Bckground eviction for multi-tier

byrnedj · byrnedj · commit bb35a05190b0 · 2024-06-25T08:21:38.000-07:00
Part 4.
-------------------------------
batch eviction / promotion
 - these changes are pretty significant so we would
   avoid squashing this commit in any prior
   background evictor patch
diff --git a/cachelib/allocator/CacheAllocator.h b/cachelib/allocator/CacheAllocator.h
diff --git a/cachelib/allocator/MM2Q.h b/cachelib/allocator/MM2Q.h
@@ -461,6 +461,18 @@ class MM2Q {
     //          is unchanged.
     bool add(T& node) noexcept;
 
+    // helper function to add the node under the container lock
+    void addNodeLocked(T& node, const Time& currTime);
+    
+    // adds the given nodes into the container and marks each as being present in
+    // the container. The nodes are added to the head of the lru.
+    //
+    // @param vector of nodes  The nodes to be added to the container.
+    // @return  number of nodes added - it is up to user to verify all
+    //          expected nodes have been added.
+    template <typename It>
+    uint32_t addBatch(It begin, It end) noexcept;
+
     // removes the node from the lru and sets it previous and next to nullptr.
     //
     // @param node  The node to be removed from the container.
@@ -895,16 +907,41 @@ bool MM2Q::Container<T, HookPtr>::add(T& node) noexcept {
     if (node.isInMMContainer()) {
       return false;
     }
+    addNodeLocked(node, currTime);
+    return true;
+  });
+}
 
-    markHot(node);
-    unmarkCold(node);
-    unmarkTail(node);
-    lru_.getList(LruType::Hot).linkAtHead(node);
-    rebalance();
+// adds the node to the list assuming not in 
+// container and holding container lock
+template <typename T, MM2Q::Hook<T> T::*HookPtr>
+void MM2Q::Container<T, HookPtr>::addNodeLocked(T& node, const Time& currTime) {
+  XDCHECK(!node.isInMMContainer());
+  markHot(node);
+  unmarkCold(node);
+  unmarkTail(node);
+  lru_.getList(LruType::Hot).linkAtHead(node);
+  rebalance();
+
+  node.markInMMContainer();
+  setUpdateTime(node, currTime);
+}
 
-    node.markInMMContainer();
-    setUpdateTime(node, currTime);
-    return true;
+template <typename T, MM2Q::Hook<T> T::*HookPtr>
+template <typename It>
+uint32_t MM2Q::Container<T, HookPtr>::addBatch(It begin, It end) noexcept {
+  const auto currTime = static_cast<Time>(util::getCurrentTimeSec());
+  return lruMutex_->lock_combine([this, begin, end, currTime]() {
+    uint32_t i = 0;
+    for (auto itr = begin; itr != end; itr++) {
+      T* node = *itr;
+      if (node->isInMMContainer()) {
+        return i;
+      }
+      addNodeLocked(*node,currTime);
+      i++;
+    }
+    return i;
   });
 }
 
@@ -935,6 +972,7 @@ MM2Q::Container<T, HookPtr>::withPromotionIterator(F&& fun) {
     fun(LockedIterator{LockHolder{}, lru_.begin(LruType::Hot)});
   });
 }
+
 template <typename T, MM2Q::Hook<T> T::*HookPtr>
 template <typename F>
 void MM2Q::Container<T, HookPtr>::withContainerLock(F&& fun) {
diff --git a/cachelib/allocator/MMLru.h b/cachelib/allocator/MMLru.h
@@ -337,6 +337,18 @@ class MMLru {
     //          is unchanged.
     bool add(T& node) noexcept;
 
+    // helper function to add the node under the container lock
+    void addNodeLocked(T& node, const Time& currTime);
+    
+    // adds the given nodes into the container and marks each as being present in
+    // the container. The nodes are added to the head of the lru.
+    //
+    // @param vector of nodes  The nodes to be added to the container.
+    // @return  number of nodes added - it is up to user to verify all
+    //          expected nodes have been added.
+    template <typename It>
+    uint32_t addBatch(It begin, It end) noexcept;
+
     // removes the node from the lru and sets it previous and next to nullptr.
     //
     // @param node  The node to be removed from the container.
@@ -690,19 +702,46 @@ bool MMLru::Container<T, HookPtr>::add(T& node) noexcept {
     if (node.isInMMContainer()) {
       return false;
     }
-    if (config_.lruInsertionPointSpec == 0 || insertionPoint_ == nullptr) {
-      lru_.linkAtHead(node);
-    } else {
-      lru_.insertBefore(*insertionPoint_, node);
-    }
-    node.markInMMContainer();
-    setUpdateTime(node, currTime);
-    unmarkAccessed(node);
-    updateLruInsertionPoint();
+    addNodeLocked(node,currTime);
     return true;
   });
 }
 
+template <typename T, MMLru::Hook<T> T::*HookPtr>
+void MMLru::Container<T, HookPtr>::addNodeLocked(T& node, const Time& currTime) {
+  XDCHECK(!node.isInMMContainer());
+  if (config_.lruInsertionPointSpec == 0 || insertionPoint_ == nullptr) {
+    lru_.linkAtHead(node);
+  } else {
+    lru_.insertBefore(*insertionPoint_, node);
+  }
+  node.markInMMContainer();
+  setUpdateTime(node, currTime);
+  unmarkAccessed(node);
+  updateLruInsertionPoint();
+}
+
+template <typename T, MMLru::Hook<T> T::*HookPtr>
+template <typename It>
+uint32_t MMLru::Container<T, HookPtr>::addBatch(It begin, It end) noexcept {
+  const auto currTime = static_cast<Time>(util::getCurrentTimeSec());
+  return lruMutex_->lock_combine([this, begin, end, currTime]() {
+    uint32_t i = 0;
+    for (auto itr = begin; itr != end; ++itr) {
+      T* node = *itr;
+      XDCHECK(!node->isInMMContainer());
+      if (node->isInMMContainer()) {
+        throw std::runtime_error(
+          folly::sformat("Was not able to add all new items, failed item {}",
+                          node->toString()));
+      }
+      addNodeLocked(*node,currTime);
+      i++;
+    }
+    return i;
+  });
+}
+
 template <typename T, MMLru::Hook<T> T::*HookPtr>
 typename MMLru::Container<T, HookPtr>::LockedIterator
 MMLru::Container<T, HookPtr>::getEvictionIterator() const noexcept {
diff --git a/cachelib/allocator/MMTinyLFU.h b/cachelib/allocator/MMTinyLFU.h
@@ -376,6 +376,18 @@ class MMTinyLFU {
     //          if the node was already in the contianer. On error state of node
     //          is unchanged.
     bool add(T& node) noexcept;
+    
+    // helper function to add the node under the container lock
+    void addNodeLocked(T& node, const Time& currTime);
+    
+    // adds the given nodes into the container and marks each as being present in
+    // the container. The nodes are added to the head of the lru.
+    //
+    // @param vector of nodes  The nodes to be added to the container.
+    // @return  number of nodes added - it is up to user to verify all
+    //          expected nodes have been added.
+    template <typename It>
+    uint32_t addBatch(It begin, It end) noexcept;
 
     // removes the node from the lru and sets it previous and next to nullptr.
     //
@@ -861,7 +873,15 @@ bool MMTinyLFU::Container<T, HookPtr>::add(T& node) noexcept {
   if (node.isInMMContainer()) {
     return false;
   }
+  addNodeLocked(node, currTime);
+  return true;
+}
 
+// adds the node to the list assuming not in 
+// container and holding container lock
+template <typename T, MMTinyLFU::Hook<T> T::*HookPtr>
+void MMTinyLFU::Container<T, HookPtr>::addNodeLocked(T& node, const Time &currTime) {
+  XDCHECK(!node.isInMMContainer());
   auto& tinyLru = lru_.getList(LruType::Tiny);
   tinyLru.linkAtHead(node);
   markTiny(node);
@@ -889,7 +909,23 @@ bool MMTinyLFU::Container<T, HookPtr>::add(T& node) noexcept {
   node.markInMMContainer();
   setUpdateTime(node, currTime);
   unmarkAccessed(node);
-  return true;
+}
+
+template <typename T, MMTinyLFU::Hook<T> T::*HookPtr>
+template <typename It>
+uint32_t MMTinyLFU::Container<T, HookPtr>::addBatch(It begin, It end) noexcept {
+  const auto currTime = static_cast<Time>(util::getCurrentTimeSec());
+  LockHolder l(lruMutex_);
+  uint32_t i = 0;
+  for (auto itr = begin; itr != end; itr++) {
+    T* node = *itr;
+    if (node->isInMMContainer()) {
+      return i;
+    }
+    addNodeLocked(*node, currTime);
+    i++;
+  }
+  return i;
 }
 
 template <typename T, MMTinyLFU::Hook<T> T::*HookPtr>
diff --git a/cachelib/allocator/memory/AllocationClass.cpp b/cachelib/allocator/memory/AllocationClass.cpp
@@ -140,6 +140,26 @@ void* AllocationClass::addSlabAndAllocate(Slab* slab) {
   });
 }
 
+std::vector<void*> AllocationClass::addSlabAndAllocateBatch(Slab* slab, uint64_t batch) {
+  XDCHECK_NE(nullptr, slab);
+  std::vector<void*> allocs;
+  allocs.reserve(batch);
+  lock_->lock_combine([this, slab, batch, &allocs]() {
+    addSlabLocked(slab);
+    uint64_t total = 0;
+    while (total < batch) {
+     void *alloc = allocateLocked();
+     if (alloc != nullptr) {
+       allocs.push_back(alloc);
+       total++;
+     } else {
+       break;
+     }
+    }
+  });
+  return allocs;
+}
+
 void* AllocationClass::allocateFromCurrentSlabLocked() noexcept {
   XDCHECK(canAllocateFromCurrentSlabLocked());
   void* ret = currSlab_->memoryAtOffset(currOffset_);
@@ -159,6 +179,26 @@ void* AllocationClass::allocate() {
   return lock_->lock_combine([this]() -> void* { return allocateLocked(); });
 }
 
+std::vector<void*> AllocationClass::allocateBatch(uint64_t batch) {
+  std::vector<void*> allocs;
+  if (!canAllocate_) {
+    return allocs;
+  }
+  lock_->lock_combine([this, &allocs, batch]() { 
+    uint64_t total = 0;
+    while (total < batch) {
+      void *alloc = allocateLocked();
+      if (alloc != nullptr) {
+        allocs.push_back(alloc);
+        total++;
+      } else {
+        break;
+      }
+    }
+  });
+  return allocs;
+}
+
 void* AllocationClass::allocateLocked() {
   // fast path for case when the cache is mostly full.
   if (freedAllocations_.empty() && freeSlabs_.empty() &&
diff --git a/cachelib/allocator/memory/AllocationClass.h b/cachelib/allocator/memory/AllocationClass.h
@@ -116,6 +116,7 @@ class AllocationClass {
   //          don't have any free memory. The caller will have to add a slab
   //          to this slab class to make further allocations out of it.
   void* allocate();
+  std::vector<void*> allocateBatch(uint64_t batch);
 
   // @param ctx     release context for the slab owning this alloc
   // @param memory  memory to check
@@ -227,6 +228,7 @@ class AllocationClass {
   // @param slab    a new slab to be added. This can NOT be nullptr.
   // @return  new allocation. This cannot fail.
   void* addSlabAndAllocate(Slab* slab);
+  std::vector<void*> addSlabAndAllocateBatch(Slab* slab, uint64_t batch);
 
   // Releasing a slab is a two step process.
   // 1. Mark a slab for release, by calling `startSlabRelease`.
diff --git a/cachelib/allocator/memory/MemoryAllocator.cpp b/cachelib/allocator/memory/MemoryAllocator.cpp
@@ -71,6 +71,16 @@ void* MemoryAllocator::allocate(PoolId id, uint32_t size) {
   return mp.allocate(size);
 }
 
+void* MemoryAllocator::allocateByCid(PoolId id, ClassId cid) {
+  auto& mp = memoryPoolManager_.getPoolById(id);
+  return mp.allocateByCid(cid);
+}
+
+std::vector<void*> MemoryAllocator::allocateByCidBatch(PoolId id, ClassId cid, uint64_t batch) {
+  auto& mp = memoryPoolManager_.getPoolById(id);
+  return mp.allocateByCidBatch(cid, batch);
+}
+
 void* MemoryAllocator::allocateZeroedSlab(PoolId id) {
   if (!config_.enableZeroedSlabAllocs) {
     throw std::logic_error("Zeroed Slab allcoation is not enabled");
diff --git a/cachelib/allocator/memory/MemoryAllocator.h b/cachelib/allocator/memory/MemoryAllocator.h
@@ -167,6 +167,8 @@ class MemoryAllocator {
   // @throw std::invalid_argument if the poolId is invalid or the size is
   //        invalid.
   void* allocate(PoolId id, uint32_t size);
+  void* allocateByCid(PoolId id, ClassId cid);
+  std::vector<void*> allocateByCidBatch(PoolId id, ClassId cid, uint64_t batch);
 
   // Allocate a zeroed Slab
   //
diff --git a/cachelib/allocator/memory/MemoryPool.cpp b/cachelib/allocator/memory/MemoryPool.cpp
diff --git a/cachelib/allocator/memory/MemoryPool.h b/cachelib/allocator/memory/MemoryPool.h
