Adding Tenure bytes deviation to modify concurrent kickoff

- Added heuristics to keep track of bytes moving from nursery to tenure, along with the deviation of this value.
- `_avgTenureBytes` keeps track of average bytes that are being tenured, while `_avgTenureBytesDeviation` keeps track of the average amount of deviation in bytes being tenured.
- Adding this deviation into the calculation for concurrent kickoff
- The default boost for the deviation on top of the existing kickoff logic is 2x. This field is called `tenureBytesDeviationBoost`
- This change results in far less aborts/percolates, but at the expense of more global GC cycles
- Will be accompanied by a change on the j9 side for explicitly setting `tenureBytesDeviationBoost`. Setting this to 0 will effectively disable it's effects
- This calculation is done for both SOA bytes (if OMR_GC_LARGE_OBJECT_AREA flag is set), and tenured bytes in general if not set

Signed-off-by: Cedric Hansen <cedric.hansen@ibm.com>

gc/base/GCExtensionsBase.hpp

@@ -267,6 +267,7 @@ class MM_GCExtensionsBase : public MM_BaseVirtual {
 	uintptr_t oldHeapSizeOnLastGlobalGC;
 	uintptr_t freeOldHeapSizeOnLastGlobalGC;
 	float concurrentKickoffTenuringHeadroom; /**< percentage of free memory remaining in tenure heap. Used in conjunction with free memory to determine concurrent mark kickoff */
+	float tenureBytesDeviationBoost; /**< boost factor for tenuring deviation used for concurrent mark kickoff math */
 #endif /* OMR_GC_MODRON_SCAVENGER */
 #if defined(OMR_GC_REALTIME)
 	MM_RememberedSetSATB* sATBBarrierRememberedSet; /**< The snapshot at the beginning barrier remembered set used for the write barrier */
@@ -1348,6 +1349,7 @@ class MM_GCExtensionsBase : public MM_BaseVirtual {
 		, oldHeapSizeOnLastGlobalGC(UDATA_MAX)
 		, freeOldHeapSizeOnLastGlobalGC(UDATA_MAX)
 		, concurrentKickoffTenuringHeadroom((float)0.02)
+		, tenureBytesDeviationBoost((float)2)
 #endif /* OMR_GC_MODRON_SCAVENGER */		
 #if defined(OMR_GC_REALTIME)
 		, sATBBarrierRememberedSet(NULL)

gc/base/standard/ConcurrentGC.cpp

@@ -1836,12 +1836,13 @@ MM_ConcurrentGC::potentialFreeSpace(MM_EnvironmentBase *env, MM_AllocateDescript
 			headRoom = (uintptr_t)(_extensions->concurrentKickoffTenuringHeadroom * _extensions->lastGlobalGCFreeBytesLOA);
 		} else {
 			assume0(SOA == _meteringType);
-			nurseryPromotion = scavengerStats->_avgTenureSOABytes == 0 ? 1 : scavengerStats->_avgTenureSOABytes;
+			nurseryPromotion = scavengerStats->_avgTenureSOABytes == 0 ? 1 : (uintptr_t)(scavengerStats->_avgTenureSOABytes + (env->getExtensions()->tenureBytesDeviationBoost * scavengerStats->_avgTenureBytesDeviation)); 
+
 			currentOldFree = oldSubspace->getApproximateActiveFreeMemorySize() - oldSubspace->getApproximateActiveFreeLOAMemorySize();
 			headRoom = (uintptr_t)(_extensions->concurrentKickoffTenuringHeadroom * (_extensions->getLastGlobalGCFreeBytes() - _extensions->lastGlobalGCFreeBytesLOA));
 		}
 #else
-		nurseryPromotion = scavengerStats->_avgTenureBytes == 0 ? 1: scavengerStats->_avgTenureBytes;
+		nurseryPromotion = scavengerStats->_avgTenureBytes == 0 ? 1: (uintptr_t)(scavengerStats->_avgTenureBytes + (env->getExtensions()->tenureBytesDeviationBoost * scavengerStats->_avgTenureBytesDeviation));
 		currentOldFree = oldSubspace->getApproximateActiveFreeMemorySize();
 		headRoom = (uintptr_t)(_extensions->concurrentKickoffTenuringHeadroom * _extensions->getLastGlobalGCFreeBytes()); 
 #endif

gc/base/standard/Scavenger.cpp

@@ -832,10 +832,27 @@ MM_Scavenger::calcGCStats(MM_EnvironmentStandard *env)
 		if (scavengerGCStats->_gcCount > 1 ) {
 			scavengerGCStats->_avgInitialFree = (uintptr_t)MM_Math::weightedAverage((float)scavengerGCStats->_avgInitialFree, (float)initialFree, INITIAL_FREE_HISTORY_WEIGHT);
 			scavengerGCStats->_avgTenureBytes = (uintptr_t)MM_Math::weightedAverage((float)scavengerGCStats->_avgTenureBytes, (float)scavengerGCStats->_tenureAggregateBytes, TENURE_BYTES_HISTORY_WEIGHT);
+			scavengerGCStats->_tenureBytesDeviation = (float)(scavengerGCStats->_tenureAggregateBytes) - (scavengerGCStats->_avgTenureBytes);
+			scavengerGCStats->_avgTenureBytesDeviation = (uintptr_t)sqrtf(MM_Math::weightedAverage((float)(scavengerGCStats->_avgTenureBytesDeviation * scavengerGCStats->_avgTenureBytesDeviation), 
+																						(float)(scavengerGCStats->_tenureBytesDeviation * scavengerGCStats->_tenureBytesDeviation), TENURE_BYTES_HISTORY_WEIGHT));
+
+			if (_extensions->debugConcurrentMark) {
+				OMRPORT_ACCESS_FROM_OMRPORT(env->getPortLibrary());
+            	omrtty_printf("Tenure bytes: %d\n_avgTenureBytes: %zu\n_tenureBytesDeviation: %f\n_avgTenureBytesDeviation: %zu\n",scavengerGCStats->_tenureAggregateBytes, scavengerGCStats->_avgTenureBytes, scavengerGCStats->_tenureBytesDeviation, scavengerGCStats->_avgTenureBytesDeviation);
+			}	
+
 #if defined(OMR_GC_LARGE_OBJECT_AREA)
 			scavengerGCStats->_avgTenureSOABytes = (uintptr_t)MM_Math::weightedAverage((float)scavengerGCStats->_avgTenureSOABytes,
 																		(float)(scavengerGCStats->_tenureAggregateBytes - scavengerGCStats->_tenureLOABytes),
-																		TENURE_BYTES_HISTORY_WEIGHT);
+																		TENURE_BYTES_HISTORY_WEIGHT);	
+
+			scavengerGCStats->_tenureBytesDeviation = (float)(scavengerGCStats->_tenureAggregateBytes - scavengerGCStats->_tenureLOABytes) - (scavengerGCStats->_avgTenureBytes);																														
+			scavengerGCStats->_avgTenureBytesDeviation = (uintptr_t)sqrtf(MM_Math::weightedAverage((float)(scavengerGCStats->_avgTenureBytesDeviation * scavengerGCStats->_avgTenureBytesDeviation), 
+																						(float)(scavengerGCStats->_tenureBytesDeviation * scavengerGCStats->_tenureBytesDeviation), TENURE_BYTES_HISTORY_WEIGHT));
+			if (_extensions->debugConcurrentMark) {
+				OMRPORT_ACCESS_FROM_OMRPORT(env->getPortLibrary());
+            	omrtty_printf("SOA bytes: %d\n_avgTenureBytes: %zu\n_tenureBytesDeviation: %f\n_avgTenureBytesDeviation: %zu\n",(scavengerGCStats->_tenureAggregateBytes - scavengerGCStats->_tenureLOABytes), scavengerGCStats->_avgTenureBytes,scavengerGCStats->_tenureBytesDeviation, scavengerGCStats->_avgTenureBytesDeviation);
+			}																
 			scavengerGCStats->_avgTenureLOABytes = (uintptr_t)MM_Math::weightedAverage((float)scavengerGCStats->_avgTenureLOABytes,
 																		(float)scavengerGCStats->_tenureLOABytes,
 																		TENURE_BYTES_HISTORY_WEIGHT);

gc/stats/ScavengerStats.cpp

@@ -67,11 +67,15 @@ MM_ScavengerStats::MM_ScavengerStats()
 #endif /* J9MODRON_TGC_PARALLEL_STATISTICS */
 	,_avgInitialFree(0)
 	,_avgTenureBytes(0)
+	,_avgTenureBytesDeviation(0)
+	,_tenureBytesDeviation((float)0)
 	,_tiltRatio(0)
 	,_nextScavengeWillPercolate(false)
 #if defined(OMR_GC_LARGE_OBJECT_AREA)	
 	,_avgTenureLOABytes(0)
 	,_avgTenureSOABytes(0)
+	,_avgTenureSOABytesDeviation(0)
+	,_tenureSOABytesDeviation((float)0)
 #endif /* OMR_GC_LARGE_OBJECT_AREA */
 	,_flipDiscardBytes(0)
 	,_tenureDiscardBytes(0)

gc/stats/ScavengerStats.hpp

@@ -103,6 +103,8 @@ class MM_ScavengerStats
 	 */ 
 	uintptr_t _avgInitialFree;
 	uintptr_t _avgTenureBytes;
+	uintptr_t _avgTenureBytesDeviation; /**< The average, weighted deviation of the tenureBytes*/
+	float _tenureBytesDeviation; /**< The amount by which the previous tenureBytes deviated from _avgTenureBytes */
 
 	uintptr_t _tiltRatio;	/**< use to pass tiltRatio to verbose */