1
1
openmpi/ompi/mca/coll/libnbc/nbc_ialltoall.c
Torsten Hoefler 6b22641669 added LibNBC (http://www.unixer.de/NBC) as collv1 (blocking) component.
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.
2006-08-24 16:47:18 +00:00

301 строка
13 KiB
C

#include "nbc.h"
static __inline__ int a2a_sched_linear(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule *schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm);
static __inline__ int a2a_sched_pairwise(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule *schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm);
static __inline__ int a2a_sched_diss(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule* schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, NBC_Handle *handle);
#ifdef NBC_CACHE_SCHEDULE
/* tree comparison function for schedule cache */
int NBC_Alltoall_args_compare(NBC_Alltoall_args *a, NBC_Alltoall_args *b, void *param) {
if( (a->sendbuf == b->sendbuf) &&
(a->sendcount == b->sendcount) &&
(a->sendtype == b->sendtype) &&
(a->recvbuf == b->recvbuf) &&
(a->recvcount == b->recvcount) &&
(a->recvtype == b->recvtype) ) {
return 0;
}
if( a->sendbuf < b->sendbuf ) {
return -1;
}
return +1;
}
#endif
/* simple linear MPI_Ialltoall the (simple) algorithm just sends to all nodes */
int NBC_Ialltoall(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, NBC_Handle *handle) {
int rank, p, res, a2asize, sndsize, datasize;
NBC_Schedule *schedule;
MPI_Aint rcvext, sndext;
#ifdef NBC_CACHE_SCHEDULE
NBC_Alltoall_args *args, *found, search;
#endif
char *rbuf, *sbuf, inplace;
enum {NBC_A2A_LINEAR, NBC_A2A_PAIRWISE, NBC_A2A_DISS} 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(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_size(sendtype, &sndsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
/* algorithm selection */
a2asize = sndsize*sendcount*p;
/* this number is optimized for TCP on odin.cs.indiana.edu */
if((p <= 8) && ((a2asize < 1<<17) || (sndsize*sendcount < 1<<12))) {
/* just send as fast as we can if we have less than 8 peers, if the
* total communicated size is smaller than 1<<17 *and* if we don't
* have eager messages (msgsize < 1<<13) */
alg = NBC_A2A_LINEAR;
} else if(a2asize < (1<<12)*p) {
alg = NBC_A2A_DISS;
} else
alg = NBC_A2A_LINEAR; /*NBC_A2A_PAIRWISE*/;
if(!inplace) {
/* copy my data to receive buffer */
rbuf = ((char *)recvbuf) + (rank*recvcount*rcvext);
sbuf = ((char *)sendbuf) + (rank*sendcount*sndext);
res = NBC_Copy(sbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* allocate temp buffer if we need one */
if(alg == NBC_A2A_DISS) {
/* only A2A_DISS needs buffers */
if(NBC_Type_intrinsic(sendtype)) {
datasize = sndext*sendcount;
} else {
res = MPI_Pack_size(sendcount, sendtype, comm, &datasize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack_size() (%i)\n", res); return res; }
}
/* allocate temporary buffers */
if(p % 2 == 0) {
handle->tmpbuf=malloc(datasize*p*2);
} else {
/* we cannot divide p by two, so alloc more to be safe ... */
handle->tmpbuf=malloc(datasize*(p/2+1)*2*2);
}
/* phase 1 - rotate n data blocks upwards into the tmpbuffer */
if(NBC_Type_intrinsic(sendtype)) {
/* contiguous - just copy (1st copy) */
memcpy(handle->tmpbuf, (char*)sendbuf+datasize*rank, datasize*(p-rank));
if(rank != 0) memcpy((char*)handle->tmpbuf+datasize*(p-rank), sendbuf, datasize*(rank));
} else {
int pos=0;
/* non-contiguous - pack */
res = MPI_Pack((char*)sendbuf+rank*sendcount*sndext, (p-rank)*sendcount, sendtype, handle->tmpbuf, (p-rank)*datasize, &pos, comm);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack() (%i)\n", res); return res; }
if(rank != 0) {
pos = 0;
MPI_Pack(sendbuf, rank*sendcount, sendtype, (char*)handle->tmpbuf+datasize*(p-rank), rank*datasize, &pos, comm);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack() (%i)\n", res); return res; }
}
}
} else {
handle->tmpbuf=NULL;
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.sendcount=sendcount;
search.sendtype=sendtype;
search.recvbuf=recvbuf;
search.recvcount=recvcount;
search.recvtype=recvtype;
found = hb_tree_search(handle->comminfo->NBC_Dict[NBC_ALLTOALL], &search);
if(found == NULL) {
#endif
/* not found - generate new schedule */
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_A2A_LINEAR:
res = a2a_sched_linear(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
break;
case NBC_A2A_DISS:
res = a2a_sched_diss(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, handle);
break;
case NBC_A2A_PAIRWISE:
res = a2a_sched_pairwise(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
break;
}
if (NBC_OK != res) { return res; }
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { 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->sendcount=sendcount;
args->sendtype=sendtype;
args->recvbuf=recvbuf;
args->recvcount=recvcount;
args->recvtype=recvtype;
args->schedule=schedule;
res = hb_tree_insert (handle->comminfo->NBC_Dict[NBC_ALLTOALL], 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_ALLTOALL] > SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe(handle->comminfo->NBC_Dict[NBC_ALLTOALL], &handle->comminfo->NBC_Dict_size[NBC_ALLTOALL]);
/*if(!rank) printf("[%i] removing %i elements - new size: %i \n", rank, SCHED_DICT_UPPER-SCHED_DICT_LOWER, handle->comminfo->NBC_Alltoall_size);*/
}
/*if(!rank) printf("[%i] added new schedule to tree - number %i\n", rank, handle->comminfo->NBC_Dict_size[NBC_ALLTOALL]);*/
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
static __inline__ int a2a_sched_pairwise(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule* schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
int res, r, sndpeer, rcvpeer;
char *rbuf, *sbuf;
res = NBC_OK;
if(p < 2) return res;
for(r=1;r<p;r++) {
sndpeer = (rank+r)%p;
rcvpeer = (rank-r+p)%p;
rbuf = ((char *) recvbuf) + (rcvpeer*recvcount*rcvext);
res = NBC_Sched_recv(rbuf, false, recvcount, recvtype, rcvpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
sbuf = ((char *) sendbuf) + (sndpeer*sendcount*sndext);
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, sndpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
if (r < p) {
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barr() (%i)\n", res); return res; }
}
}
return res;
}
static __inline__ int a2a_sched_linear(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule* schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
int res, r;
char *rbuf, *sbuf;
res = NBC_OK;
for(r=0;r<p;r++) {
/* easy algorithm */
if ((r == rank)) { continue; }
rbuf = ((char *) recvbuf) + (r*recvcount*rcvext);
res = NBC_Sched_recv(rbuf, false, recvcount, recvtype, r, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
sbuf = ((char *) sendbuf) + (r*sendcount*sndext);
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, r, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
return res;
}
static __inline__ int a2a_sched_diss(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule* schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, NBC_Handle *handle) {
int res, i, r, speer, rpeer, pos, datasize, offset, virtp;
char *rbuf, *rtmpbuf, *stmpbuf;
res = NBC_OK;
if(p < 2) return res;
if(NBC_Type_intrinsic(sendtype)) {
datasize = sndext*sendcount;
} else {
res = MPI_Pack_size(sendcount, sendtype, comm, &datasize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack_size() (%i)\n", res); return res; }
}
/* allocate temporary buffers */
if(p % 2 == 0) {
rtmpbuf = (char*)handle->tmpbuf+datasize*p;
stmpbuf = (char*)handle->tmpbuf+datasize*(p+p/2);
} else {
/* we cannot divide p by two, so alloc more to be safe ... */
virtp = (p/2+1)*2;
rtmpbuf = (char*)handle->tmpbuf+datasize*p;
stmpbuf = (char*)handle->tmpbuf+datasize*(p+virtp/2);
}
/* phase 2 - communicate */
/*printf("[%i] temp buffer is at %lu of size %i, maxround: %i\n", rank, (unsigned long)handle->tmpbuf, (int)datasize*p*(1<<maxround), maxround);*/
for(r = 1; r < p; r<<=1) {
offset = 0;
for(i=1; i<p; i++) {
/* test if bit r is set in rank number i */
if((i&r) == r) {
/* copy data to sendbuffer (2nd copy) - could be avoided using iovecs */
/*printf("[%i] round %i: copying element %i to buffer %lu\n", rank, r, i, (unsigned long)(stmpbuf+offset));*/
NBC_Sched_copy((void*)(long)(i*datasize), true, datasize, MPI_BYTE, stmpbuf+offset-(unsigned long)handle->tmpbuf, true, datasize, MPI_BYTE, schedule);
offset += datasize;
}
}
speer = ( rank + r) % p;
/* add p because modulo does not work with negative values */
rpeer = ((rank - r)+p) % p;
/*printf("[%i] receiving %i bytes from host %i into rbuf %lu\n", rank, offset, rpeer, (unsigned long)rtmpbuf);*/
res = NBC_Sched_recv(rtmpbuf-(unsigned long)handle->tmpbuf, true, offset, MPI_BYTE, rpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
/*printf("[%i] sending %i bytes to host %i from sbuf %lu\n", rank, offset, speer, (unsigned long)stmpbuf);*/
res = NBC_Sched_send(stmpbuf-(unsigned long)handle->tmpbuf, true, offset, MPI_BYTE, speer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
/* unpack from buffer */
offset = 0;
for(i=1; i<p; i++) {
/* test if bit r is set in rank number i */
if((i&r) == r) {
/* copy data to tmpbuffer (3rd copy) - could be avoided using iovecs */
NBC_Sched_copy(rtmpbuf+offset-(unsigned long)handle->tmpbuf, true, datasize, MPI_BYTE, (void*)(long)(i*datasize), true, datasize, MPI_BYTE, schedule);
offset += datasize;
}
}
}
/* phase 3 - reorder - data is now in wrong order in handle->tmpbuf -
* reorder it into recvbuf */
for(i=0; i<p; i++) {
rbuf = (char*)recvbuf+((rank-i+p)%p)*recvcount*rcvext;
pos = 0;
res = NBC_Sched_unpack((void*)(long)(i*datasize), true, recvcount, recvtype, rbuf, false, schedule);
if (NBC_OK != res) { printf("MPI Error in NBC_Sched_pack() (%i)\n", res); return res; }
}
return NBC_OK;
}