static void SendSpanningChildren(int size, char *msg, int rankToAssign, int startNode) {
#if CMK_BROADCAST_SPANNING_TREE
int i, oldRank;
char *newmsg;
oldRank = CMI_DEST_RANK(msg);
/* doing this is to avoid the multiple assignment in the following for loop */
CMI_DEST_RANK(msg) = rankToAssign;
/* first send msgs to other nodes */
CmiAssert(startNode >=0 && startNode<CmiNumNodes());
for (i=1; i<=BROADCAST_SPANNING_FACTOR; i++) {
int nd = CmiMyNode()-startNode;
if (nd<0) nd+=CmiNumNodes();
nd = BROADCAST_SPANNING_FACTOR*nd + i;
if (nd > CmiNumNodes() - 1) break;
nd += startNode;
nd = nd%CmiNumNodes();
CmiAssert(nd>=0 && nd!=CmiMyNode());
#if CMK_BROADCAST_USE_CMIREFERENCE
CmiReference(msg);
CmiSendNetworkFunc(CmiNodeFirst(nd), size, msg, BCAST_SYNC);
#else
newmsg = CopyMsg(msg, size);
CmiSendNetworkFunc(CmiNodeFirst(nd), size, newmsg, BCAST_SYNC);
#endif
}
CMI_DEST_RANK(msg) = oldRank;
#endif
}
void HbmLB::ReceiveStats(double t, int frompe, int fromlevel)
{
#if CMK_LBDB_ON
FindNeighbors();
int atlevel = fromlevel + 1;
CmiAssert(tree->isroot(CkMyPe(), atlevel));
DEBUGF(("[%d] ReceiveStats from PE %d from level: %d\n", CkMyPe(), frompe, fromlevel));
int neighborIdx = NeighborIndex(frompe, atlevel);
CmiAssert(neighborIdx==0 || neighborIdx==1);
LevelData *lData = levelData[atlevel];
lData->statsList[neighborIdx] = t;
int &stats_msg_count = levelData[atlevel]->stats_msg_count;
stats_msg_count ++;
DEBUGF(("[%d] ReceiveStats at level: %d %d/%d\n", CkMyPe(), atlevel, stats_msg_count, levelData[atlevel]->nChildren));
if (stats_msg_count == levelData[atlevel]->nChildren)
{
stats_msg_count = 0;
int parent = levelData[atlevel]->parent;
// load balancing
thisProxy[CkMyPe()].Loadbalancing(atlevel);
}
#endif
}
void setupSharedBuffers(){
int i=0;
CmiAssert((sysvshmContext->recvbufnames=(int *)malloc(sizeof(int)*(sysvshmContext->nodesize)))!=NULL);
CmiAssert((sysvshmContext->sendbufnames=(int *)malloc(sizeof(int)*(sysvshmContext->nodesize)))!=NULL);
CmiAssert((sysvshmContext->semarray= (ushort *)malloc(sizeof(int)*(sysvshmContext->nodesize)))!=NULL);
for(i=0;i<sysvshmContext->nodesize;i++){
sysvshmContext->semarray[i]=1;
if(i != sysvshmContext->noderank){
sysvshmContext->sendbufnames[i]=SHMSNDNAME(i);
sysvshmContext->recvbufnames[i]=SHMRCVNAME(i);
}
}
createShmObjectsAndSems(&(sysvshmContext->recvBufs),sysvshmContext->recvbufnames,0);
createShmObjectsAndSems(&(sysvshmContext->sendBufs),sysvshmContext->sendbufnames,1);
for(i=0;i<sysvshmContext->nodesize;i++){
if(i != sysvshmContext->noderank){
CmiAssert(sysvshmContext->sendBufs[i].header->count == 0);
sysvshmContext->sendBufs[i].header->count = 0;
sysvshmContext->sendBufs[i].header->bytes = 0;
}
}
}
/// Delete event and reconnect surrounding events in queue
void eventQueue::DeleteEvent(Event *ev)
{
#ifdef EQ_SANITIZE
sanitize();
#endif
Event *tmp;
CmiAssert(ev != currentPtr);
CmiAssert(ev->spawnedList == NULL);
CmiAssert(ev != frontPtr);
CmiAssert(ev != backPtr);
// if ev is earliest straggler, see if there is another
if (RBevent == ev) {
RBevent = NULL;
tmp = ev->next;
while ((tmp != currentPtr) && (tmp != backPtr) && (tmp->done == 1))
tmp = tmp->next;
if ((tmp != currentPtr) && (tmp != backPtr) && (tmp->done == 0))
RBevent = tmp;
}
// connect surrounding events
ev->prev->next = ev->next;
ev->next->prev = ev->prev;
POSE_TimeType ts = ev->timestamp;
if (!ev->done) eventCount--;
else mem_usage--;
delete ev; // then delete the event
if (ts == largest) FindLargest();
#ifdef EQ_SANITIZE
sanitize();
#endif
}
/** \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);
}
void pushSendQ(SysvshmSendQ *q,OutgoingMsg msg){
if(q->numEntries == q->size){
/* need to resize */
OutgoingMsg *oldData = q->data;
int newSize = q->size<<1;
q->data = (OutgoingMsg *)malloc(sizeof(OutgoingMsg)*newSize);
/* copy head to the beginning of the new array */
CmiAssert(q->begin == q->end);
CmiAssert(q->begin < q->size);
memcpy(&(q->data[0]),&(oldData[q->begin]),sizeof(OutgoingMsg)*(q->size - q->begin));
if(q->end != 0){
memcpy(&(q->data[(q->size - q->begin)]),&(oldData[0]),sizeof(OutgoingMsg)*(q->end));
}
free(oldData);
q->begin = 0;
q->end = q->size;
q->size = newSize;
}
q->data[q->end] = msg;
(q->end)++;
if(q->end >= q->size){
q->end -= q->size;
}
q->numEntries++;
}
void ProxyPatch::boxClosed(int box) {
if (box == 1) { // force Box
// Note: delay the deletion of proxyDataMsg (of the
// current step) until the next step. This is done
// for the sake of atom migration (ProxyDataMsg)
// as the ProxyPatch has to unregister the atoms
// of the previous step in the AtomMap data structure
// also denotes end of gbis phase 3
sendResults();
} else if ( box == 5) {//end phase 1
} else if ( box == 8) {//end phase 2
} else if (box == 9) {
//nothing
} else if (box == 10) {
// LCPO do nothing
}
if ( ! --boxesOpen ) {
DebugM(2,patchID << ": " << "Checking message buffer.\n");
if(proxyMsgBufferStatus == PROXYALLMSGBUFFERED) {
CmiAssert(curProxyMsg != NULL);
DebugM(3,"Patch " << patchID << " processing buffered proxy ALL data.\n");
receiveAll(curProxyMsg);
}else if(proxyMsgBufferStatus == PROXYDATAMSGBUFFERED) {
CmiAssert(curProxyMsg != NULL);
DebugM(3,"Patch " << patchID << " processing buffered proxy data.\n");
receiveData(curProxyMsg);
}
} else {
DebugM(3,"ProxyPatch " << patchID << ": " << boxesOpen << " boxes left to close.\n");
}
}
void createShmObject(char *name,int size,char **pPtr){
int fd=-1;
int flags; // opening flags for shared object
int open_repeat_count = 0;
flags= O_RDWR | O_CREAT; // open file in read-write mode and create it if its not there
while(fd<0 && open_repeat_count < 100){
open_repeat_count++;
fd = shm_open(name,flags, S_IRUSR | S_IWUSR); // create the shared object with permissions for only the user to read and write
if(fd < 0 && open_repeat_count > 10){
fprintf(stderr,"Error(attempt=%d) from shm_open %s while opening %s \n",open_repeat_count, strerror(errno),name);
fflush(stderr);
}
}
CmiAssert(fd >= 0);
ftruncate(fd,size); //set the size of the shared memory object
*pPtr = mmap(NULL,size,PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
CmiAssert(*pPtr != NULL);
close(fd);
}
/** Recursively packs/sizes entire subheap; DOES NOT UNPACK HEAP!!! */
void HeapNode::pup(PUP::er &p)
{
CmiAssert(this != NULL);
CmiAssert(!p.isUnpacking());
e->pup(p);
if (left) left->pup(p);
if (right) right->pup(p);
}
void ProxyPatch::boxClosed(int box) {
ProxyGBISP1ResultMsg *msg1;
ProxyGBISP2ResultMsg *msg2;
if (box == 1) { // force Box
// Note: delay the deletion of proxyDataMsg (of the
// current step) until the next step. This is done
// for the sake of atom migration (ProxyDataMsg)
// as the ProxyPatch has to unregister the atoms
// of the previous step in the AtomMap data structure
// also denotes end of gbis phase 3
sendResults();
} else if (box == 5) { //end phase 1
//this msg should only have nonzero atoms if flags.doNonbonded
int msgAtoms = (flags.doNonbonded) ? numAtoms : 0;
msg1 = new (msgAtoms, PRIORITY_SIZE) ProxyGBISP1ResultMsg;
for (int i = 0; i < msgAtoms; i++) {
msg1->psiSum[i] = psiSum[i];
}
msg1->patch = patchID;
msg1->psiSumLen = msgAtoms;
msg1->origPe = CkMyPe();
SET_PRIORITY(msg1, flags.sequence, GB1_PROXY_RESULTS_PRIORITY + PATCH_PRIORITY(patchID));
ProxyMgr::Object()->sendResult(msg1);
} else if (box == 8) { //end phase 2
//this msg should only have nonzero atoms if flags.doFullElectrostatics
int msgAtoms = (flags.doFullElectrostatics) ? numAtoms : 0;
msg2 = new (msgAtoms, PRIORITY_SIZE) ProxyGBISP2ResultMsg;
for (int i = 0; i < msgAtoms; i++) {
msg2->dEdaSum[i] = dEdaSum[i];
}
msg2->patch = patchID;
msg2->dEdaSumLen = msgAtoms;
msg2->origPe = CkMyPe();
SET_PRIORITY(msg2, flags.sequence, GB2_PROXY_RESULTS_PRIORITY + PATCH_PRIORITY(patchID));
ProxyMgr::Object()->sendResult(msg2);
} else if (box == 9) {
//nothing
} else if (box == 10) {
// LCPO do nothing
}
if (!--boxesOpen) {
DebugM(2, patchID << ": " << "Checking message buffer.\n");
if (proxyMsgBufferStatus == PROXYALLMSGBUFFERED) {
CmiAssert(curProxyMsg != NULL);
DebugM(3, "Patch " << patchID << " processing buffered proxy ALL data.\n");
receiveAll(curProxyMsg);
} else if (proxyMsgBufferStatus == PROXYDATAMSGBUFFERED) {
CmiAssert(curProxyMsg != NULL);
DebugM(3, "Patch " << patchID << " processing buffered proxy data.\n");
receiveData(curProxyMsg);
}
} else {
DebugM(3, "ProxyPatch " << patchID << ": " << boxesOpen << " boxes left to close.\n");
}
}
void CmiSendMessageSysvshm(OutgoingMsg ogm,OtherNode node,int rank,unsigned int broot){
struct sembuf sb;
#if SYSVSHM_STATS
double _startSendTime = CmiWallTimer();
#endif
int dstRank = SysvshmRank(ogm->dst);
MEMDEBUG(CmiMemoryCheck());
DgramHeaderMake(ogm->data,rank,ogm->src,Cmi_charmrun_pid,1, broot);
MACHSTATE4(3,"Send Msg Sysvshm ogm %p size %d dst %d dstRank %d",ogm,ogm->size,ogm->dst,dstRank);
CmiAssert(dstRank >=0 && dstRank != sysvshmContext->noderank);
sharedBufData *dstBuf = &(sysvshmContext->sendBufs[dstRank]);
ACQUIRENW(sysvshmContext->noderank);
if(semop(dstBuf->semid, &sb, 1)<0) {
/**failed to get the lock
insert into q and retain the message*/
pushSendQ(sysvshmContext->sendQs[dstRank],ogm);
ogm->refcount++;
MEMDEBUG(CmiMemoryCheck());
return;
}else{
/***
* We got the lock for this buffer
* first write all the messages in the sendQ and then write this guy
* */
if(sysvshmContext->sendQs[dstRank]->numEntries == 0){
/* send message user event */
int ret = sendMessage(ogm,dstBuf,sysvshmContext->sendQs[dstRank]);
MACHSTATE(3,"Sysvshm Send succeeded immediately");
}else{
ogm->refcount+=2;/*this message should not get deleted when the queue is flushed*/
pushSendQ(sysvshmContext->sendQs[dstRank],ogm);
MACHSTATE3(3,"Sysvshm ogm %p pushed to sendQ length %d refcount %d",ogm,sysvshmContext->sendQs[dstRank]->numEntries,ogm->refcount);
int sent = flushSendQ(dstRank);
ogm->refcount--; /*if it has been sent, can be deleted by caller, if not will be deleted when queue is flushed*/
MACHSTATE1(3,"Sysvshm flushSendQ sent %d messages",sent);
}
/* unlock the recvbuffer*/
RELEASE(sysvshmContext->noderank);
CmiAssert(semop(dstBuf->semid, &sb, 1)>=0);
}
#if SYSVSHM_STATS
sysvshmContext->sendCount ++;
sysvshmContext->sendTime += (CmiWallTimer()-_startSendTime);
#endif
MEMDEBUG(CmiMemoryCheck());
};
int ScriptTcl::Tcl_replicaAtomSendrecv(ClientData clientData, Tcl_Interp *interp, int argc, char **argv) {
ScriptTcl *script = (ScriptTcl *)clientData;
script->initcheck();
if ( ! Node::Object()->simParameters->replicaUniformPatchGrids ) {
Tcl_SetResult(interp,"replicaUniformPatchGrids is required for atom exchange",TCL_VOLATILE);
return TCL_ERROR;
}
if ( argc < 2 || argc > 3 ) {
Tcl_SetResult(interp,"bad arg count; args: dest ?source?",TCL_VOLATILE);
return TCL_ERROR;
}
int dest = -1;
if ( sscanf(argv[1], "%d", &dest) != 1 ) {
Tcl_SetResult(interp,"bad dest; args: dest ?source?",TCL_VOLATILE);
return TCL_ERROR;
}
int source = -1;
if ( argc == 3 ) {
if ( sscanf(argv[2], "%d", &source) != 1 ) {
Tcl_SetResult(interp,"bad source; args: dest ?source?",TCL_VOLATILE);
return TCL_ERROR;
}
}
#if CMK_HAS_PARTITION
if (dest != CmiMyPartition()) {
DataMessage *recvMsg = NULL;
replica_sendRecv((char*)&(script->state->lattice), sizeof(Lattice), dest, CkMyPe(), &recvMsg, source, CkMyPe());
CmiAssert(recvMsg != NULL);
memcpy(&(script->state->lattice), recvMsg->data, recvMsg->size);
CmiFree(recvMsg);
}
#endif
char str[40];
sprintf(str, "%d", dest);
script->setParameter("scriptArg1", str);
sprintf(str, "%d", source);
script->setParameter("scriptArg2", str);
CkpvAccess(_qd)->create(2 * PatchMap::Object()->numPatches());
script->runController(SCRIPT_ATOMSENDRECV);
#if CMK_HAS_PARTITION
if (dest != CmiMyPartition()) {
DataMessage *recvMsg = NULL;
ControllerState *cstate = script->state->controller;
replica_sendRecv((char*)cstate, sizeof(ControllerState), dest, CkMyPe(), &recvMsg, source, CkMyPe());
CmiAssert(recvMsg != NULL);
memcpy(cstate, recvMsg->data, recvMsg->size);
CmiFree(recvMsg);
}
#endif
return TCL_OK;
}
// return the seq-th load balancer string name of
// it can be specified in either compile time or runtime
// runtime has higher priority
const char *LBDatabase::loadbalancer(int seq) {
if (lbRegistry.runtime_lbs.length()) {
CmiAssert(seq < lbRegistry.runtime_lbs.length());
return lbRegistry.runtime_lbs[seq];
}
else {
CmiAssert(seq < lbRegistry.compile_lbs.length());
return lbRegistry.compile_lbs[seq];
}
}
// migration done at current lbLevel
void HbmLB::MigrationDone(int balancing)
{
#if CMK_LBDB_ON
int i, j;
LevelData *lData = levelData[0];
DEBUGF(("[%d] HbmLB::MigrationDone lbLevel:%d numLevels:%d!\n", CkMyPe(), lbLevel, tree->numLevels()));
CmiAssert(newObjs.size() == lData->migrates_expected);
#if 0
if (lbLevel == tree->numLevels()-1) {
theLbdb->incStep();
// reset
lData->clear();
}
else {
lData->migrates_expected = -1;
lData->migrates_completed = 0;
lData->obj_completed = 0;
}
#else
lData->migrates_expected = -1;
lData->migrates_completed = 0;
lData->obj_completed = 0;
#endif
CkVec<LDObjData> &oData = myStats.objData;
// update
int count=0;
for (i=0; i<oData.size(); i++)
if (oData[i].handle.handle == -100) count++;
CmiAssert(count == newObjs.size());
for (i=0; i<oData.size(); i++) {
if (oData[i].handle.handle == -100) {
LDObjHandle &handle = oData[i].handle;
for (j=0; j<newObjs.size(); j++) {
if (handle.omID() == newObjs[j].omID() &&
handle.objID() == newObjs[j].objID()) {
handle = newObjs[j];
break;
}
}
CmiAssert(j<newObjs.size());
}
}
newObjs.free();
thisProxy[lData->parent].NotifyObjectMigrationDone(0, lbLevel);
#endif
}
void ComputeMap::loadComputeMap(const char *fname)
{
FILE *fp = fopen(fname, "r");
CmiAssert(fp != NULL);
int n;
fscanf(fp, "%d\n", &n);
CmiAssert(n == nComputes);
for(int i=0; i < nComputes; i++)
{
fscanf(fp, "%d\n", &computeData[i].node);
}
fclose(fp);
}
/// Remove heap node matching evID
int HeapNode::remove(eventID evID, POSE_TimeType timestamp)
{
#ifdef EH_SANITIZE
sanitize();
#endif
CmiAssert(this != NULL);
CmiAssert(!(this->e->evID == evID));
int found = 0; // return status
HeapNode *tmp;
if (left) { // search left subheap first
if (timestamp < left->e->timestamp) found = 0; // subheap elements too high
else if ((timestamp == left->e->timestamp) && (evID == left->e->evID)) {
// found element on top
tmp = left; // save a pointer to it
if (left->left)
left = left->left->conjoin(left->right); // remove it from heap
else left = left->right;
subheapsize--;
tmp->left = tmp->right = NULL; // foil recursive destructor
delete tmp; // delete it
found = 1;
}
else if (timestamp >= left->e->timestamp) { // need to look deeper
found = left->remove(evID, timestamp);
if (found) subheapsize--; // must decrement heap size
}
}
if (found) return 1; // found in left subheap; exit with status 1
else if (right) { // search in right subheap
if (timestamp < right->e->timestamp) found = 0; //subheap elements too high
else if ((timestamp == right->e->timestamp) && (evID == right->e->evID)) {
// found element on top
tmp = right; // save a pointer to it
if (right->left)
right = right->left->conjoin(right->right); // remove it from heap
else right = right->right;
subheapsize--;
tmp->left = tmp->right = NULL; // foil recursive destructor
delete tmp; // delete it
found = 1;
}
else if (timestamp >= right->e->timestamp) { // need to look deeper
found = right->remove(evID, timestamp); // set return status
if (found) subheapsize--; // must decrement heap size
}
}
#ifdef EH_SANITIZE
sanitize();
#endif
return found; // exit with found status
}
// remove the obj-th object from database
void BaseLB::LDStats::removeObject(int obj)
{
CmiAssert(obj < objData.size());
LDObjData odata = objData[obj];
LDObjKey okey; // build a key
okey.omID() = odata.omID();
okey.objID() = odata.objID();
objData.remove(obj);
from_proc.remove(obj);
to_proc.remove(obj);
n_objs --;
if (odata.migratable) n_migrateobjs --;
// search for sender, can be multiple sender
int removed = 0;
for (int com=0; com<n_comm; com++) {
LDCommData &cdata = commData[com-removed];
if(!cdata.from_proc() && cdata.sender == okey) {
commData.remove(com-removed);
removed++;
}
}
n_comm -= removed;
}
static inline void compute_energy_orthogonal_ckloop(int first, int last, void *result, int paraNum, void *param){
CmiAssert(first==last);
void **params = (void **)param;
PmeKSpace *kspace = (PmeKSpace *)params[0];
float *q_arr = (float *)params[1];
double *recips = (double *)params[2];
double *partialEnergy = (double *)params[3];
double *partialVirial = (double *)params[4];
int *unitDist = (int *)params[5];
int i = first;
int unit = unitDist[0];
int remains = unitDist[1];
int k1from, k1to;
if(i<remains){
k1from = i*(unit+1);
k1to = k1from+unit;
}else{
k1from = remains*(unit+1)+(i-remains)*unit;
k1to = k1from+unit-1;
}
double *pEnergy = partialEnergy+i;
double *pVirial = partialVirial+i*6;
kspace->compute_energy_orthogonal_subset(q_arr, recips, pVirial, pEnergy, k1from, k1to);
}
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 SayHi(HiMsg *m)
{
redNo ++;
CmiAssert(m->data[0] == 22 && m->data[1] == 28);
CkGetSectionInfo(sid, m);
CkMulticastMgr *mg = CProxy_CkMulticastMgr(mCastGrpId).ckLocalBranch();
int dataSize = (int)(CompleteMsgSize);
int* data = new int [dataSize];
int fragSize = dataSize/NumFrags;
CkAssert (0 != fragSize);
for (int i=0; i<dataSize; i++) {
data [i] = thisIndex;
}
CkCallback cb(CkIndex_Hello::cb_client(NULL), CkArrayIndex1D(0), thisProxy);
mg->contribute(dataSize*sizeof(int), data,CkReduction::sum_int, sid, cb, fragSize*sizeof(int));
// data[0] = thisIndex+2;
// data[1] = thisIndex+2;
// mg->contribute(2*sizeof(int), &data,CkReduction::max_int, sid, sizeof(int));
// data[0] = thisIndex+1;
// data[1] = thisIndex+1;
// mg->contribute(2*sizeof(int), &data,CkReduction::product_int, sid, sizeof(int));
delete m;
if (1)
ckMigrate((CkMyPe()+1)%CkNumPes());
}
void TraceBluegene::beginExecute(envelope *e)
{
if (e==NULL || !genTimeLog) return;
BgTimeLog* log = tTIMELINE[tTIMELINE.length()-1];
CmiAssert(log!=NULL);
log->setCharmEP(e->getEpIdx());
}
void static inline handoverSysvshmMessage(char *newmsg,int total_size,int rank,int broot){
CmiAssert(rank == 0);
#if CMK_BROADCAST_SPANNING_TREE
if (rank == DGRAM_BROADCAST
#if CMK_NODE_QUEUE_AVAILABLE
|| rank == DGRAM_NODEBROADCAST
#endif
){
SendSpanningChildren(NULL, 0, total_size, newmsg,broot,rank);
}
#elif CMK_BROADCAST_HYPERCUBE
if (rank == DGRAM_BROADCAST
#if CMK_NODE_QUEUE_AVAILABLE
|| rank == DGRAM_NODEBROADCAST
#endif
){
SendHypercube(NULL, 0, total_size, newmsg,broot,rank);
}
#endif
switch (rank) {
case DGRAM_BROADCAST: {
CmiPushPE(0, newmsg);
break;
}
default:
{
CmiPushPE(rank, newmsg);
}
} /* end of switch */
}
void emptyRecvBuf(sharedBufData *recvBuf){
int numMessages = recvBuf->header->count;
int i=0;
char *ptr=recvBuf->data;
for(i=0;i<numMessages;i++){
int size;
int rank, srcpe, seqno, magic, i;
unsigned int broot;
char *msg = ptr;
char *newMsg;
DgramHeaderBreak(msg, rank, srcpe, magic, seqno, broot);
size = CmiMsgHeaderGetLength(msg);
newMsg = (char *)CmiAlloc(size);
memcpy(newMsg,msg,size);
handoverSysvshmMessage(newMsg,size,rank,broot);
ptr += size;
MACHSTATE3(3,"message of size %d recvd ends at ptr-data %d total bytes %d bytes %d",size,ptr-recvBuf->data,recvBuf->header->bytes);
}
CmiAssert(ptr - recvBuf->data == recvBuf->header->bytes);
recvBuf->header->count=0;
recvBuf->header->bytes=0;
}
inline void emptyAllRecvBufs(){
struct sembuf sb;
int i;
int j,ret;
union semun {
int val;
struct semid_ds *buf;
ushort array[1];
} arg;
for(i=0;i<sysvshmContext->nodesize;i++){
if(i != sysvshmContext->noderank){
sharedBufData *recvBuf = &(sysvshmContext->recvBufs[i]);
if(recvBuf->header->count > 0){
#if SYSVSHM_STATS
sysvshmContext->lockRecvCount++;
#endif
ACQUIRE(i);
if(semop(recvBuf->semid, &sb, 1)>=0) {
MACHSTATE1(3,"emptyRecvBuf to be called for rank %d",i);
emptyRecvBuf(recvBuf);
RELEASE(i);
CmiAssert((semop(recvBuf->semid, &sb, 1))>=0);
}
}
}
}
};
void TraceBluegene::addBackwardDep(void *log)
{
if(!genTimeLog || log==NULL) return;
BgTimeLog *parentLogPtr = BgLastLog(tTIMELINEREC);
CmiAssert(parentLogPtr);
BgAddBackwardDep(parentLogPtr, (BgTimeLog*)log);
}
// called on every processor
void CentralLB::SendStats()
{
#if CMK_LBDB_ON
CmiAssert(statsMsg != NULL);
reduction_started = 0;
#if USE_LDB_SPANNING_TREE
if(CkNumPes()>1024)
{
if (CkMyPe() == cur_ld_balancer)
thisProxy[CkMyPe()].ReceiveStats(statsMsg);
else
thisProxy[CkMyPe()].ReceiveStatsViaTree(statsMsg);
}
else
#endif
{
DEBUGF(("[%d] calling ReceiveStats on step %d \n",CmiMyPe(),step()));
thisProxy[cur_ld_balancer].ReceiveStats(statsMsg);
}
statsMsg = NULL;
#ifdef __BIGSIM__
BgEndStreaming();
#endif
{
// enfore the barrier to wait until centralLB says no
LDOMHandle h;
h.id.id.idx = 0;
theLbdb->getLBDB()->RegisteringObjects(h);
}
#endif
}
void CentralLB::ProcessAtSync()
{
#if CMK_LBDB_ON
if (reduction_started) return; // reducton in progress
CmiAssert(CmiNodeAlive(CkMyPe()));
if (CkMyPe() == cur_ld_balancer) {
start_lb_time = CkWallTimer();
}
#if (defined(_FAULT_MLOG_) || defined(_FAULT_CAUSAL_))
initMlogLBStep(thisgroup);
#endif
// build message
BuildStatsMsg();
#if USE_REDUCTION
// reduction to get total number of objects and comm
// so that processor 0 can pre-allocate load balancing database
int counts[2];
counts[0] = theLbdb->GetObjDataSz();
counts[1] = theLbdb->GetCommDataSz();
CkCallback cb(CkIndex_CentralLB::ReceiveCounts((CkReductionMsg*)NULL),
thisProxy[0]);
contribute(2*sizeof(int), counts, CkReduction::sum_int, cb);
reduction_started = 1;
#else
SendStats();
#endif
#endif
}
/// Change currentPtr to point to event e
void eventQueue::SetCurrentPtr(Event *e) {
Event *tmp = e;
// Check possible error conditions
CmiAssert((e->done == 0) || (e->done == -1)); // e is not done or a sentinel
CmiAssert(currentPtr->done != 2);
currentPtr = e; // move currentPtr to e
if ((currentPtr == backPtr) && (eqh->top)) { // moved currentPtr to backPtr
// get next event from heap and put it in the queue
tmp = eqh->GetAndRemoveTopEvent();
tmp->next = currentPtr;
tmp->prev = currentPtr->prev;
currentPtr->prev->next = tmp;
currentPtr->prev = tmp;
currentPtr = tmp;
}
}
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);
}
void _CkExit(void)
{
CmiAssert(CkMyPe() == 0);
// Shuts down Converse handlers for the upper layers on this processor
//
CkNumberHandler(_charmHandlerIdx,(CmiHandler)_discardHandler);
CkNumberHandler(_bocHandlerIdx, (CmiHandler)_discardHandler);
DEBUGF(("[%d] CkExit - _exitStarted:%d %d\n", CkMyPe(), _exitStarted, _exitHandlerIdx));
if(CkMyPe()==0) {
if(_exitStarted)
CsdScheduler(-1);
envelope *env = _allocEnv(ReqStatMsg);
env->setSrcPe(CkMyPe());
CmiSetHandler(env, _exitHandlerIdx);
/*FAULT_EVAC*/
CmiSyncBroadcastAllAndFree(env->getTotalsize(), (char *)env);
} else {
envelope *env = _allocEnv(ExitMsg);
env->setSrcPe(CkMyPe());
CmiSetHandler(env, _exitHandlerIdx);
CmiSyncSendAndFree(0, env->getTotalsize(), (char *)env);
}
#if ! CMK_BIGSIM_THREAD
_TRACE_END_EXECUTE();
//Wait for stats, which will call ConverseExit when finished:
CsdScheduler(-1);
#endif
}
