CcdCallbacks: only call timers when necessary (#3537)

Added CcdCondFn to avoid timers in condition-based callbacks and
changed the return type of CcdRaiseCondition to enable only calling
timers when necessary, decreasing the scheduling overhead. Also only
call CcdCallBacks() when there are periodic callbacks registered.

Define a macro for backwards compatibility, i.e. users can check
"#if CCD_COND_FN_EXISTS" in application code.

benchmarks/charm++/communication_overhead/overhead_test.C

@@ -14,10 +14,10 @@ int nCycles = 100; // Number of iterations for each message size
 
 CProxy_TestDriver testDriverProxy;
 
-void idleStartHandler(void *timingGroupObj, double start);
-void idleEndHandler(void *timingGroupObj, double cur);
-void idleStartHandlerArray(void *timingGroupObj, double start);
-void idleEndHandlerArray(void *timingGroupObj, double cur);
+void idleStartHandler(void *timingGroupObj);
+void idleEndHandler(void *timingGroupObj);
+void idleStartHandlerArray(void *timingGroupObj);
+void idleEndHandlerArray(void *timingGroupObj);
 
 class SimpleMessage: public CMessage_SimpleMessage {
 public:
@@ -357,28 +357,28 @@ public:
 
 };
 
-void idleStartHandler(void *timingGroupObj, double start) {
+void idleStartHandler(void *timingGroupObj) {
   CommunicationGroup *localInstance = (CommunicationGroup *) timingGroupObj;
-  localInstance->startIdleTime = start;
+  localInstance->startIdleTime = CkWallTimer();
 }
 
-void idleEndHandler(void *timingGroupObj, double cur) {
+void idleEndHandler(void *timingGroupObj) {
   CommunicationGroup *localInstance = (CommunicationGroup *) timingGroupObj;
   if(localInstance->startIdleTime > 0) {
-    localInstance->iterationIdleTime += cur - localInstance->startIdleTime;
+    localInstance->iterationIdleTime += CkWallTimer() - localInstance->startIdleTime;
     localInstance->startIdleTime = -1;
   }
 }
 
-void idleStartHandlerArray(void *timingGroupObj, double start) {
+void idleStartHandlerArray(void *timingGroupObj) {
   CommunicationArray *localInstance = (CommunicationArray *) timingGroupObj;
-  localInstance->startIdleTime = start;
+  localInstance->startIdleTime = CkWallTimer();
 }
 
-void idleEndHandlerArray(void *timingGroupObj, double cur) {
+void idleEndHandlerArray(void *timingGroupObj) {
   CommunicationArray *localInstance = (CommunicationArray *) timingGroupObj;
   if(localInstance->startIdleTime > 0) {
-    localInstance->iterationIdleTime += cur - localInstance->startIdleTime;
+    localInstance->iterationIdleTime += CkWallTimer() - localInstance->startIdleTime;
     localInstance->startIdleTime = -1;
   }
 }

benchmarks/converse/machinetest/multiping.C

@@ -49,18 +49,18 @@ CpvDeclare(double, IdleStartTime);
 CpvDeclare(double, IdleTime);
 
 //Register Idle time handlers
-void ApplIdleStart(void *, double start)
+void ApplIdleStart(void *)
 {
-  CpvAccess(IdleStartTime)= start; //CmiWallTimer();
+  CpvAccess(IdleStartTime) = CmiWallTimer();
   return;
 }
 
-void ApplIdleEnd(void *, double cur)
+void ApplIdleEnd(void *)
 {
   if(CpvAccess(IdleStartTime) < 0)
     return;
 
-  CpvAccess(IdleTime) += cur /*CmiWallTimer()*/-CpvAccess(IdleStartTime);
+  CpvAccess(IdleTime) += CmiWallTimer() - CpvAccess(IdleStartTime);
   CpvAccess(IdleStartTime)=-1;
   return;
 }

benchmarks/converse/machinetest/pingall.C

@@ -35,18 +35,18 @@ CpvDeclare(double, IdleStartTime);
 CpvDeclare(double, IdleTime);
 
 //Registering idle handlers
-void ApplIdleStart(void *, double start)
+void ApplIdleStart(void *)
 {
-  CpvAccess(IdleStartTime)= start; //CmiWallTimer();
+  CpvAccess(IdleStartTime) = CmiWallTimer();
   return;
 }
 
-void ApplIdleEnd(void *, double cur)
+void ApplIdleEnd(void *)
 {
   if(CpvAccess(IdleStartTime) < 0)
     return;
 
-  CpvAccess(IdleTime) += cur /*CmiWallTimer()*/-CpvAccess(IdleStartTime);
+  CpvAccess(IdleTime) += CmiWallTimer() - CpvAccess(IdleStartTime);
   CpvAccess(IdleStartTime)=-1;
   return;
 }

doc/charm++/manual.rst

@@ -4395,8 +4395,7 @@ whenidle
    priority or speculative) in the absence of other work. ``whenidle``
    entry methods must return a ``bool`` value, indicating whether the
    entry method should be called when the processor is idle again, and
-   accept a ``double`` argument representing the current timestamp. An
-   example can be found in ``examples/charm++/whenidle``.
+   take no arguments. An example can be found in ``examples/charm++/whenidle``.
 
 python
    entry methods are enabled to be called from python scripts as

doc/converse/manual.rst

@@ -2002,18 +2002,18 @@ CcdUSERMAX
 
 .. code-block:: c++
 
-   int CcdCallOnCondition(int condnum, CcdVoidFn fnp, void* arg)
+   int CcdCallOnCondition(int condnum, CcdCondFn fnp, void* arg)
 
 This call instructs the system to call the function indicated by the
 function pointer ``fnp``, with the specified argument ``arg``, when
 the condition indicated by ``condnum`` is raised next. Multiple
 functions may be registered for the same condition number.
-``CcdVoidFn`` is a function pointer with the signature ``void fnp(void
-*userParam, double curWallTime)``
+``CcdCondFn`` is a function pointer with the signature ``void fnp(void
+*userParam)``
 
 .. code-block:: c++
 
-  int CcdCallOnConditionKeep(int condnum, CcdVoidFn fnp, void* arg)
+  int CcdCallOnConditionKeep(int condnum, CcdCondFn fnp, void* arg)
 
 As above, but the association is permanent- the given function will
 be called again whenever this condition is raised.
@@ -2032,20 +2032,11 @@ callbacks from within ccd callbacks.
 
 .. code-block:: c++
 
-  double CcdRaiseCondition(int condNum)
-
+  void CcdRaiseCondition(int condNum)
 
 When this function is called, it invokes all the functions whose
 pointers were registered for the ``condNum`` via a *prior* call to
-``CcdCallOnCondition`` or ``CcdCallOnConditionKeep``. The function
-internally calls ``CmiWallTimer`` and returns this value. When using
-``CcdRaiseCondition``, the return value can be used to determine the
-current walltime avoiding an additional call to ``CmiWallTimer``.
-However, it is important to note that the walltime value returned
-by ``CcdRaiseCondition`` could be stale by the time it is returned
-since registered functions are executed between the timer call and
-the return. For this reason, this walltime value returned should be
-used in situations where an exact or current timer value is not desired.
+``CcdCallOnCondition`` or ``CcdCallOnConditionKeep``.
 
 .. code-block:: c++
 
@@ -2056,6 +2047,11 @@ be called at least ``msLater`` milliseconds later. The registered
 function ``fnp`` is actually called the first time the scheduler gets
 control after ``deltaT`` milliseconds have elapsed. The default
 polling resolution for timed callbacks is 5 ms.
+``CcdVoidFn`` is a function pointer with the signature ``void fnp(void
+*userParam, double currWallTime)``. Note the extra wall-time parameter,
+which differs from ``CcdCondFn``, because the runtime system internally
+calls a timer in order to invoke periodic callbacks. It passes that
+time into the user callback as well to avoid the need for extra timer calls.
 
 .. code-block:: c++
 

examples/charm++/whenidle/idlework.C

@@ -15,7 +15,7 @@ void Test::registerIdleWork() {
     CkCallWhenIdle(CkIndex_Test::idleProgress(0), this);
 }
 
-bool Test::idleProgress(double time) {
+bool Test::idleProgress() {
     static bool calledBefore = false; 
     CkPrintf("[TEST] Idle Work Called, CalledBefore=%d.\n", (int)calledBefore);
     calledBefore = !calledBefore;

examples/charm++/whenidle/idlework.ci

@@ -6,7 +6,7 @@ mainmodule idlework {
   array [1D] Test {
     entry Test();
     entry void registerIdleWork();
-    entry [whenidle] bool idleProgress(double);
+    entry [whenidle] bool idleProgress();
   };
 }
 

examples/charm++/whenidle/idlework.h

@@ -5,7 +5,7 @@ class Test : public CBase_Test {
 public:
   Test(CkMigrateMessage *m) {}
   Test();
-  bool idleProgress(double time);
+  bool idleProgress();
   void registerIdleWork();
 };
 

src/arch/cuda/hybridAPI/hapi_impl.cpp

@@ -145,7 +145,7 @@ void hapiInit(char** argv) {
 
 #ifndef HAPI_CUDA_CALLBACK
     // Register polling function to be invoked at every scheduler loop
-    CcdCallOnConditionKeep(CcdSCHEDLOOP, hapiPollEvents, NULL);
+    CcdCallOnConditionKeep(CcdSCHEDLOOP, (CcdCondFn)hapiPollEvents, NULL);
 #endif
   }
 
@@ -1381,7 +1381,7 @@ void hapiClearInstrument() {
 // all successive completed events in the queue starting from the front.
 // TODO Maybe we should make one pass of all events in the queue instead,
 // since there might be completed events later in the queue.
-void hapiPollEvents(void* param, double cur_time) {
+void hapiPollEvents(void* param) {
 #ifndef HAPI_CUDA_CALLBACK
   if (CpvAccess(n_hapi_events) <= 0) return;
 

src/arch/cuda/hybridAPI/hapi_impl.h

@@ -18,7 +18,7 @@ void hapiInit(char** argv);
 void hapiExit();
 
 // Polls for GPU work completion. Does not do anything if HAPI_CUDA_CALLBACK is defined.
-void hapiPollEvents(void* param, double cur_time);
+void hapiPollEvents(void* param);
 
 // BufferPool constructs for mempool implementation.
 // Data and metadata reside in same chunk of memory.

src/arch/gni/machine.C

@@ -1900,19 +1900,19 @@ void LrtsPostCommonInit(int everReturn)
 
 #if !CMK_SMP
     if (useDynamicSMSG)
-    CcdCallOnConditionKeep(CcdPERIODIC_10ms, (CcdVoidFn) PumpDatagramConnection, NULL);
+    CcdCallOnConditionKeep(CcdPERIODIC_10ms, (CcdCondFn) PumpDatagramConnection, NULL);
 #endif
 
 #if ! LARGEPAGE
     if (_checkProgress)
 #if CMK_SMP
     if (CmiMyRank() == 0)
 #endif
-    CcdCallOnConditionKeep(CcdPERIODIC_2minute, (CcdVoidFn) CheckProgress, NULL);
+    CcdCallOnConditionKeep(CcdPERIODIC_2minute, (CcdCondFn) CheckProgress, NULL);
 #endif
 
 #if !LARGEPAGE
-    CcdCallOnCondition(CcdTOPOLOGY_AVAIL, (CcdVoidFn)set_limit, NULL);
+    CcdCallOnCondition(CcdTOPOLOGY_AVAIL, (CcdCondFn)set_limit, NULL);
 #endif
 }
 

src/arch/netlrts/machine.C

@@ -1863,14 +1863,14 @@ void LrtsPostCommonInit(int everReturn)
 #if CMK_SHARED_VARS_UNAVAILABLE
   if (Cmi_netpoll) /*Repeatedly call CommServer*/
     CcdCallOnConditionKeep(CcdPERIODIC, 
-        (CcdVoidFn) CommunicationPeriodic, NULL);
+        (CcdCondFn) CommunicationPeriodic, NULL);
   else /*Only need this for retransmits*/
     CcdCallOnConditionKeep(CcdPERIODIC_10ms, 
-        (CcdVoidFn) CommunicationPeriodic, NULL);
+        (CcdCondFn) CommunicationPeriodic, NULL);
 #endif
 
   if (CmiMyRank()==0 && Cmi_charmrun_fd!=-1) {
-    CcdCallOnConditionKeep(CcdPERIODIC_10ms, (CcdVoidFn) CmiStdoutFlush, NULL);
+    CcdCallOnConditionKeep(CcdPERIODIC_10ms, (CcdCondFn) CmiStdoutFlush, NULL);
 #if CMK_SHARED_VARS_UNAVAILABLE && !CMK_TIMER_USE_WIN32API
     if (!Cmi_asyncio) {
     CcdCallFnAfter((CcdVoidFn)pingCharmrunPeriodic,NULL,1000);
@@ -1909,7 +1909,7 @@ void LrtsPostCommonInit(int everReturn)
   /* Call the function to periodically call the token adapt function */
   CcdCallFnAfter((CcdVoidFn)TokenUpdatePeriodic, NULL, 2000); // magic number of 2000ms
   CcdCallOnConditionKeep(CcdPERIODIC_10s,   // magic number of PERIOD 10s
-        (CcdVoidFn) TokenUpdatePeriodic, NULL);
+        (CcdCondFn) TokenUpdatePeriodic, NULL);
 #endif
 
 #ifdef CMK_RANDOMLY_CORRUPT_MESSAGES

src/arch/pami/machine.C

@@ -856,7 +856,7 @@ void PerrorExit (const char *err) {
     exit (-1);
 }
 
-static void CmiNotifyIdleCcd(void *ignored1, double ignored2)
+static void CmiNotifyIdleCcd(void *ignored)
 {
     CmiNotifyIdle();
 }

src/arch/util/machine-common-core.C

@@ -1580,12 +1580,12 @@ static void ConverseRunPE(int everReturn) {
 #if CMK_SMP
     {
       CmiIdleState *sidle=CmiNotifyGetState();
-      CcdCallOnConditionKeep(CcdPROCESSOR_BEGIN_IDLE,(CcdVoidFn)CmiNotifyBeginIdle,(void *)sidle);
-      CcdCallOnConditionKeep(CcdPROCESSOR_STILL_IDLE,(CcdVoidFn)CmiNotifyStillIdle,(void *)sidle);
+      CcdCallOnConditionKeep(CcdPROCESSOR_BEGIN_IDLE,(CcdCondFn)CmiNotifyBeginIdle,(void *)sidle);
+      CcdCallOnConditionKeep(CcdPROCESSOR_STILL_IDLE,(CcdCondFn)CmiNotifyStillIdle,(void *)sidle);
     }
 #else
-    CcdCallOnConditionKeep(CcdPROCESSOR_BEGIN_IDLE,(CcdVoidFn)CmiNotifyBeginIdle, NULL);
-    CcdCallOnConditionKeep(CcdPROCESSOR_STILL_IDLE,(CcdVoidFn)CmiNotifyStillIdle, NULL);
+    CcdCallOnConditionKeep(CcdPROCESSOR_BEGIN_IDLE,(CcdCondFn)CmiNotifyBeginIdle, NULL);
+    CcdCallOnConditionKeep(CcdPROCESSOR_STILL_IDLE,(CcdCondFn)CmiNotifyStillIdle, NULL);
 #endif
 
 

src/arch/verbs/machine.C

@@ -1860,14 +1860,14 @@ void LrtsPostCommonInit(int everReturn)
 #if CMK_SHARED_VARS_UNAVAILABLE
   if (Cmi_netpoll) /*Repeatedly call CommServer*/
     CcdCallOnConditionKeep(CcdPERIODIC, 
-        (CcdVoidFn) CommunicationPeriodic, NULL);
+        (CcdCondFn) CommunicationPeriodic, NULL);
   else /*Only need this for retransmits*/
     CcdCallOnConditionKeep(CcdPERIODIC_10ms, 
-        (CcdVoidFn) CommunicationPeriodic, NULL);
+        (CcdCondFn) CommunicationPeriodic, NULL);
 #endif
 
   if (CmiMyRank()==0 && Cmi_charmrun_fd!=-1) {
-    CcdCallOnConditionKeep(CcdPERIODIC_10ms, (CcdVoidFn) CmiStdoutFlush, NULL);
+    CcdCallOnConditionKeep(CcdPERIODIC_10ms, (CcdCondFn) CmiStdoutFlush, NULL);
 #if CMK_SHARED_VARS_UNAVAILABLE
     if (!Cmi_asyncio) {
     /* gm cannot live with setitimer */
@@ -1908,7 +1908,7 @@ void LrtsPostCommonInit(int everReturn)
   /* Call the function to periodically call the token adapt function */
   CcdCallFnAfter((CcdVoidFn)TokenUpdatePeriodic, NULL, 2000); // magic number of 2000ms
   CcdCallOnConditionKeep(CcdPERIODIC_10s,   // magic number of PERIOD 10s
-        (CcdVoidFn) TokenUpdatePeriodic, NULL);
+        (CcdCondFn) TokenUpdatePeriodic, NULL);
 #endif
 
 #ifdef CMK_RANDOMLY_CORRUPT_MESSAGES

src/ck-core/ck.C

@@ -2771,8 +2771,8 @@ private:
   }
 };
 
-void CkMessageReplayQuiescence(void *rep, double time);
-void CkMessageDetailReplayDone(void *rep, double time);
+void CkMessageReplayQuiescence(void *rep);
+void CkMessageDetailReplayDone(void *rep);
 
 class CkMessageReplay : public CkMessageWatcher {
   int counter;
@@ -2899,7 +2899,7 @@ public:
 	  getNext();
 	  REPLAYDEBUG("Constructing ckMessageReplay: "<< nextPE <<" "<< nextSize <<" "<<nextEvent);
 #if CMI_QD
-	  if (CkMyPe()==0) CmiStartQD(CkMessageReplayQuiescence, this);
+	  if (CkMyPe()==0) CmiStartQD((CcdCondFn)CkMessageReplayQuiescence, this);
 #endif
 	}
 	~CkMessageReplay() {fclose(f);}
@@ -2998,7 +2998,7 @@ public:
 
     CsdEnqueue(getNext());
 
-    CcdCallOnCondition(CcdPROCESSOR_STILL_IDLE, (CcdVoidFn)CkMessageDetailReplayDone, (void*)this);
+    CcdCallOnCondition(CcdPROCESSOR_STILL_IDLE, (CcdCondFn)CkMessageDetailReplayDone, (void*)this);
   }
   virtual bool process(envelope **env,CkCoreState *ck) {
     void *msg = getNext();
@@ -3007,13 +3007,13 @@ public:
   }
 };
 
-void CkMessageReplayQuiescence(void *rep, double time) {
+void CkMessageReplayQuiescence(void *rep) {
   CkPrintf("[%d] Quiescence detected\n",CkMyPe());
   CkMessageReplay *replay = (CkMessageReplay*)rep;
-  //CmiStartQD(CkMessageReplayQuiescence, replay);
+  //CmiStartQD((CcdCondFn)CkMessageReplayQuiescence, replay);
 }
 
-void CkMessageDetailReplayDone(void *rep, double time) {
+void CkMessageDetailReplayDone(void *rep) {
   CkMessageDetailReplay *replay = (CkMessageDetailReplay *)rep;
   CkPrintf("[%d] Detailed replay finished after %f seconds. Exiting.\n",CkMyPe(),CkWallTimer()-replay->starttime);
   ConverseExit();

src/ck-core/ckarray.C

@@ -529,7 +529,7 @@ inline void CkArray::springCleaning(void)
   setupSpringCleaning();
 }
 
-void CkArray::staticSpringCleaning(void* forArray, double curWallTime)
+void CkArray::staticSpringCleaning(void* forArray)
 {
   ((CkArray*)forArray)->springCleaning();
 }
@@ -539,7 +539,7 @@ void CkArray::setupSpringCleaning()
   // set up broadcast cleaner
   if (!stableLocations)
     springCleaningCcd =
-        CcdCallOnCondition(CcdPERIODIC_1minute, staticSpringCleaning, (void*)this);
+        CcdCallOnCondition(CcdPERIODIC_1minute, (CcdCondFn)CkArray::staticSpringCleaning, (void*)this);
 }
 
 /********************* Little CkArray Utilities ******************/

src/ck-core/ckarray.h

@@ -860,7 +860,7 @@ class CkArray : public CkReductionMgr
 
   // Spring cleaning
   void springCleaning(void);
-  static void staticSpringCleaning(void* forWhom, double curWallTime);
+  static void staticSpringCleaning(void* forWhom);
   void setupSpringCleaning();
   int springCleaningCcd;