void ConvComlibManager::doneCreating() {
ComlibPrintf("Called doneCreating\n");
if (busy) {
// we have to delay the table broadcast because we are in the middle of another one
doneCreatingScheduled = CmiTrue;
return;
}
// if we reach here it means we are not busy and we can proceed
busy = CmiTrue;
acksReceived = CmiNumPes() - 1;
int count = 0;
for (int i=1; i<=nstrats; ++i) {
if (strategyTable[i].isNew) {
count++;
}
}
if (count > 0) {
// create the wrapper and link the strategies there
StrategyWrapper sw(count);
count = 0;
for (int i=1; i<=nstrats; ++i) {
if (strategyTable[i].isNew) {
sw.position[count] = i;
sw.replace[count] = CmiFalse;
sw.strategy[count] = strategyTable[i].strategy;
count++;
CkpvAccess(conv_com_object).inSync(i);
}
}
// pup the wrapper into a message
PUP::sizer ps;
ps|sw;
char *msg = (char*)CmiAlloc(ps.size() + CmiReservedHeaderSize);
PUP::toMem pm(msg+CmiReservedHeaderSize);
//int size = ps.size();
//pm|size;
pm|sw;
//for (int i=CmiReservedHeaderSize; i<CmiReservedHeaderSize+size; ++i) {
// CmiPrintf("%x",((char*)msg)[i]);
//}
//CmiPrintf("\n");
CmiSetHandler(msg, CkpvAccess(comlib_receive_table));
CmiSyncBroadcastAndFree(ps.size()+CmiReservedHeaderSize, msg);
/* NOT USED NOW!
// call the finalizeCreation after the strategies has been packed
for (int i=0; i<strategyTable.size(); ++i) {
if (strategyTable[i].isNew) strategyTable[i].strategy->finalizeCreation();
}
*/
} else {
busy = CmiFalse;
}
}
static void bdcastRO(void){
int i;
//Determine the size of the RODataMessage
PUP::sizer ps;
for(i=0;i<_readonlyTable.size();i++) _readonlyTable[i]->pupData(ps);
//Allocate and fill out the RODataMessage
envelope *env = _allocEnv(RODataMsg, ps.size());
PUP::toMem pp((char *)EnvToUsr(env));
for(i=0;i<_readonlyTable.size();i++) _readonlyTable[i]->pupData(pp);
env->setCount(++_numInitMsgs);
env->setSrcPe(CkMyPe());
CmiSetHandler(env, _roRestartHandlerIdx);
CmiSyncBroadcastAndFree(env->getTotalsize(), (char *)env);
}
std::pair< int, std::unique_ptr<char[]> >
serialize( const std::vector< std::size_t >& v )
// *****************************************************************************
// Serialize std::vectors to raw memory stream
//! \param[in] v Vector
//! \return Pair of the length and the raw stream containing the serialized
//! vectors
//! \author J. Bakosi
// *****************************************************************************
{
// Prepare for serializing vectors to a raw binary stream, compute size
PUP::sizer sizer;
sizer | const_cast< std::vector< std::size_t >& >( v );
// Create raw character stream to store the serialized vectors
std::unique_ptr<char[]> flatData = tk::make_unique<char[]>( sizer.size() );
// Serialize vector, each message will contain a vector
PUP::toMem packer( flatData.get() );
packer | const_cast< std::vector< std::size_t >& >( v );
// Return size of and raw stream
return { sizer.size(), std::move(flatData) };
}
/**
This is the main charm setup routine. It's called
on all processors after Converse initialization.
This routine gets passed to Converse from "main.C".
The main purpose of this routine is to set up the objects
and Ckpv's used during a regular Charm run. See the comment
at the top of the file for overall flow.
*/
void _initCharm(int unused_argc, char **argv)
{
int inCommThread = (CmiMyRank() == CmiMyNodeSize());
DEBUGF(("[%d,%.6lf ] _initCharm started\n",CmiMyPe(),CmiWallTimer()));
CkpvInitialize(size_t *, _offsets);
CkpvAccess(_offsets) = new size_t[32];
CkpvInitialize(PtrQ*,_buffQ);
CkpvInitialize(PtrVec*,_bocInitVec);
CkpvInitialize(void*, _currentChare);
CkpvInitialize(int, _currentChareType);
CkpvInitialize(CkGroupID, _currentGroup);
CkpvInitialize(void *, _currentNodeGroupObj);
CkpvInitialize(CkGroupID, _currentGroupRednMgr);
CkpvInitialize(GroupTable*, _groupTable);
CkpvInitialize(GroupIDTable*, _groupIDTable);
CkpvInitialize(CmiImmediateLockType, _groupTableImmLock);
CkpvInitialize(bool, _destroyingNodeGroup);
CkpvAccess(_destroyingNodeGroup) = false;
CkpvInitialize(UInt, _numGroups);
CkpvInitialize(int, _numInitsRecd);
CkpvInitialize(int, _initdone);
CkpvInitialize(char**, Ck_argv); CkpvAccess(Ck_argv)=argv;
CkpvInitialize(MsgPool*, _msgPool);
CkpvInitialize(CkCoreState *, _coreState);
/*
Added for evacuation-sayantan
*/
#ifndef __BIGSIM__
CpvInitialize(char *,_validProcessors);
#endif
CkpvInitialize(char ,startedEvac);
CpvInitialize(int,serializer);
_initChareTables(); // for checkpointable plain chares
CksvInitialize(UInt, _numNodeGroups);
CksvInitialize(GroupTable*, _nodeGroupTable);
CksvInitialize(GroupIDTable, _nodeGroupIDTable);
CksvInitialize(CmiImmediateLockType, _nodeGroupTableImmLock);
CksvInitialize(CmiNodeLock, _nodeLock);
CksvInitialize(PtrVec*,_nodeBocInitVec);
CksvInitialize(UInt,_numInitNodeMsgs);
CkpvInitialize(int,_charmEpoch);
CkpvAccess(_charmEpoch)=0;
CksvInitialize(int, _triggersSent);
CksvAccess(_triggersSent) = 0;
CkpvInitialize(_CkOutStream*, _ckout);
CkpvInitialize(_CkErrStream*, _ckerr);
CkpvInitialize(Stats*, _myStats);
CkpvAccess(_groupIDTable) = new GroupIDTable(0);
CkpvAccess(_groupTable) = new GroupTable;
CkpvAccess(_groupTable)->init();
CkpvAccess(_groupTableImmLock) = CmiCreateImmediateLock();
CkpvAccess(_numGroups) = 1; // make 0 an invalid group number
CkpvAccess(_buffQ) = new PtrQ();
CkpvAccess(_bocInitVec) = new PtrVec();
CkpvAccess(_currentNodeGroupObj) = NULL;
if(CkMyRank()==0)
{
CksvAccess(_numNodeGroups) = 1; //make 0 an invalid group number
CksvAccess(_numInitNodeMsgs) = 0;
CksvAccess(_nodeLock) = CmiCreateLock();
CksvAccess(_nodeGroupTable) = new GroupTable();
CksvAccess(_nodeGroupTable)->init();
CksvAccess(_nodeGroupTableImmLock) = CmiCreateImmediateLock();
CksvAccess(_nodeBocInitVec) = new PtrVec();
}
CkCallbackInit();
CmiNodeAllBarrier();
#if ! CMK_BIGSIM_CHARM
initQd(argv); // bigsim calls it in ConverseCommonInit
#endif
CkpvAccess(_coreState)=new CkCoreState();
CkpvAccess(_numInitsRecd) = 0;
CkpvAccess(_initdone) = 0;
CkpvAccess(_ckout) = new _CkOutStream();
CkpvAccess(_ckerr) = new _CkErrStream();
_charmHandlerIdx = CkRegisterHandler((CmiHandler)_bufferHandler);
_initHandlerIdx = CkRegisterHandler((CmiHandler)_initHandler);
CkNumberHandlerEx(_initHandlerIdx, (CmiHandlerEx)_initHandler, CkpvAccess(_coreState));
_roRestartHandlerIdx = CkRegisterHandler((CmiHandler)_roRestartHandler);
_exitHandlerIdx = CkRegisterHandler((CmiHandler)_exitHandler);
//added for interoperabilitY
_libExitHandlerIdx = CkRegisterHandler((CmiHandler)_libExitHandler);
_bocHandlerIdx = CkRegisterHandler((CmiHandler)_initHandler);
CkNumberHandlerEx(_bocHandlerIdx, (CmiHandlerEx)_initHandler, CkpvAccess(_coreState));
#ifdef __BIGSIM__
if(BgNodeRank()==0)
#endif
_infoIdx = CldRegisterInfoFn((CldInfoFn)_infoFn);
_triggerHandlerIdx = CkRegisterHandler((CmiHandler)_triggerHandler);
_ckModuleInit();
CldRegisterEstimator((CldEstimator)_charmLoadEstimator);
_futuresModuleInit(); // part of futures implementation is a converse module
_loadbalancerInit();
_metabalancerInit();
#if CMK_MEM_CHECKPOINT
init_memcheckpt(argv);
#endif
initCharmProjections();
#if CMK_TRACE_IN_CHARM
// initialize trace module in ck
traceCharmInit(argv);
#endif
CkpvInitialize(int, envelopeEventID);
CkpvAccess(envelopeEventID) = 0;
CkMessageWatcherInit(argv,CkpvAccess(_coreState));
/**
The rank-0 processor of each node calls the
translator-generated "_register" routines.
_register routines call the charm.h "CkRegister*" routines,
which record function pointers and class information for
all Charm entities, like Chares, Arrays, and readonlies.
There's one _register routine generated for each
.ci file. _register routines *must* be called in the
same order on every node, and *must not* be called by
multiple threads simultaniously.
*/
#ifdef __BIGSIM__
if(BgNodeRank()==0)
#else
if(CkMyRank()==0)
#endif
{
SDAG::registerPUPables();
CmiArgGroup("Charm++",NULL);
_parseCommandLineOpts(argv);
_registerInit();
CkRegisterMsg("System", 0, 0, CkFreeMsg, sizeof(int));
CkRegisterChareInCharm(CkRegisterChare("null", 0, TypeChare));
CkIndex_Chare::__idx=CkRegisterChare("Chare", sizeof(Chare), TypeChare);
CkRegisterChareInCharm(CkIndex_Chare::__idx);
CkIndex_Group::__idx=CkRegisterChare("Group", sizeof(Group), TypeGroup);
CkRegisterChareInCharm(CkIndex_Group::__idx);
CkRegisterEp("null", (CkCallFnPtr)_nullFn, 0, 0, 0+CK_EP_INTRINSIC);
/**
These _register calls are for the built-in
Charm .ci files, like arrays and load balancing.
If you add a .ci file to charm, you'll have to
add a call to the _register routine here, or make
your library into a "-module".
*/
_registerCkFutures();
_registerCkArray();
_registerLBDatabase();
_registerMetaBalancer();
_registerCkCallback();
_registertempo();
_registerwaitqd();
_registerCkCheckpoint();
#if CMK_MEM_CHECKPOINT
_registerCkMemCheckpoint();
#endif
/*
Setup Control Point Automatic Tuning Framework.
By default it is enabled as a part of charm,
however it won't enable its tracing module
unless a +CPEnableMeasurements command line argument
is specified. See trace-common.C for more info
Thus there should be no noticable overhead to
always having the control point framework linked
in.
*/
#if CMK_WITH_CONTROLPOINT
_registerPathHistory();
_registerControlPoints();
_registerTraceControlPoints();
#endif
/**
CkRegisterMainModule is generated by the (unique)
"mainmodule" .ci file. It will include calls to
register all the .ci files.
*/
CkRegisterMainModule();
/**
_registerExternalModules is actually generated by
charmc at link time (as "moduleinit<pid>.C").
This generated routine calls the _register functions
for the .ci files of libraries linked using "-module".
This funny initialization is most useful for AMPI/FEM
programs, which don't have a .ci file and hence have
no other way to control the _register process.
*/
_registerExternalModules(argv);
_registerDone();
}
/* The following will happen on every virtual processor in BigEmulator, not just on once per real processor */
if (CkMyRank() == 0) {
CpdBreakPointInit();
}
CmiNodeAllBarrier();
// Execute the initcalls registered in modules
_initCallTable.enumerateInitCalls();
#if CMK_CHARMDEBUG
CpdFinishInitialization();
#endif
//CmiNodeAllBarrier();
CkpvAccess(_myStats) = new Stats();
CkpvAccess(_msgPool) = new MsgPool();
CmiNodeAllBarrier();
#if !(__FAULT__)
CmiBarrier();
CmiBarrier();
CmiBarrier();
#endif
#if CMK_SMP_TRACE_COMMTHREAD
_TRACE_BEGIN_COMPUTATION();
#else
if (!inCommThread) {
_TRACE_BEGIN_COMPUTATION();
}
#endif
#ifdef ADAPT_SCHED_MEM
if(CkMyRank()==0){
memCriticalEntries = new int[numMemCriticalEntries];
int memcnt=0;
for(int i=0; i<_entryTable.size(); i++){
if(_entryTable[i]->isMemCritical){
memCriticalEntries[memcnt++] = i;
}
}
}
#endif
#if (defined(_FAULT_MLOG_) || defined(_FAULT_CAUSAL_))
_messageLoggingInit();
#endif
#ifndef __BIGSIM__
/*
FAULT_EVAC
*/
CpvAccess(_validProcessors) = new char[CkNumPes()];
for(int vProc=0;vProc<CkNumPes();vProc++){
CpvAccess(_validProcessors)[vProc]=1;
}
_ckEvacBcastIdx = CkRegisterHandler((CmiHandler)_ckEvacBcast);
_ckAckEvacIdx = CkRegisterHandler((CmiHandler)_ckAckEvac);
#endif
CkpvAccess(startedEvac) = 0;
CpvAccess(serializer) = 0;
evacuate = 0;
CcdCallOnCondition(CcdSIGUSR1,(CcdVoidFn)CkDecideEvacPe,0);
#if (defined(_FAULT_MLOG_) || defined(_FAULT_CAUSAL_))
CcdCallOnCondition(CcdSIGUSR2,(CcdVoidFn)CkMlogRestart,0);
#endif
if(_raiseEvac){
processRaiseEvacFile(_raiseEvacFile);
/*
if(CkMyPe() == 2){
// CcdCallOnConditionKeep(CcdPERIODIC_10s,(CcdVoidFn)CkDecideEvacPe,0);
CcdCallFnAfter((CcdVoidFn)CkDecideEvacPe, 0, 10000);
}
if(CkMyPe() == 3){
CcdCallFnAfter((CcdVoidFn)CkDecideEvacPe, 0, 10000);
}*/
}
if (CkMyRank() == 0) {
TopoManager_init();
}
CmiNodeAllBarrier();
if (!_replaySystem) {
CkFtFn faultFunc_restart = CkRestartMain;
if (faultFunc == NULL || faultFunc == faultFunc_restart) { // this is not restart from memory
// these two are blocking calls for non-bigsim
#if ! CMK_BIGSIM_CHARM
CmiInitCPUAffinity(argv);
CmiInitMemAffinity(argv);
#endif
}
CmiInitCPUTopology(argv);
#if CMK_SHARED_VARS_POSIX_THREADS_SMP
if (CmiCpuTopologyEnabled()) {
int *pelist;
int num;
CmiGetPesOnPhysicalNode(0, &pelist, &num);
#if !CMK_MULTICORE && !CMK_SMP_NO_COMMTHD
// Count communication threads, if present
// XXX: Assuming uniformity of node size here
num += num/CmiMyNodeSize();
#endif
if (!_Cmi_forceSpinOnIdle && num > CmiNumCores())
{
if (CmiMyPe() == 0)
CmiPrintf("\nCharm++> Warning: the number of SMP threads (%d) is greater than the number of physical cores (%d), so threads will sleep while idling. Use +CmiSpinOnIdle or +CmiSleepOnIdle to control this directly.\n\n", num, CmiNumCores());
CmiLock(CksvAccess(_nodeLock));
if (! _Cmi_sleepOnIdle) _Cmi_sleepOnIdle = 1;
CmiUnlock(CksvAccess(_nodeLock));
}
}
#endif
}
if(CmiMyPe() == 0) {
char *topoFilename;
if(CmiGetArgStringDesc(argv,"+printTopo",&topoFilename,"topo file name"))
{
std::stringstream sstm;
sstm << topoFilename << "." << CmiMyPartition();
std::string result = sstm.str();
FILE *fp;
fp = fopen(result.c_str(), "w");
if (fp == NULL) {
CkPrintf("Error opening %s file, writing to stdout\n", topoFilename);
fp = stdout;
}
TopoManager_printAllocation(fp);
fclose(fp);
}
}
#if CMK_USE_PXSHM && ( CMK_CRAYXE || CMK_CRAYXC ) && CMK_SMP
// for SMP on Cray XE6 (hopper) it seems pxshm has to be initialized
// again after cpuaffinity is done
if (CkMyRank() == 0) {
CmiInitPxshm(argv);
}
CmiNodeAllBarrier();
#endif
//CldCallback();
#if CMK_BIGSIM_CHARM && CMK_CHARMDEBUG
// Register the BG handler for CCS. Notice that this is put into a variable shared by
// the whole real processor. This because converse needs to find it. We check that all
// virtual processors register the same index for this handler.
CpdBgInit();
#endif
if (faultFunc) {
#if CMK_WITH_STATS
if (CkMyPe()==0) _allStats = new Stats*[CkNumPes()];
#endif
if (!inCommThread) {
CkArgMsg *msg = (CkArgMsg *)CkAllocMsg(0, sizeof(CkArgMsg), 0);
msg->argc = CmiGetArgc(argv);
msg->argv = argv;
faultFunc(_restartDir, msg);
CkFreeMsg(msg);
}
}else if(CkMyPe()==0){
#if CMK_WITH_STATS
_allStats = new Stats*[CkNumPes()];
#endif
register size_t i, nMains=_mainTable.size();
for(i=0;i<nMains;i++) /* Create all mainchares */
{
register int size = _chareTable[_mainTable[i]->chareIdx]->size;
register void *obj = malloc(size);
_MEMCHECK(obj);
_mainTable[i]->setObj(obj);
CkpvAccess(_currentChare) = obj;
CkpvAccess(_currentChareType) = _mainTable[i]->chareIdx;
register CkArgMsg *msg = (CkArgMsg *)CkAllocMsg(0, sizeof(CkArgMsg), 0);
msg->argc = CmiGetArgc(argv);
msg->argv = argv;
_entryTable[_mainTable[i]->entryIdx]->call(msg, obj);
#if (defined(_FAULT_MLOG_) || defined(_FAULT_CAUSAL_))
CpvAccess(_currentObj) = (Chare *)obj;
#endif
}
_mainDone = 1;
_STATS_RECORD_CREATE_CHARE_N(nMains);
_STATS_RECORD_PROCESS_CHARE_N(nMains);
for(i=0;i<_readonlyMsgs.size();i++) /* Send out readonly messages */
{
register void *roMsg = (void *) *((char **)(_readonlyMsgs[i]->pMsg));
if(roMsg==0)
continue;
//Pack the message and send it to all other processors
register envelope *env = UsrToEnv(roMsg);
env->setSrcPe(CkMyPe());
env->setMsgtype(ROMsgMsg);
env->setRoIdx(i);
CmiSetHandler(env, _initHandlerIdx);
CkPackMessage(&env);
CmiSyncBroadcast(env->getTotalsize(), (char *)env);
CpvAccess(_qd)->create(CkNumPes()-1);
//For processor 0, unpack and re-set the global
CkUnpackMessage(&env);
_processROMsgMsg(env);
_numInitMsgs++;
}
//Determine the size of the RODataMessage
PUP::sizer ps;
for(i=0;i<_readonlyTable.size();i++) _readonlyTable[i]->pupData(ps);
//Allocate and fill out the RODataMessage
envelope *env = _allocEnv(RODataMsg, ps.size());
PUP::toMem pp((char *)EnvToUsr(env));
for(i=0;i<_readonlyTable.size();i++) _readonlyTable[i]->pupData(pp);
env->setCount(++_numInitMsgs);
env->setSrcPe(CkMyPe());
CmiSetHandler(env, _initHandlerIdx);
DEBUGF(("[%d,%.6lf] RODataMsg being sent of size %d \n",CmiMyPe(),CmiWallTimer(),env->getTotalsize()));
CmiSyncBroadcastAndFree(env->getTotalsize(), (char *)env);
CpvAccess(_qd)->create(CkNumPes()-1);
_initDone();
}
DEBUGF(("[%d,%d%.6lf] inCommThread %d\n",CmiMyPe(),CmiMyRank(),CmiWallTimer(),inCommThread));
// when I am a communication thread, I don't participate initDone.
if (inCommThread) {
CkNumberHandlerEx(_bocHandlerIdx,(CmiHandlerEx)_processHandler,
CkpvAccess(_coreState));
CkNumberHandlerEx(_charmHandlerIdx,(CmiHandlerEx)_processHandler
,
CkpvAccess(_coreState));
_processBufferedMsgs();
}
#if CMK_CHARMDEBUG
// Should not use CpdFreeze inside a thread (since this processor is really a user-level thread)
if (CpvAccess(cpdSuspendStartup))
{
//CmiPrintf("In Parallel Debugging mode .....\n");
CpdFreeze();
}
#endif
#if __FAULT__
if(killFlag){
readKillFile();
}
#endif
}