Add the randomized allocation sampling feature

This feature allows profilers to do allocation profiling based off randomized samples. It has better theoretical and empirically observed accuracy than our current allocation profiling approaches while also maintaining low performance overhead. It is designed for use in production profiling scenarios. For more information about usage and implementation, see the included doc docs/design/features/RandomizedAllocationSampling.md

Author: noahfalk

docs/design/features/RandomizedAllocationSampling.md

@@ -1333,17 +1333,19 @@ namespace ETW
 #define ETWLoaderStaticLoad 0 // Static reference load
 #define ETWLoaderDynamicLoad 1 // Dynamic assembly load
 
+#if defined (FEATURE_EVENT_TRACE)
+EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_PROVIDER_DOTNET_Context;
+EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_PRIVATE_PROVIDER_DOTNET_Context;
+EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_RUNDOWN_PROVIDER_DOTNET_Context;
+EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_STRESS_PROVIDER_DOTNET_Context;
+#endif // FEATURE_EVENT_TRACE
+
 #if defined(FEATURE_EVENT_TRACE) && !defined(HOST_UNIX)
 //
 // The ONE and only ONE global instantiation of this class
 //
 extern ETW::CEtwTracer *  g_pEtwTracer;
 
-EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_PROVIDER_DOTNET_Context;
-EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_PRIVATE_PROVIDER_DOTNET_Context;
-EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_RUNDOWN_PROVIDER_DOTNET_Context;
-EXTERN_C DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_STRESS_PROVIDER_DOTNET_Context;
-
 //
 // Special Handling of Startup events
 //

src/coreclr/inc/eventtracebase.h

@@ -3,6 +3,7 @@ set(SOURCES
     dn-u16.cpp
     ${CLR_SRC_NATIVE_DIR}/minipal/utf8.c
     ${CLR_SRC_NATIVE_DIR}/minipal/time.c
+    ${CLR_SRC_NATIVE_DIR}/minipal/xoshiro128pp.c
 )
 
 if(NOT CLR_CROSS_COMPONENTS_BUILD)

src/coreclr/minipal/Windows/CMakeLists.txt

@@ -29,6 +29,12 @@
 
 #include "gcdesc.h"
 
+#ifdef FEATURE_EVENT_TRACE
+   #include "clretwallmain.h"
+#else // FEATURE_EVENT_TRACE
+   #include "etmdummy.h"
+#endif // FEATURE_EVENT_TRACE
+
 #define RH_LARGE_OBJECT_SIZE 85000
 
 MethodTable g_FreeObjectEEType;
@@ -471,6 +477,29 @@ EXTERN_C int64_t QCALLTYPE RhGetTotalAllocatedBytesPrecise()
     return allocated;
 }
 
+void FireAllocationSampled(GC_ALLOC_FLAGS flags, size_t size, size_t samplingBudgetOffset, Object* orObject)
+{
+#ifdef FEATURE_EVENT_TRACE
+    void* typeId = GetLastAllocEEType();
+    // Note: Just as for AllocationTick, the type name cannot be retrieved
+    WCHAR* name = nullptr;
+
+    if (typeId != nullptr)
+    {
+        unsigned int allocKind =
+            (flags & GC_ALLOC_PINNED_OBJECT_HEAP) ? 2 :
+            (flags & GC_ALLOC_LARGE_OBJECT_HEAP) ? 1 :
+            0;  // SOH
+        unsigned int heapIndex = 0;
+#ifdef BACKGROUND_GC
+        gc_heap* hp = gc_heap::heap_of((BYTE*)orObject);
+        heapIndex = hp->heap_number;
+#endif
+        FireEtwAllocationSampled(allocKind, GetClrInstanceId(), typeId, name, heapIndex, (BYTE*)orObject, size, samplingBudgetOffset);
+    }
+#endif
+}
+
 static Object* GcAllocInternal(MethodTable* pEEType, uint32_t uFlags, uintptr_t numElements, Thread* pThread)
 {
     ASSERT(!pThread->IsDoNotTriggerGcSet());
@@ -539,8 +568,47 @@ static Object* GcAllocInternal(MethodTable* pEEType, uint32_t uFlags, uintptr_t
     // Save the MethodTable for instrumentation purposes.
     tls_pLastAllocationEEType = pEEType;
 
-    Object* pObject = GCHeapUtilities::GetGCHeap()->Alloc(pThread->GetAllocContext(), cbSize, uFlags);
-    pThread->GetEEAllocContext()->UpdateCombinedLimit();
+    // check for dynamic allocation sampling
+    ee_alloc_context* pEEAllocContext = pThread->GetEEAllocContext();
+    gc_alloc_context* pAllocContext = pEEAllocContext->GetGCAllocContext();
+    bool isSampled = false;
+    size_t availableSpace = 0;
+    size_t samplingBudget = 0;
+
+    bool isRandomizedSamplingEnabled = ee_alloc_context::IsRandomizedSamplingEnabled();
+    if (isRandomizedSamplingEnabled)
+    {
+        // The number bytes we can allocate before we need to emit a sampling event.
+        // This calculation is only valid if combined_limit < alloc_limit.
+        samplingBudget = (size_t)(pEEAllocContext->combined_limit - pAllocContext->alloc_ptr);
+
+        // The number of bytes available in the current allocation context
+        availableSpace = (size_t)(pAllocContext->alloc_limit - pAllocContext->alloc_ptr);
+
+        // Check to see if the allocated object overlaps a sampled byte
+        // in this AC. This happens when both:
+        // 1) The AC contains a sampled byte (combined_limit < alloc_limit)
+        // 2) The object is large enough to overlap it (samplingBudget < aligned_size)
+        //
+        // Note that the AC could have no remaining space for allocations (alloc_ptr =
+        // alloc_limit = combined_limit). When a thread hasn't done any SOH allocations
+        // yet it also starts in an empty state where alloc_ptr = alloc_limit =
+        // combined_limit = nullptr. The (1) check handles both of these situations
+        // properly as an empty AC can not have a sampled byte inside of it.
+        isSampled =
+            (pEEAllocContext->combined_limit < pAllocContext->alloc_limit) &&
+            (samplingBudget < cbSize);
+
+        // if the object overflows the AC, we need to sample the remaining bytes
+        // the sampling budget only included at most the bytes inside the AC
+        if (cbSize > availableSpace && !isSampled)
+        {
+            samplingBudget = ee_alloc_context::ComputeGeometricRandom() + availableSpace;
+            isSampled = (samplingBudget < cbSize);
+        }
+    }
+
+    Object* pObject = GCHeapUtilities::GetGCHeap()->Alloc(pAllocContext, cbSize, uFlags);
     if (pObject == NULL)
         return NULL;
 
@@ -551,6 +619,19 @@ static Object* GcAllocInternal(MethodTable* pEEType, uint32_t uFlags, uintptr_t
         ((Array*)pObject)->InitArrayLength((uint32_t)numElements);
     }
 
+    if (isSampled)
+    {
+        FireAllocationSampled((GC_ALLOC_FLAGS)uFlags, cbSize, samplingBudget, pObject);
+    }
+
+    // There are a variety of conditions that may have invalidated the previous combined_limit value
+    // such as not allocating the object in the AC memory region (UOH allocations), moving the AC, adding
+    // extra alignment padding, allocating a new AC, or allocating an object that consumed the sampling budget.
+    // Rather than test for all the different invalidation conditions individually we conservatively always
+    // recompute it. If sampling isn't enabled this inlined function is just trivially setting
+    // combined_limit=alloc_limit.
+    pEEAllocContext->UpdateCombinedLimit(isRandomizedSamplingEnabled);
+
     if (uFlags & GC_ALLOC_USER_OLD_HEAP)
         GCHeapUtilities::GetGCHeap()->PublishObject((uint8_t*)pObject);
 

src/coreclr/nativeaot/Runtime/GCHelpers.cpp

@@ -12,6 +12,11 @@
 
 void EventTracing_Initialize() { }
 
+bool IsRuntimeProviderEnabled(uint8_t level, uint64_t keyword)
+{
+    return false;
+}
+
 void ETW::GCLog::FireGcStart(ETW_GC_INFO * pGcInfo) { }
 
 #ifdef FEATURE_ETW

src/coreclr/nativeaot/Runtime/disabledeventtrace.cpp

@@ -1,5 +1,6 @@
 # Native runtime events supported by aot runtime.
 
+AllocationSampled
 BGC1stConEnd
 BGC1stNonConEnd
 BGC1stSweepEnd

src/coreclr/nativeaot/Runtime/eventpipe/gen-eventing-event-inc.lst

@@ -37,6 +37,11 @@ DOTNET_TRACE_CONTEXT MICROSOFT_WINDOWS_DOTNETRUNTIME_PRIVATE_PROVIDER_DOTNET_Con
     MICROSOFT_WINDOWS_DOTNETRUNTIME_PRIVATE_PROVIDER_EVENTPIPE_Context
 };
 
+bool IsRuntimeProviderEnabled(uint8_t level, uint64_t keyword)
+{
+    return RUNTIME_PROVIDER_CATEGORY_ENABLED(level, keyword);
+}
+
 volatile LONGLONG ETW::GCLog::s_l64LastClientSequenceNumber = 0;
 
 //---------------------------------------------------------------------------------------
@@ -245,4 +250,4 @@ void EventPipeEtwCallbackDotNETRuntimePrivate(
     _Inout_opt_ PVOID CallbackContext)
 {
     EtwCallbackCommon(DotNETRuntimePrivate, ControlCode, Level, MatchAnyKeyword, FilterData, true);
-}
+}

src/coreclr/nativeaot/Runtime/eventtrace.cpp

@@ -30,6 +30,8 @@ void InitializeEventTracing();
 
 #ifdef FEATURE_EVENT_TRACE
 
+bool IsRuntimeProviderEnabled(uint8_t level, uint64_t keyword);
+
 // !!!!!!! NOTE !!!!!!!!
 // The flags must match those in the ETW manifest exactly
 // !!!!!!! NOTE !!!!!!!!
@@ -102,6 +104,7 @@ struct ProfilingScanContext;
 #define CLR_GCHEAPSURVIVALANDMOVEMENT_KEYWORD 0x400000
 #define CLR_MANAGEDHEAPCOLLECT_KEYWORD 0x800000
 #define CLR_GCHEAPANDTYPENAMES_KEYWORD 0x1000000
+#define CLR_ALLOCATIONSAMPLING_KEYWORD 0x80000000000
 
 //
 // Using KEYWORDZERO means when checking the events category ignore the keyword

src/coreclr/nativeaot/Runtime/eventtracebase.h

@@ -4,6 +4,7 @@
 #include "common.h"
 #include "gctoclreventsink.h"
 #include "thread.h"
+#include "eventtracebase.h"
 
 GCToCLREventSink g_gcToClrEventSink;
 
@@ -174,6 +175,14 @@ void GCToCLREventSink::FireGCAllocationTick_V4(uint64_t allocationAmount,
 {
     LIMITED_METHOD_CONTRACT;
 
+#ifdef FEATURE_EVENT_TRACE
+    if (IsRuntimeProviderEnabled(TRACE_LEVEL_INFORMATION, CLR_ALLOCATIONSAMPLING_KEYWORD))
+    {
+        // skip AllocationTick if AllocationSampled is emitted
+        return;
+    }
+#endif // FEATURE_EVENT_TRACE
+
     void * typeId = GetLastAllocEEType();
     WCHAR * name = nullptr;
 

src/coreclr/nativeaot/Runtime/gctoclreventsink.cpp

@@ -35,6 +35,13 @@ static Thread* g_RuntimeInitializingThread;
 
 #endif //!DACCESS_COMPILE
 
+ee_alloc_context::PerThreadRandom::PerThreadRandom()
+{
+    minipal_xoshiro128pp_init(&random_state, (uint32_t)PalGetTickCount64());
+}
+
+thread_local ee_alloc_context::PerThreadRandom ee_alloc_context::t_random = PerThreadRandom();
+
 PInvokeTransitionFrame* Thread::GetTransitionFrame()
 {
     if (ThreadStore::GetSuspendingThread() == this)

src/coreclr/nativeaot/Runtime/thread.cpp

@@ -6,6 +6,7 @@
 
 #include "StackFrameIterator.h"
 #include "slist.h" // DefaultSListTraits
+#include <minipal/xoshiro128pp.h>
 
 struct gc_alloc_context;
 class RuntimeInstance;
@@ -113,7 +114,19 @@ struct ee_alloc_context
 
     gc_alloc_context* GetGCAllocContext();
     uint8_t* GetCombinedLimit();
-    void UpdateCombinedLimit();
+    void UpdateCombinedLimit(bool samplingEnabled);
+    static bool IsRandomizedSamplingEnabled();
+    static int ComputeGeometricRandom();
+
+    struct PerThreadRandom
+    {
+        minipal_xoshiro128pp random_state;
+
+        PerThreadRandom();
+        float NextFloat();
+    };
+
+    static thread_local PerThreadRandom t_random;
 };
 
 

src/coreclr/nativeaot/Runtime/thread.h

@@ -3,7 +3,9 @@
 
 #ifndef DACCESS_COMPILE
 
+#include "eventtracebase.h"
 
+const uint32_t SamplingDistributionMean = (100 * 1024);
 
 inline gc_alloc_context* ee_alloc_context::GetGCAllocContext()
 {
@@ -22,11 +24,43 @@ struct _thread_inl_gc_alloc_context
     uint8_t* alloc_limit;
 };
 
-inline void ee_alloc_context::UpdateCombinedLimit()
+
+inline bool ee_alloc_context::IsRandomizedSamplingEnabled()
+{
+#ifdef FEATURE_EVENT_TRACE
+    return IsRuntimeProviderEnabled(TRACE_LEVEL_INFORMATION, CLR_ALLOCATIONSAMPLING_KEYWORD);
+#else
+    return false;
+#endif // FEATURE_EVENT_TRACE
+}
+
+inline void ee_alloc_context::UpdateCombinedLimit(bool samplingEnabled)
+{
+    if (!samplingEnabled)
+    {
+        combined_limit = ((_thread_inl_gc_alloc_context*)GetGCAllocContext())->alloc_limit;
+    }
+    else
+    {
+        // compute the next sampling limit based on a geometric distribution
+        uint8_t* sampling_limit = ((_thread_inl_gc_alloc_context*)GetGCAllocContext())->alloc_ptr + ComputeGeometricRandom();
+
+        // if the sampling limit is larger than the allocation context, no sampling will occur in this AC
+        combined_limit = min(sampling_limit, ((_thread_inl_gc_alloc_context*)GetGCAllocContext())->alloc_limit);
+    }
+}
+
+inline int ee_alloc_context::ComputeGeometricRandom()
+{
+    // compute a random sample from the Geometric distribution
+    float probability = t_random.NextFloat();
+    int threshold = (int)(-log(1 - probability) * SamplingDistributionMean);
+    return threshold;
+}
+
+inline float ee_alloc_context::PerThreadRandom::NextFloat()
 {
-    // The randomized allocation sampling feature is being submitted in stages. For now sampling is never enabled so
-    // combined_limit is always the same as alloc_limit.
-    combined_limit = ((_thread_inl_gc_alloc_context*)GetGCAllocContext())->alloc_limit;
+    return (float)minipal_xoshiro128pp_next(&random_state) * (1.0F/UINT32_MAX);
 }
 
 // Set the m_pDeferredTransitionFrame field for GC allocation helpers that setup transition frame