Implementing DynamicIterator for TransformRepNoLock

Summary: What @haobo done with TransformRep, now in TransformRepNoLock. Similar implementation, except that I made DynamicIterator a subclass of Iterator which makes me have less iterator initializations.

Test Plan: ./prefix_test. Seeing huge savings vs. TransformRep again!

Reviewers: dhruba, haobo, sdong, kailiu

Reviewed By: haobo

CC: leveldb, haobo

Differential Revision: https://reviews.facebook.net/D13953
This commit is contained in:
Igor Canadi 2013-11-08 00:31:09 -08:00
parent fd075d6edd
commit 8b3379dc0a
4 changed files with 130 additions and 109 deletions

View File

@ -93,9 +93,8 @@ class PrefixTest {
auto prefix_extractor = NewFixedPrefixTransform(8); auto prefix_extractor = NewFixedPrefixTransform(8);
options.prefix_extractor = prefix_extractor; options.prefix_extractor = prefix_extractor;
if (FLAGS_use_nolock_version) { if (FLAGS_use_nolock_version) {
options.memtable_factory = options.memtable_factory.reset(NewHashSkipListRepFactory(
std::make_shared<rocksdb::PrefixHashRepNoLockFactory>( prefix_extractor, FLAGS_bucket_count));
prefix_extractor, FLAGS_bucket_count);
} else { } else {
options.memtable_factory = options.memtable_factory =
std::make_shared<rocksdb::PrefixHashRepFactory>( std::make_shared<rocksdb::PrefixHashRepFactory>(

View File

@ -63,6 +63,11 @@ class SkipList {
// The returned iterator is not valid. // The returned iterator is not valid.
explicit Iterator(const SkipList* list); explicit Iterator(const SkipList* list);
// Change the underlying skiplist used for this iterator
// This enables us not changing the iterator without deallocating
// an old one and then allocating a new one
void SetList(const SkipList* list);
// Returns true iff the iterator is positioned at a valid node. // Returns true iff the iterator is positioned at a valid node.
bool Valid() const; bool Valid() const;
@ -194,6 +199,11 @@ SkipList<Key,Comparator>::NewNode(const Key& key, int height) {
template<typename Key, class Comparator> template<typename Key, class Comparator>
inline SkipList<Key,Comparator>::Iterator::Iterator(const SkipList* list) { inline SkipList<Key,Comparator>::Iterator::Iterator(const SkipList* list) {
SetList(list);
}
template<typename Key, class Comparator>
inline void SkipList<Key,Comparator>::Iterator::SetList(const SkipList* list) {
list_ = list; list_ = list;
node_ = nullptr; node_ = nullptr;
} }

View File

@ -260,67 +260,11 @@ public:
} }
}; };
// NO LOCKS VERSION
// The same as TransformRepFactory except it doesn't use locks. // The same as TransformRepFactory except it doesn't use locks.
// Experimental, will replace TransformRepFactory once we are sure // Experimental, will replace TransformRepFactory once we are sure
// it performs better // it performs better
class TransformRepNoLockFactory : public MemTableRepFactory { extern MemTableRepFactory* NewHashSkipListRepFactory(
public: const SliceTransform* transform, size_t bucket_count = 1000000);
explicit TransformRepNoLockFactory(const SliceTransform* transform,
size_t bucket_count)
: transform_(transform),
bucket_count_(bucket_count) { }
virtual ~TransformRepNoLockFactory() { delete transform_; }
virtual std::shared_ptr<MemTableRep> CreateMemTableRep(
MemTableRep::KeyComparator&, Arena*) override;
virtual const char* Name() const override {
return "TransformRepNoLockFactory";
}
const SliceTransform* GetTransform() { return transform_; }
protected:
const SliceTransform* transform_;
const size_t bucket_count_;
};
// UnsortedReps bin user keys based on an identity function transform -- that
// is, transform(key) = key. This optimizes for point look-ups.
//
// Parameters: See TransformRepNoLockFactory.
class UnsortedRepNoLockFactory : public TransformRepNoLockFactory {
public:
explicit UnsortedRepNoLockFactory(size_t bucket_count = 1000000)
: TransformRepNoLockFactory(NewNoopTransform(),
bucket_count) { }
virtual const char* Name() const override {
return "UnsortedRepNoLockFactory";
}
};
// PrefixHashReps bin user keys based on a fixed-size prefix. This optimizes for
// short ranged scans over a given prefix.
//
// Parameters: See TransformRepNoLockFactory.
class PrefixHashRepNoLockFactory : public TransformRepNoLockFactory {
public:
explicit PrefixHashRepNoLockFactory(const SliceTransform* prefix_extractor,
size_t bucket_count = 1000000)
: TransformRepNoLockFactory(prefix_extractor, bucket_count)
{ }
virtual std::shared_ptr<MemTableRep> CreateMemTableRep(
MemTableRep::KeyComparator&, Arena*) override;
virtual const char* Name() const override {
return "PrefixHashRepNoLockFactory";
}
};
} }

View File

@ -16,9 +16,9 @@
namespace rocksdb { namespace rocksdb {
namespace { namespace {
class TransformRepNoLock : public MemTableRep { class HashSkipListRep : public MemTableRep {
public: public:
TransformRepNoLock(MemTableRep::KeyComparator& compare, Arena* arena, HashSkipListRep(MemTableRep::KeyComparator& compare, Arena* arena,
const SliceTransform* transform, size_t bucket_size); const SliceTransform* transform, size_t bucket_size);
virtual void Insert(const char* key) override; virtual void Insert(const char* key) override;
@ -27,17 +27,21 @@ class TransformRepNoLock : public MemTableRep {
virtual size_t ApproximateMemoryUsage() override; virtual size_t ApproximateMemoryUsage() override;
virtual ~TransformRepNoLock(); virtual ~HashSkipListRep();
virtual std::shared_ptr<MemTableRep::Iterator> GetIterator() override; virtual std::shared_ptr<MemTableRep::Iterator> GetIterator() override;
virtual std::shared_ptr<MemTableRep::Iterator> GetIterator( virtual std::shared_ptr<MemTableRep::Iterator> GetIterator(
const Slice& slice) override; const Slice& slice) override;
std::shared_ptr<MemTableRep::Iterator> GetTransformIterator( virtual std::shared_ptr<MemTableRep::Iterator> GetPrefixIterator(
const Slice& transformed); const Slice& prefix) override;
virtual std::shared_ptr<MemTableRep::Iterator> GetDynamicPrefixIterator()
override;
private: private:
friend class DynamicIterator;
typedef SkipList<const char*, MemTableRep::KeyComparator&> Bucket; typedef SkipList<const char*, MemTableRep::KeyComparator&> Bucket;
size_t bucket_size_; size_t bucket_size_;
@ -76,50 +80,72 @@ class TransformRepNoLock : public MemTableRep {
virtual ~Iterator() { virtual ~Iterator() {
// if we own the list, we should also delete it // if we own the list, we should also delete it
if (own_list_) { if (own_list_) {
assert(list_ != nullptr);
delete list_; delete list_;
} }
}; }
// Returns true iff the iterator is positioned at a valid node. // Returns true iff the iterator is positioned at a valid node.
virtual bool Valid() const { virtual bool Valid() const {
return iter_.Valid(); return list_ != nullptr && iter_.Valid();
} }
// Returns the key at the current position. // Returns the key at the current position.
// REQUIRES: Valid() // REQUIRES: Valid()
virtual const char* key() const { virtual const char* key() const {
assert(Valid());
return iter_.key(); return iter_.key();
} }
// Advances to the next position. // Advances to the next position.
// REQUIRES: Valid() // REQUIRES: Valid()
virtual void Next() { virtual void Next() {
assert(Valid());
iter_.Next(); iter_.Next();
} }
// Advances to the previous position. // Advances to the previous position.
// REQUIRES: Valid() // REQUIRES: Valid()
virtual void Prev() { virtual void Prev() {
assert(Valid());
iter_.Prev(); iter_.Prev();
} }
// Advance to the first entry with a key >= target // Advance to the first entry with a key >= target
virtual void Seek(const char* target) { virtual void Seek(const char* target) {
iter_.Seek(target); if (list_ != nullptr) {
iter_.Seek(target);
}
} }
// Position at the first entry in collection. // Position at the first entry in collection.
// Final state of iterator is Valid() iff collection is not empty. // Final state of iterator is Valid() iff collection is not empty.
virtual void SeekToFirst() { virtual void SeekToFirst() {
iter_.SeekToFirst(); if (list_ != nullptr) {
iter_.SeekToFirst();
}
} }
// Position at the last entry in collection. // Position at the last entry in collection.
// Final state of iterator is Valid() iff collection is not empty. // Final state of iterator is Valid() iff collection is not empty.
virtual void SeekToLast() { virtual void SeekToLast() {
iter_.SeekToLast(); if (list_ != nullptr) {
iter_.SeekToLast();
}
}
protected:
void Reset(Bucket* list) {
if (own_list_) {
assert(list_ != nullptr);
delete list_;
}
list_ = list;
iter_.SetList(list);
own_list_ = false;
} }
private: private:
// if list_ is nullptr, we should NEVER call any methods on iter_
// if list_ is nullptr, this Iterator is not Valid()
Bucket* list_; Bucket* list_;
Bucket::Iterator iter_; Bucket::Iterator iter_;
// here we track if we own list_. If we own it, we are also // here we track if we own list_. If we own it, we are also
@ -127,6 +153,40 @@ class TransformRepNoLock : public MemTableRep {
bool own_list_; bool own_list_;
}; };
class DynamicIterator : public HashSkipListRep::Iterator {
public:
explicit DynamicIterator(const HashSkipListRep& memtable_rep)
: HashSkipListRep::Iterator(nullptr, false),
memtable_rep_(memtable_rep) {}
// Advance to the first entry with a key >= target
virtual void Seek(const char* target) {
auto transformed = memtable_rep_.transform_->Transform(
memtable_rep_.UserKey(target));
Reset(memtable_rep_.GetBucket(transformed));
HashSkipListRep::Iterator::Seek(target);
}
// Position at the first entry in collection.
// Final state of iterator is Valid() iff collection is not empty.
virtual void SeekToFirst() {
// Prefix iterator does not support total order.
// We simply set the iterator to invalid state
Reset(nullptr);
}
// Position at the last entry in collection.
// Final state of iterator is Valid() iff collection is not empty.
virtual void SeekToLast() {
// Prefix iterator does not support total order.
// We simply set the iterator to invalid state
Reset(nullptr);
}
private:
// the underlying memtable
const HashSkipListRep& memtable_rep_;
};
class EmptyIterator : public MemTableRep::Iterator { class EmptyIterator : public MemTableRep::Iterator {
// This is used when there wasn't a bucket. It is cheaper than // This is used when there wasn't a bucket. It is cheaper than
// instantiating an empty bucket over which to iterate. // instantiating an empty bucket over which to iterate.
@ -150,17 +210,7 @@ class TransformRepNoLock : public MemTableRep {
std::shared_ptr<EmptyIterator> empty_iterator_; std::shared_ptr<EmptyIterator> empty_iterator_;
}; };
class PrefixHashRepNoLock : public TransformRepNoLock { HashSkipListRep::HashSkipListRep(MemTableRep::KeyComparator& compare,
public:
PrefixHashRepNoLock(MemTableRep::KeyComparator& compare, Arena* arena,
const SliceTransform* transform, size_t bucket_size)
: TransformRepNoLock(compare, arena, transform, bucket_size) { }
virtual std::shared_ptr<MemTableRep::Iterator> GetPrefixIterator(
const Slice& prefix) override;
};
TransformRepNoLock::TransformRepNoLock(MemTableRep::KeyComparator& compare,
Arena* arena, const SliceTransform* transform, size_t bucket_size) Arena* arena, const SliceTransform* transform, size_t bucket_size)
: bucket_size_(bucket_size), : bucket_size_(bucket_size),
transform_(transform), transform_(transform),
@ -175,11 +225,11 @@ TransformRepNoLock::TransformRepNoLock(MemTableRep::KeyComparator& compare,
} }
} }
TransformRepNoLock::~TransformRepNoLock() { HashSkipListRep::~HashSkipListRep() {
delete[] buckets_; delete[] buckets_;
} }
TransformRepNoLock::Bucket* TransformRepNoLock::GetInitializedBucket( HashSkipListRep::Bucket* HashSkipListRep::GetInitializedBucket(
const Slice& transformed) { const Slice& transformed) {
size_t hash = GetHash(transformed); size_t hash = GetHash(transformed);
auto bucket = GetBucket(hash); auto bucket = GetBucket(hash);
@ -191,14 +241,14 @@ TransformRepNoLock::Bucket* TransformRepNoLock::GetInitializedBucket(
return bucket; return bucket;
} }
void TransformRepNoLock::Insert(const char* key) { void HashSkipListRep::Insert(const char* key) {
assert(!Contains(key)); assert(!Contains(key));
auto transformed = transform_->Transform(UserKey(key)); auto transformed = transform_->Transform(UserKey(key));
auto bucket = GetInitializedBucket(transformed); auto bucket = GetInitializedBucket(transformed);
bucket->Insert(key); bucket->Insert(key);
} }
bool TransformRepNoLock::Contains(const char* key) const { bool HashSkipListRep::Contains(const char* key) const {
auto transformed = transform_->Transform(UserKey(key)); auto transformed = transform_->Transform(UserKey(key));
auto bucket = GetBucket(transformed); auto bucket = GetBucket(transformed);
if (bucket == nullptr) { if (bucket == nullptr) {
@ -207,11 +257,11 @@ bool TransformRepNoLock::Contains(const char* key) const {
return bucket->Contains(key); return bucket->Contains(key);
} }
size_t TransformRepNoLock::ApproximateMemoryUsage() { size_t HashSkipListRep::ApproximateMemoryUsage() {
return sizeof(buckets_); return sizeof(buckets_);
} }
std::shared_ptr<MemTableRep::Iterator> TransformRepNoLock::GetIterator() { std::shared_ptr<MemTableRep::Iterator> HashSkipListRep::GetIterator() {
auto list = new Bucket(compare_, arena_); auto list = new Bucket(compare_, arena_);
for (size_t i = 0; i < bucket_size_; ++i) { for (size_t i = 0; i < bucket_size_; ++i) {
auto bucket = GetBucket(i); auto bucket = GetBucket(i);
@ -225,38 +275,56 @@ std::shared_ptr<MemTableRep::Iterator> TransformRepNoLock::GetIterator() {
return std::make_shared<Iterator>(list); return std::make_shared<Iterator>(list);
} }
std::shared_ptr<MemTableRep::Iterator> TransformRepNoLock::GetTransformIterator( std::shared_ptr<MemTableRep::Iterator> HashSkipListRep::GetPrefixIterator(
const Slice& transformed) { const Slice& prefix) {
auto bucket = GetBucket(transformed); auto bucket = GetBucket(prefix);
if (bucket == nullptr) { if (bucket == nullptr) {
return empty_iterator_; return empty_iterator_;
} }
return std::make_shared<Iterator>(bucket, false); return std::make_shared<Iterator>(bucket, false);
} }
std::shared_ptr<MemTableRep::Iterator> TransformRepNoLock::GetIterator( std::shared_ptr<MemTableRep::Iterator> HashSkipListRep::GetIterator(
const Slice& slice) { const Slice& slice) {
auto transformed = transform_->Transform(slice); return GetPrefixIterator(transform_->Transform(slice));
return GetTransformIterator(transformed); }
std::shared_ptr<MemTableRep::Iterator>
HashSkipListRep::GetDynamicPrefixIterator() {
return std::make_shared<DynamicIterator>(*this);
} }
} // anon namespace } // anon namespace
std::shared_ptr<MemTableRep> TransformRepNoLockFactory::CreateMemTableRep( class HashSkipListRepFactory : public MemTableRepFactory {
MemTableRep::KeyComparator& compare, Arena* arena) { public:
return std::make_shared<TransformRepNoLock>(compare, arena, transform_, explicit HashSkipListRepFactory(const SliceTransform* transform,
bucket_count_); size_t bucket_count = 1000000)
} : transform_(transform),
bucket_count_(bucket_count) { }
std::shared_ptr<MemTableRep> PrefixHashRepNoLockFactory::CreateMemTableRep( virtual ~HashSkipListRepFactory() { delete transform_; }
MemTableRep::KeyComparator& compare, Arena* arena) {
return std::make_shared<PrefixHashRepNoLock>(compare, arena, transform_,
bucket_count_);
}
std::shared_ptr<MemTableRep::Iterator> PrefixHashRepNoLock::GetPrefixIterator( virtual std::shared_ptr<MemTableRep> CreateMemTableRep(
const Slice& prefix) { MemTableRep::KeyComparator& compare, Arena* arena) override {
return TransformRepNoLock::GetTransformIterator(prefix); return std::make_shared<HashSkipListRep>(compare, arena, transform_,
bucket_count_);
}
virtual const char* Name() const override {
return "HashSkipListRepFactory";
}
const SliceTransform* GetTransform() { return transform_; }
private:
const SliceTransform* transform_;
const size_t bucket_count_;
};
MemTableRepFactory* NewHashSkipListRepFactory(
const SliceTransform* transform, size_t bucket_count) {
return new HashSkipListRepFactory(transform, bucket_count);
} }
} // namespace rocksdb } // namespace rocksdb