CmiCommHandle CmiAsyncBroadcastFn(int size, char *msg) {
#if ENSURE_MSG_PAIRORDER
/* Not sure how to add the msg seq no for async broadcast messages --Chao Mei */
/* so abort here ! */
CmiAssert(0);
return 0;
#else
int i, rank;
int mype = CmiMyPe();
#if ENABLE_CONVERSE_QD
CQdCreate(CpvAccess(cQdState), CmiNumPes()-1);
#endif
MACHSTATE1(3,"[%d] Sending async broadcast message from {",CmiMyNode());
CMI_BROADCAST_ROOT(msg) = 0;
void *handle = malloc(sizeof(int));
*((int *)handle) = CmiNumPes()-1;
for (i=mype+1; i<CmiNumPes(); i++) {
CMI_DEST_RANK(msg) = CmiRankOf(i);
lapiSendFn(CmiNodeOf(i), size, msg, DeliveredMsg, handle);
}
for (i=0; i<mype; i++) {
CMI_DEST_RANK(msg) = CmiRankOf(i);
lapiSendFn(CmiNodeOf(i), size, msg, DeliveredMsg, handle);
}
MACHSTATE(3,"} Sending async broadcast message end");
return handle;
#endif
}
void CldEnqueue(int pe, void *msg, int infofn)
{
int len, queueing, priobits; unsigned int *prioptr;
CldInfoFn ifn; CldPackFn pfn;
peinfo *pinf = &(CpvAccess(peinf));
ifn = (CldInfoFn)CmiHandlerToFunction(infofn);
ifn(msg, &pfn, &len, &queueing, &priobits, &prioptr);
if (pe != CLD_ANYWHERE) {
if (pfn && (CmiNodeOf(pe) != CmiMyNode())) {
pfn(&msg);
ifn(msg, &pfn, &len, &queueing, &priobits, &prioptr);
}
CmiSetInfo(msg, infofn);
CmiSetXHandler(msg, CmiGetHandler(msg));
CmiSetHandler(msg, pinf->EnqueueHandler);
if (pe==CLD_BROADCAST) CmiSyncBroadcastAndFree(len, msg);
else if (pe==CLD_BROADCAST_ALL) CmiSyncBroadcastAllAndFree(len, msg);
else CmiSyncSendAndFree(pe, len, msg);
} else {
CmiSetInfo(msg, infofn);
CmiSetXHandler(msg, CmiGetHandler(msg));
CmiSetHandler(msg, pinf->HopHandler);
CsdEnqueueGeneral(msg, queueing, priobits, prioptr);
}
}
void BGQTorusManager::populateLocalNodes() {
if(CmiNumPartitions() == 1) return;
CmiLock(bgq_lock);
if(bgq_isLocalSet) {
CmiUnlock(bgq_lock);
return;
}
if(bgq_localNodes == NULL)
bgq_localNodes = (int *)malloc(CmiNumNodesGlobal()*sizeof(int));
CmiAssert(bgq_localNodes != NULL);
for(int i = 0; i < CmiNumNodesGlobal(); i++)
bgq_localNodes[i] = -1;
for(int i = 0; i < CmiNumNodes(); i++) {
int a, b, c, d, e, t;
int global;
rankToCoordinates(CmiNodeFirst(i), a, b, c, d, e, t);
global = CmiNodeOf(coordinatesToRank(a, b, c, d, e, t));
bgq_localNodes[global] = i;
}
bgq_isLocalSet = 1;
CmiUnlock(bgq_lock);
}
/** \function pidtonid
* finds nids for pids 1 to CmiNumPes and stores them in an array
* correspondingly also creates an array for nids to pids
*/
void pidtonid(int numpes) {
CmiLock(cray_lock);
if (pid2nid != NULL) {
CmiUnlock(cray_lock);
return; /* did once already */
}
getDimension(&maxNID,&maxX,&maxY,&maxZ);
int numCores = CmiNumCores();
pid2nid = (int *)malloc(sizeof(int) * numpes);
#if XT4_TOPOLOGY || XT5_TOPOLOGY || XE6_TOPOLOGY
int i, nid, ret;
CmiAssert(rca_coords == NULL);
rca_coords = (rca_mesh_coord_t *)malloc(sizeof(rca_mesh_coord_t)*(maxNID+1));
for (i=0; i<maxNID; i++) {
rca_coords[i].mesh_x = rca_coords[i].mesh_y = rca_coords[i].mesh_z = -1;
}
for (i=0; i<numpes; i++) {
PMI_Get_nid(CmiGetNodeGlobal(CmiNodeOf(i),CmiMyPartition()), &nid);
pid2nid[i] = nid;
CmiAssert(nid < maxNID);
ret = rca_get_meshcoord(nid, &rca_coords[nid]);
CmiAssert(ret != -1);
}
#endif
CmiUnlock(cray_lock);
}
/* send msg along the hypercube in broadcast. (Sameer) */
static void SendHyperCubeProc(int size, char *msg) {
int startpe = CMI_BROADCAST_ROOT(msg)-1;
int startnode = CmiNodeOf(startpe);
#if CMK_SMP
if (startpe > CmiNumPes()) startnode = startpe - CmiNumPes();
#endif
SendHyperCube(size, msg, 0, startnode);
#if CMK_SMP
/* second send msgs to my peers on this node */
SendToPeers(size, msg);
#endif
}
void CmiDirect_manytomany_initialize_send ( void * h,
unsigned tag,
unsigned idx,
unsigned displ,
unsigned bytes,
unsigned rank )
{
BGPCmiDirectM2mHandle *handle = (BGPCmiDirectM2mHandle *) h;
assert ( tag < MAX_CONN );
handle->m2m_sndlens [tag][idx] = bytes;
handle->m2m_sdispls [tag][idx] = displ;
handle->m2m_ranks [tag][idx] = CmiGetNodeGlobal(CmiNodeOf(rank),CmiMyPartition());
handle->m2m_permutation[tag][idx] = (idx+1)%handle->m2m_nsndranks[tag];
}
void CldHopHandler(char *msg)
{
peinfo *pinf = &(CpvAccess(peinf));
int len, queueing, priobits; unsigned int *prioptr;
CldInfoFn ifn; CldPackFn pfn; int pe;
if (pinf->rebalance) {
/* do pe = ((lrand48()&0x7FFFFFFF)%CmiNumPes()); */
do pe = ((CrnRand()&0x7FFFFFFF)%CmiNumPes());
while (pe == pinf->mype);
ifn = (CldInfoFn)CmiHandlerToFunction(CmiGetInfo(msg));
ifn(msg, &pfn, &len, &queueing, &priobits, &prioptr);
if (pfn && CmiNodeOf(pe) != CmiMyNode()) {
pfn(&msg);
ifn(msg, &pfn, &len, &queueing, &priobits, &prioptr);
}
CmiSyncSendAndFree(pe, len, msg);
pinf->rebalance--;
} else {
CmiSetHandler(msg, CmiGetXHandler(msg));
CmiHandleMessage(msg);
}
}
void CmiInitCPUAffinity(char **argv)
{
static skt_ip_t myip;
int ret, i, exclude;
hostnameMsg *msg;
char *pemap = NULL;
char *commap = NULL;
char *pemapfile = NULL;
int show_affinity_flag;
int affinity_flag = CmiGetArgFlagDesc(argv,"+setcpuaffinity",
"set cpu affinity");
while (CmiGetArgIntDesc(argv,"+excludecore", &exclude, "avoid core when setting cpuaffinity")) {
if (CmiMyRank() == 0) add_exclude(exclude);
affinity_flag = 1;
}
if (CmiGetArgStringDesc(argv, "+pemapfile", &pemapfile, "define pe to core mapping file")) {
FILE *fp;
char buf[128];
pemap = (char*)malloc(1024);
fp = fopen(pemapfile, "r");
if (fp == NULL) CmiAbort("pemapfile does not exist");
while (!feof(fp)) {
if (fgets(buf, 128, fp)) {
if (buf[strlen(buf)-1] == '\n') buf[strlen(buf)-1] = 0;
strcat(pemap, buf);
}
}
fclose(fp);
if (CmiMyPe()==0) CmiPrintf("Charm++> read from pemap file '%s': %s\n", pemapfile, pemap);
}
CmiGetArgStringDesc(argv, "+pemap", &pemap, "define pe to core mapping");
if (pemap!=NULL && excludecount>0)
CmiAbort("Charm++> +pemap can not be used with +excludecore.\n");
CmiGetArgStringDesc(argv, "+commap", &commap, "define comm threads to core mapping");
if (pemap!=NULL || commap!=NULL) affinity_flag = 1;
show_affinity_flag = CmiGetArgFlagDesc(argv,"+showcpuaffinity",
"print cpu affinity");
cpuAffinityHandlerIdx =
CmiRegisterHandler((CmiHandler)cpuAffinityHandler);
cpuAffinityRecvHandlerIdx =
CmiRegisterHandler((CmiHandler)cpuAffinityRecvHandler);
if (CmiMyRank() ==0) {
affLock = CmiCreateLock();
}
#if CMK_BLUEGENEP || CMK_BLUEGENEQ
if(affinity_flag){
affinity_flag = 0;
if(CmiMyPe()==0) CmiPrintf("Charm++> cpu affinity setting is not needed on Blue Gene, thus ignored.\n");
}
if(show_affinity_flag){
show_affinity_flag = 0;
if(CmiMyPe()==0) CmiPrintf("Charm++> printing cpu affinity is not supported on Blue Gene.\n");
}
#endif
if (!affinity_flag) {
if (show_affinity_flag) CmiPrintCPUAffinity();
return;
}
if (CmiMyPe() == 0) {
CmiPrintf("Charm++> cpu affinity enabled. \n");
if (excludecount > 0) {
CmiPrintf("Charm++> cpuaffinity excludes core: %d", excludecore[0]);
for (i=1; i<excludecount; i++) CmiPrintf(" %d", excludecore[i]);
CmiPrintf(".\n");
}
if (pemap!=NULL)
CmiPrintf("Charm++> cpuaffinity PE-core map : %s\n", pemap);
}
if (CmiMyPe() >= CmiNumPes()) { /* this is comm thread */
/* comm thread either can float around, or pin down to the last rank.
however it seems to be reportedly slower if it is floating */
CmiNodeAllBarrier();
if (commap != NULL) {
int mycore = search_pemap(commap, CmiMyPeGlobal()-CmiNumPesGlobal());
if(CmiMyPe()-CmiNumPes()==0) printf("Charm++> set comm %d on node %d to core #%d\n", CmiMyPe()-CmiNumPes(), CmiMyNode(), mycore);
if (-1 == CmiSetCPUAffinity(mycore))
CmiAbort("set_cpu_affinity abort!");
CmiNodeAllBarrier();
if (show_affinity_flag) CmiPrintCPUAffinity();
return; /* comm thread return */
}
else {
/* if (CmiSetCPUAffinity(CmiNumCores()-1) == -1) CmiAbort("set_cpu_affinity abort!"); */
#if !CMK_CRAYXT && !CMK_CRAYXE && !CMK_CRAYXC && !CMK_BLUEGENEQ
if (pemap == NULL) {
#if CMK_MACHINE_PROGRESS_DEFINED
while (affinity_doneflag < CmiMyNodeSize()) CmiNetworkProgress();
#else
#if CMK_SMP
#error "Machine progress call needs to be implemented for cpu affinity!"
#endif
#endif
}
#endif
#if CMK_CRAYXT || CMK_CRAYXE || CMK_CRAYXC
/* if both pemap and commmap are NULL, will compute one */
if (pemap != NULL)
#endif
{
CmiNodeAllBarrier();
if (show_affinity_flag) CmiPrintCPUAffinity();
return; /* comm thread return */
}
}
}
if (pemap != NULL && CmiMyPe()<CmiNumPes()) { /* work thread */
int mycore = search_pemap(pemap, CmiMyPeGlobal());
if(show_affinity_flag) CmiPrintf("Charm++> set PE %d on node %d to core #%d\n", CmiMyPe(), CmiMyNode(), mycore);
if (mycore >= CmiNumCores()) {
CmiPrintf("Error> Invalid core number %d, only have %d cores (0-%d) on the node. \n", mycore, CmiNumCores(), CmiNumCores()-1);
CmiAbort("Invalid core number");
}
if (CmiSetCPUAffinity(mycore) == -1) CmiAbort("set_cpu_affinity abort!");
CmiNodeAllBarrier();
CmiNodeAllBarrier();
/* if (show_affinity_flag) CmiPrintCPUAffinity(); */
return;
}
#if CMK_CRAYXT || CMK_CRAYXE || CMK_CRAYXC
{
int numCores = CmiNumCores();
int myid = getXTNodeID(CmiMyNodeGlobal(), CmiNumNodesGlobal());
int myrank;
int pe, mype = CmiMyPeGlobal();
int node = CmiMyNodeGlobal();
int nnodes = 0;
#if CMK_SMP
if (CmiMyPe() >= CmiNumPes()) { /* this is comm thread */
int node = CmiMyPe() - CmiNumPes();
mype = CmiGetPeGlobal(CmiNodeFirst(node) + CmiMyNodeSize() - 1, CmiMyPartition()); /* last pe on SMP node */
node = CmiGetNodeGlobal(node, CmiMyPartition());
}
#endif
pe = mype - 1;
while (pe >= 0) {
int n = CmiNodeOf(pe);
if (n != node) { nnodes++; node = n; }
if (getXTNodeID(n, CmiNumNodesGlobal()) != myid) break;
pe --;
}
CmiAssert(numCores > 0);
myrank = (mype - pe - 1 + nnodes)%numCores;
#if CMK_SMP
if (CmiMyPe() >= CmiNumPes())
myrank = (myrank + 1)%numCores;
#endif
if (-1 != CmiSetCPUAffinity(myrank)) {
DEBUGP(("Processor %d is bound to core #%d on node #%d\n", CmiMyPe(), myrank, mynode));
}
else{
CmiPrintf("Processor %d set affinity failed!\n", CmiMyPe());
CmiAbort("set cpu affinity abort!\n");
}
}
if (CmiMyPe() < CmiNumPes())
CmiNodeAllBarrier();
CmiNodeAllBarrier();
#else
/* get my ip address */
if (CmiMyRank() == 0)
{
#if CMK_HAS_GETHOSTNAME
myip = skt_my_ip(); /* not thread safe, so only calls on rank 0 */
#else
CmiAbort("Can not get unique name for the compute nodes. \n");
#endif
}
CmiNodeAllBarrier();
/* prepare a msg to send */
msg = (hostnameMsg *)CmiAlloc(sizeof(hostnameMsg));
CmiSetHandler((char *)msg, cpuAffinityHandlerIdx);
msg->pe = CmiMyPe();
msg->ip = myip;
msg->ncores = CmiNumCores();
DEBUGP(("PE %d's node has %d number of cores. \n", CmiMyPe(), msg->ncores));
msg->rank = 0;
CmiSyncSendAndFree(0, sizeof(hostnameMsg), (void *)msg);
if (CmiMyPe() == 0) {
int i;
hostTable = CmmNew();
rankmsg = (rankMsg *)CmiAlloc(sizeof(rankMsg)+CmiNumPes()*sizeof(int)*2);
CmiSetHandler((char *)rankmsg, cpuAffinityRecvHandlerIdx);
rankmsg->ranks = (int *)((char*)rankmsg + sizeof(rankMsg));
rankmsg->nodes = (int *)((char*)rankmsg + sizeof(rankMsg) + CmiNumPes()*sizeof(int));
for (i=0; i<CmiNumPes(); i++) {
rankmsg->ranks[i] = 0;
rankmsg->nodes[i] = -1;
}
for (i=0; i<CmiNumPes(); i++) CmiDeliverSpecificMsg(cpuAffinityHandlerIdx);
}
/* receive broadcast from PE 0 */
CmiDeliverSpecificMsg(cpuAffinityRecvHandlerIdx);
CmiLock(affLock);
affinity_doneflag++;
CmiUnlock(affLock);
CmiNodeAllBarrier();
#endif
if (show_affinity_flag) CmiPrintCPUAffinity();
}
void TorusLB::strategy() {
int index;
// compute the average load by (compute load + background load) / numPesAvailable
computeAverage();
// two heaps of self and pair computes
makeTwoHeaps();
const int beginGroup = processors[0].Id;
const int endGroup = beginGroup + P;
#define INGROUP(PROC) ((PROC) >= beginGroup && (PROC) < endGroup)
computeInfo *c;
processorInfo *p, *minp;
Iterator nextP;
overLoad = 1.2;
for(int I=0; I<numComputes; I++) {
c = (computeInfo *) computePairHeap->deleteMax();
if ( ! c ) c = (computeInfo *) computeSelfHeap->deleteMax();
if(c->processor != -1) continue; // go to the next compute
if(!c) CkAbort("TorusLB: Compute Heap empty!\n");
for(int j=0; j<6; j++) {
bestPe[j] = 0;
goodPe[j] = 0;
badPe[j] = 0;
}
// Look at pes which have the compute's patches
// HYBRID check if processor is in local group
#define SELECT_REALPE(X) if INGROUP((X)) { \
selectPes(&processors[(X) - beginGroup], c); \
}
const int realPe1 = patches[c->patch1].processor;
SELECT_REALPE(realPe1)
const int realPe2 = patches[c->patch2].processor;
if ( realPe2 != realPe1 ) {
SELECT_REALPE(realPe2)
}
// Try the processors which have the patches' proxies
p = (processorInfo *)(patches[c->patch1].proxiesOn.iterator((Iterator *)&nextP));
while(p) { // patch 1
if INGROUP(p->Id) selectPes(p, c);
p = (processorInfo *)(patches[c->patch1].proxiesOn.next((Iterator *)&nextP));
}
p = (processorInfo *)(patches[c->patch2].proxiesOn.iterator((Iterator *)&nextP));
while(p) { // patch 2
if INGROUP(p->Id) selectPes(p, c);
p = (processorInfo *)(patches[c->patch2].proxiesOn.next((Iterator *)&nextP));
}
// see if we have found a processor to place the compute on
p = 0;
if((p = bestPe[5])
#if USE_TOPOMAP
|| (p = goodPe[5])
#endif
|| (p = bestPe[4])
#if USE_TOPOMAP
|| (p = goodPe[4])
#endif
|| (p = bestPe[3])
#if USE_TOPOMAP
|| (p = goodPe[3])
#endif
|| (p = bestPe[1])
#if USE_TOPOMAP
|| (p = goodPe[1])
#endif
|| (p = bestPe[2])
#if USE_TOPOMAP
|| (p = goodPe[2])
#endif
|| (p = bestPe[0])
#if USE_TOPOMAP
|| (p = goodPe[0])
#endif
) {
assign(c, p);
continue;
}
// Try all pes on the nodes of the home patches
if ( CmiNumNodes() > 1 ) { // else not useful
double minLoad = overLoad * averageLoad;
minp = 0;
int realNode1 = CmiNodeOf(realPe1);
int nodeSize = CmiNodeSize(realNode1);
if ( nodeSize > 1 ) { // else did it already
int firstpe = CmiNodeFirst(realNode1);
for ( int rpe = firstpe; rpe < firstpe+nodeSize; ++rpe ) {
if INGROUP(rpe) {
p = &processors[rpe - beginGroup];
if ( p->available && ( p->load + c->load < minLoad ) ) {
minLoad = p->load + c->load;
minp = p;
}
}
}
}
if ( realPe2 != realPe1 ) {
int realNode2 = CmiNodeOf(realPe2);
if ( realNode2 != realNode1 ) { // else did it already
nodeSize = CmiNodeSize(realNode2);
if ( nodeSize > 1 ) {
int firstpe = CmiNodeFirst(realNode2);
for ( int rpe = firstpe; rpe < firstpe+nodeSize; ++rpe ) {
if INGROUP(rpe) {
p = &processors[rpe - beginGroup];
if ( p->available && ( p->load + c->load < minLoad ) ) {
minLoad = p->load + c->load;
minp = p;
}
}
}
}
