nasty :-( */
|
|
for(fragnum = 0; fragnum < bend; fragnum++) {
|
|
roffset = fragnum*fragcount*ext;
|
|
boffset = (fragnum-bstart)*fragcount*ext;
|
|
thiscount = fragcount;
|
|
|
|
/* first numfrag rounds ... REDUCE to rank 0 */
|
|
if(fragnum < numfrag) {
|
|
if(fragnum == numfrag-1) {
|
|
/* last fragment may not be full */
|
|
thiscount = count-fragcount*fragnum;
|
|
}
|
|
//printf("[%i] reduce %i elements from %lu\n", rank, thiscount, roffset);
|
|
|
|
/* REDUCE - PART last node does not recv */
|
|
if(rank != p-1) {
|
|
res = NBC_Sched_recv((char*)roffset, true, thiscount, datatype, rrpeer, 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(rank == 0) {
|
|
res = NBC_Sched_op((char*)recvbuf+roffset, false, (char*)sendbuf+roffset, false, (char*)roffset, true, thiscount, datatype, op, schedule);
|
|
} else {
|
|
res = NBC_Sched_op((char*)roffset, true, (char*)sendbuf+roffset, false, (char*)roffset, true, thiscount, datatype, op, schedule);
|
|
}
|
|
res = NBC_Sched_barrier(schedule);
|
|
}
|
|
|
|
/* REDUCE PART root does not send */
|
|
if(rank != 0) {
|
|
/* rank p-1 has to send out of sendbuffer :) */
|
|
if(rank == p-1) {
|
|
res = NBC_Sched_send((char*)sendbuf+roffset, false, thiscount, datatype, srpeer, schedule);
|
|
} else {
|
|
res = NBC_Sched_send((char*)roffset, true, thiscount, datatype, srpeer, 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);
|
|
}
|
|
}
|
|
|
|
/* BCAST from rank 0 */
|
|
if(fragnum >= bstart) {
|
|
//printf("[%i] bcast %i elements from %lu\n", rank, thiscount, boffset);
|
|
if(fragnum == bend-1) {
|
|
/* last fragment may not be full */
|
|
thiscount = count-fragcount*(fragnum-bstart);
|
|
}
|
|
|
|
/* BCAST PART root does not receive */
|
|
if(rank != 0) {
|
|
res = NBC_Sched_recv((char*)recvbuf+boffset, false, thiscount, datatype, rbpeer, schedule);
|
|
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
|
|
res = NBC_Sched_barrier(schedule);
|
|
}
|
|
|
|
/* BCAST PART last rank does not send */
|
|
if(rank != p-1) {
|
|
res = NBC_Sched_send((char*)recvbuf+boffset, false, thiscount, datatype, sbpeer, schedule);
|
|
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
|
|
res = NBC_Sched_barrier(schedule);
|
|
}
|
|
}
|
|
}
|
|
|
|
//NBC_PRINT_SCHED(*schedule);
|
|
|
|
return NBC_OK;
|
|
}
|