void LBMachineUtil::StatsOn()
{
const double cur_wall = CkWallTimer();
#if CMK_LB_CPUTIMER
const double cur_cpu = CkCpuTimer();
#endif
if (state == off) {
#if ! CMK_BIGSIM_CHARM
cancel_idleStart=CcdCallOnConditionKeep(
CcdPROCESSOR_BEGIN_IDLE,(CcdVoidFn)staticIdleStart,(void *)this);
cancel_idleEnd=CcdCallOnConditionKeep(
CcdPROCESSOR_END_IDLE,(CcdVoidFn)staticIdleEnd,(void *)this);
#endif
state = on;
}
if (start_totalwall != -1.) {
total_walltime += (cur_wall - start_totalwall);
#if CMK_LB_CPUTIMER
total_cputime += (cur_cpu - start_totalcpu);
#endif
}
start_totalwall = cur_wall;
#if CMK_LB_CPUTIMER
start_totalcpu = cur_cpu;
#endif
}
/**
* Initialize the callback containers
*/
void CcdModuleInit(void)
{
int i;
double curTime;
CpvInitialize(ccd_heap_elem*, ccd_heap);
CpvInitialize(ccd_cond_callbacks, conds);
CpvInitialize(ccd_periodic_callbacks, pcb);
CpvInitialize(int, ccd_heaplen);
CpvInitialize(int, ccd_heapmaxlen);
CpvInitialize(int, _ccd_numchecks);
CpvAccess(ccd_heaplen) = 0;
CpvAccess(ccd_heapmaxlen) = MAXTIMERHEAPENTRIES;
CpvAccess(ccd_heap) =
(ccd_heap_elem*) malloc(sizeof(ccd_heap_elem)*2*(MAXTIMERHEAPENTRIES + 1));
_MEMCHECK(CpvAccess(ccd_heap));
for(i=0; i<MAXNUMCONDS; i++) {
init_cblist(&(CpvAccess(conds).condcb[i]), CBLIST_INIT_LEN);
init_cblist(&(CpvAccess(conds).condcb_keep[i]), CBLIST_INIT_LEN);
}
CpvAccess(_ccd_numchecks) = 1;
CpvAccess(pcb).nSkip = 1;
curTime=CmiWallTimer();
CpvAccess(pcb).lastCheck = curTime;
for (i=0; i<CCD_PERIODIC_MAX; i++)
CpvAccess(pcb).nextCall[i]=curTime+periodicCallInterval[i];
CcdCallOnConditionKeep(CcdPROCESSOR_BEGIN_IDLE,CcdCallBacksReset,0);
CcdCallOnConditionKeep(CcdPROCESSOR_END_IDLE,CcdCallBacksReset,0);
}
static void MachinePostCommonInitForLAPI(int everReturn) {
/**
* In SMP, the machine layer usually has one comm thd, and it is
* designed to be responsible for all network communication. So
* if there's no dedicated processor for the comm thread, it has
* to share a proc with a worker thread. In this scenario,
* the worker thread needs to yield for some time to give CPU
* time to comm thread. However, in current configuration, we
* will always dedicate one proc for the comm thd, therefore,
* such yielding scheme is not necessary. Besides, avoiding
* this yielding scheme improves performance because worker
* thread doesn't need to yield and will be more responsive to
* incoming messages. So, we will always use CmiNotifyIdle
* instead.
*
* --Chao Mei
*/
/* Not registering any Idle-state related functions right now!! */
#if 0 && CMK_SMP
s=CmiNotifyGetState();
CcdCallOnConditionKeep(CcdPROCESSOR_BEGIN_IDLE,(CcdVoidFn)CmiNotifyBeginIdle,(void *)s);
CcdCallOnConditionKeep(CcdPROCESSOR_STILL_IDLE,(CcdVoidFn)CmiNotifyStillIdle,(void *)s);
#else
CcdCallOnConditionKeep(CcdPROCESSOR_STILL_IDLE,(CcdVoidFn)CmiNotifyIdle,NULL);
#if !CMK_SMP || CMK_SMP_NO_COMMTHD
/* If there's comm thread, then comm thd is responsible for advancing comm */
if (!CsvAccess(lapiInterruptMode)) {
CcdCallOnConditionKeep(CcdPERIODIC_10ms, (CcdVoidFn)AdvanceCommunicationForLAPI, NULL);
}
#endif
#endif
}
CmiStartFn mymain(int argc, char** argv)
{
if(CmiMyRank() == CmiMyNodeSize()) return 0;
CpvInitialize(int,msgSize);
CpvInitialize(int,cycleNum);
CpvInitialize(int,sizeNum);
CpvAccess(sizeNum) = 1;
CpvAccess(msgSize)= CmiMsgHeaderSizeBytes + 8;
CpvInitialize(int,exitHandler);
CpvAccess(exitHandler) = CmiRegisterHandler((CmiHandler) exitHandlerFunc);
CpvInitialize(int,node0Handler);
CpvAccess(node0Handler) = CmiRegisterHandler((CmiHandler) node0HandlerFunc);
CpvInitialize(int,node1Handler);
CpvAccess(node1Handler) = CmiRegisterHandler((CmiHandler) node1HandlerFunc);
CpvInitialize(int,ackHandler);
CpvAccess(ackHandler) = CmiRegisterHandler((CmiHandler) ackHandlerFunc);
CpvInitialize(double,startTime);
CpvInitialize(double,endTime);
CpvInitialize(double, IdleStartTime);
CpvInitialize(double, IdleTime);
CpvInitialize(int,ackCount);
CpvAccess(ackCount) = 0;
CpvInitialize(int,twoway);
CpvAccess(twoway) = 0;
CcdCallOnConditionKeep(CcdPROCESSOR_BEGIN_IDLE, ApplIdleStart, NULL);
CcdCallOnConditionKeep(CcdPROCESSOR_END_IDLE, ApplIdleEnd, NULL);
if(argc > 1)
CpvAccess(twoway) = atoi(argv[1]);
if(CmiMyPe() == 0) {
if(!CpvAccess(twoway))
CmiPrintf("Starting Pingpong with oneway traffic \n");
else
CmiPrintf("Starting Pingpong with twoway traffic\n");
}
if ((CmiMyPe() < CmiNumPes()/2) || CpvAccess(twoway))
startPingpong();
return 0;
}
void CcsInit(char **argv)
{
CpvInitialize(CkHashtable_c, ccsTab);
CpvAccess(ccsTab) = CkCreateHashtable_string(sizeof(CcsHandlerRec),5);
CpvInitialize(CcsImplHeader *, ccsReq);
CpvAccess(ccsReq) = NULL;
_ccsHandlerIdx = CmiRegisterHandler((CmiHandler)req_fw_handler);
#if CMK_BIGSIM_CHARM
CpvInitialize(int, _bgCcsHandlerIdx);
CpvAccess(_bgCcsHandlerIdx) = 0;
CpvInitialize(int, _bgCcsAck);
CpvAccess(_bgCcsAck) = 0;
#endif
CpvInitialize(int, cmiArgDebugFlag);
CpvInitialize(char *, displayArgument);
CpvInitialize(int, cpdSuspendStartup);
CpvAccess(cmiArgDebugFlag) = 0;
CpvAccess(displayArgument) = NULL;
CpvAccess(cpdSuspendStartup) = 0;
CcsBuiltinsInit(argv);
rep_fw_handler_idx = CmiRegisterHandler((CmiHandler)rep_fw_handler);
#if NODE_0_IS_CONVHOST
#if ! CMK_CMIPRINTF_IS_A_BUILTIN
print_fw_handler_idx = CmiRegisterHandler((CmiHandler)print_fw_handler);
#endif
{
int ccs_serverPort=0;
char *ccs_serverAuth=NULL;
if (CmiGetArgFlagDesc(argv,"++server", "Create a CCS server port") |
CmiGetArgIntDesc(argv,"++server-port",&ccs_serverPort, "Listen on this TCP/IP port number") |
CmiGetArgStringDesc(argv,"++server-auth",&ccs_serverAuth, "Use this CCS authentication file"))
if (CmiMyPe()==0)
{/*Create and occasionally poll on a CCS server port*/
CcsServer_new(NULL,&ccs_serverPort,ccs_serverAuth);
CcdCallOnConditionKeep(CcdPERIODIC,(CcdVoidFn)CcsServerCheck,NULL);
}
}
#endif
/* if in parallel debug mode i.e ++cpd, freeze */
if (CmiGetArgFlagDesc(argv, "+cpd", "Used *only* in conjunction with parallel debugger"))
{
CpvAccess(cmiArgDebugFlag) = 1;
if (CmiGetArgStringDesc(argv, "+DebugDisplay",&(CpvAccess(displayArgument)), "X display for gdb used only in cpd mode"))
{
if (CpvAccess(displayArgument) == NULL)
CmiPrintf("WARNING> NULL parameter for +DebugDisplay\n***");
}
else if (CmiMyPe() == 0)
{
/* only one processor prints the warning */
CmiPrintf("WARNING> x term for gdb needs to be specified as +DebugDisplay by debugger\n***\n");
}
if (CmiGetArgFlagDesc(argv, "+DebugSuspend", "Suspend execution at beginning of program")) {
CpvAccess(cpdSuspendStartup) = 1;
}
}
CcsReleaseMessages();
}
TempAwareRefineLB::TempAwareRefineLB(const CkLBOptions &opt): CentralLB(opt)
{
#ifdef TEMP_LDB
starting=CmiWallTimer();
// procsPerNode=4;
migFile=fopen("migInfo","w");
numAvailFreqs = 11;
//numAvailFreqs = 14;
//numAvailFreqs = 7;
freqs=new int[numAvailFreqs];
freqsEffect=new int[numAvailFreqs];
// for might (lab machine)
/*
freqs[0] = 2262000;
freqs[1] = 2261000;
freqs[2] = 2128000;
freqs[3] = 1995000;
freqs[4] = 1862000;
freqs[5] = 1729000;
freqs[6] = 1596000;
*/
// for tarekc cluster
freqs[0] = 2395000;
freqs[1] = 2394000;
freqs[2] = 2261000;
freqs[3] = 2128000;
freqs[4] = 1995000;
freqs[5] = 1862000;
freqs[6] = 1729000;
freqs[7] = 1596000;
freqs[8] = 1463000;
freqs[9] = 1330000;
freqs[10] = 1197000;
freqsEffect[0] = 1979886;
freqsEffect[1] = 1943017;
freqsEffect[2] = 1910989;
freqsEffect[3] = 1876619;
freqsEffect[4] = 1824126;
freqsEffect[5] = 1763990;
freqsEffect[6] = 1666773;
freqsEffect[7] = 1560224;
freqsEffect[8] = 1443154;
freqsEffect[9] = 1317009;
freqsEffect[10] = 1200000;
/*
// for grace, humility etc (lab i7 machines)
freqs[0] = 2801000;
freqs[1] = 2800000;
freqs[2] = 2667000;
freqs[3] = 2533000;
freqs[4] = 2400000;
freqs[5] = 2267000;
freqs[6] = 2133000;
freqs[7] = 2000000;
freqs[8] = 1867000;
freqs[9] = 1733000;
freqs[10] = 1600000;
freqs[11] = 1467000;
freqs[12] = 1333000;
freqs[13] = 1200000;
*/
procFreqPtr = new int[CkNumPes()];
for(int i=0;i<CkNumPes();i++)
{
char newfreq[10];
sprintf(newfreq,"%d",freqs[0]);
cpufreq_sysfs_write(newfreq,i%physicalCoresPerNode);
procFreqPtr[i]=0;
}
// logicalCoresPerChip=4;
procFreq=NULL;
procTemp=NULL;
procFreqNew=NULL;
procFreqNewEffect = NULL;
avgChipTemp=NULL;
lbname = "TempAwareRefineLB";
if (CkMyPe()==0)
CkPrintf("[%d] TempAwareRefineLB created\n",CkMyPe());
char logFile[100];
snprintf(logFile, sizeof(logFile), "temp_freq.log.%d", CkMyPe());
if ((logFD = fopen(logFile, "a"))) {
fprintf(logFD, "Time, PE, Temperature, Frequency\n");
} else {
CkAbort("Couldn't open temperature/frequency log file");
}
CcdCallOnConditionKeep(CcdPERIODIC_1second, &printCurrentTemperature, this);
#else
CmiAbort("TEMPLB ERROR: not supported without TEMP_LDB flag.\n");
#endif
}
TempAwareCommLB::TempAwareCommLB(const CkLBOptions &opt): CBase_TempAwareCommLB(opt)
{
#ifdef TEMP_LDB
lbname = "TempAwareCommLB";
if (CkMyPe()==0)
CkPrintf("[%d] TempAwareCommLB created\n",CkMyPe());
starting=CmiWallTimer();
migFile=fopen("migInfo","w");
numAvailFreqs = 11;
freqs=new int[numAvailFreqs];
freqsEffect=new int[numAvailFreqs];
// for tarekc cluster
freqs[0] = 2395000;
freqs[1] = 2394000;
freqs[2] = 2261000;
freqs[3] = 2128000;
freqs[4] = 1995000;
freqs[5] = 1862000;
freqs[6] = 1729000;
freqs[7] = 1596000;
freqs[8] = 1463000;
freqs[9] = 1330000;
freqs[10] = 1197000;
freqsEffect[0] = 1979886;
freqsEffect[1] = 1943017;
freqsEffect[2] = 1910989;
freqsEffect[3] = 1876619;
freqsEffect[4] = 1824126;
freqsEffect[5] = 1763990;
freqsEffect[6] = 1666773;
freqsEffect[7] = 1560224;
freqsEffect[8] = 1443154;
freqsEffect[9] = 1317009;
freqsEffect[10] = 1200000;
procFreqPtr = new int[CkNumPes()];
for(int i=0;i<CkNumPes();i++)
{
char newfreq[10];
sprintf(newfreq,"%d",freqs[0]);
cpufreq_sysfs_write(newfreq,i%physicalCoresPerNode);
procFreqPtr[i]=0;
}
procFreq=NULL;
procTemp=NULL;
procFreqNew=NULL;
procFreqNewEffect = NULL;
avgChipTemp=NULL;
char logFile[100];
snprintf(logFile, sizeof(logFile), "temp_freq.log.%d", CkMyPe());
if ((logFD = fopen(logFile, "a"))) {
fprintf(logFD, "Time, PE, Temperature, Frequency\n");
} else {
CkAbort("Couldn't open temperature/frequency log file");
}
CcdCallOnConditionKeep(CcdPERIODIC_1second, &printCurrentTemperature, this);
#else
#endif
}