hints for narrowing down FindFile range and avoiding checking unrelevant L0 files

Summary:
The file tree structure in Version is prebuilt and the range of each file is known.
On the Get() code path, we do binary search in FindFile() by comparing
target key with each file's largest key and also check the range for each L0 file.
With some pre-calculated knowledge, each key comparision that has been done can serve
as a hint to narrow down further searches:
(1) If a key falls within a L0 file's range, we can safely skip the next
file if its range does not overlap with the current one.
(2) If a key falls within a file's range in level L0 - Ln-1, we should only
need to binary search in the next level for files that overlap with the current one.

(1) will be able to skip some files depending one the key distribution.
(2) can greatly reduce the range of binary search, especially for bottom
levels, given that one file most likely only overlaps with N files from
the level below (where N is max_bytes_for_level_multiplier). So on level
L, we will only look at ~N files instead of N^L files.

Some inital results: measured with 500M key DB, when write is light (10k/s = 1.2M/s), this
improves QPS ~7% on top of blocked bloom. When write is heavier (80k/s =
9.6M/s), it gives us ~13% improvement.

Test Plan: make all check

Reviewers: haobo, igor, dhruba, sdong, yhchiang

Reviewed By: haobo

CC: leveldb

Differential Revision: https://reviews.facebook.net/D17205

Makefile

@@ -106,6 +106,7 @@ TESTS = \
 	backupable_db_test \
 	version_edit_test \
 	version_set_test \
+	file_indexer_test \
 	write_batch_test\
 	deletefile_test \
 	table_test \
@@ -376,6 +377,9 @@ version_edit_test: db/version_edit_test.o $(LIBOBJECTS) $(TESTHARNESS)
 version_set_test: db/version_set_test.o $(LIBOBJECTS) $(TESTHARNESS)
 	$(CXX) db/version_set_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)
 
+file_indexer_test : db/file_indexer_test.o $(LIBOBJECTS) $(TESTHARNESS)
+	$(CXX) db/file_indexer_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)
+
 reduce_levels_test: tools/reduce_levels_test.o $(LIBOBJECTS) $(TESTHARNESS)
 	$(CXX) tools/reduce_levels_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)
 

db/file_indexer.cc

@@ -0,0 +1,202 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include "db/file_indexer.h"
+#include <algorithm>
+#include "rocksdb/comparator.h"
+#include "db/version_edit.h"
+
+namespace rocksdb {
+
+FileIndexer::FileIndexer(const uint32_t num_levels,
+                         const Comparator* ucmp)
+  : num_levels_(num_levels),
+    ucmp_(ucmp),
+    next_level_index_(num_levels),
+    level_rb_(num_levels, -1) {
+}
+
+
+uint32_t FileIndexer::NumLevelIndex() {
+  return next_level_index_.size();
+}
+
+uint32_t FileIndexer::LevelIndexSize(uint32_t level) {
+  return next_level_index_[level].size();
+}
+
+void FileIndexer::GetNextLevelIndex(
+    const uint32_t level, const uint32_t file_index, const int cmp_smallest,
+    const int cmp_largest, int32_t* left_bound, int32_t* right_bound) {
+  assert(level > 0);
+
+  // Last level, no hint
+  if (level == num_levels_ - 1) {
+    *left_bound = 0;
+    *right_bound = -1;
+    return;
+  }
+
+  assert(level < num_levels_ - 1);
+  assert(static_cast<int32_t>(file_index) <= level_rb_[level]);
+
+  const auto& index = next_level_index_[level][file_index];
+
+  if (cmp_smallest < 0) {
+    *left_bound = (level > 0 && file_index > 0) ?
+      next_level_index_[level][file_index - 1].largest_lb : 0;
+    *right_bound = index.smallest_rb;
+  } else if (cmp_smallest == 0) {
+    *left_bound = index.smallest_lb;
+    *right_bound = index.smallest_rb;
+  } else if (cmp_smallest > 0 && cmp_largest < 0) {
+    *left_bound = index.smallest_lb;
+    *right_bound = index.largest_rb;
+  } else if (cmp_largest == 0) {
+    *left_bound = index.largest_lb;
+    *right_bound = index.largest_rb;
+  } else if (cmp_largest > 0) {
+    *left_bound = index.largest_lb;
+    *right_bound = level_rb_[level + 1];
+  } else {
+    assert(false);
+  }
+
+  assert(*left_bound >= 0);
+  assert(*left_bound <= *right_bound + 1);
+  assert(*right_bound <= level_rb_[level + 1]);
+}
+
+void FileIndexer::ClearIndex() {
+  for (uint32_t level = 1; level < num_levels_; ++level) {
+    next_level_index_[level].clear();
+  }
+}
+
+void FileIndexer::UpdateIndex(std::vector<FileMetaData*>* const files) {
+  if (files == nullptr) {
+    return;
+  }
+
+  // L1 - Ln-1
+  for (uint32_t level = 1; level < num_levels_ - 1; ++level) {
+    const auto& upper_files = files[level];
+    const int32_t upper_size = upper_files.size();
+    const auto& lower_files = files[level + 1];
+    level_rb_[level] = upper_files.size() - 1;
+    if (upper_size == 0) {
+      continue;
+    }
+    auto& index = next_level_index_[level];
+    index.resize(upper_size);
+
+    CalculateLB(upper_files, lower_files, &index,
+        [this](const FileMetaData* a, const FileMetaData* b) -> int {
+          return ucmp_->Compare(a->smallest.user_key(), b->largest.user_key());
+        },
+        [](IndexUnit* index, int32_t f_idx) {
+          index->smallest_lb = f_idx;
+        });
+    CalculateLB(upper_files, lower_files, &index,
+        [this](const FileMetaData* a, const FileMetaData* b) -> int {
+          return ucmp_->Compare(a->largest.user_key(), b->largest.user_key());
+        },
+        [](IndexUnit* index, int32_t f_idx) {
+          index->largest_lb = f_idx;
+        });
+    CalculateRB(upper_files, lower_files, &index,
+        [this](const FileMetaData* a, const FileMetaData* b) -> int {
+          return ucmp_->Compare(a->smallest.user_key(), b->smallest.user_key());
+        },
+        [](IndexUnit* index, int32_t f_idx) {
+          index->smallest_rb = f_idx;
+        });
+    CalculateRB(upper_files, lower_files, &index,
+        [this](const FileMetaData* a, const FileMetaData* b) -> int {
+          return ucmp_->Compare(a->largest.user_key(), b->smallest.user_key());
+        },
+        [](IndexUnit* index, int32_t f_idx) {
+          index->largest_rb = f_idx;
+        });
+  }
+  level_rb_[num_levels_ - 1] = files[num_levels_ - 1].size() - 1;
+}
+
+void FileIndexer::CalculateLB(const std::vector<FileMetaData*>& upper_files,
+    const std::vector<FileMetaData*>& lower_files,
+    std::vector<IndexUnit>* index,
+    std::function<int(const FileMetaData*, const FileMetaData*)> cmp_op,
+    std::function<void(IndexUnit*, int32_t)> set_index) {
+  const int32_t upper_size = upper_files.size();
+  const int32_t lower_size = lower_files.size();
+  int32_t upper_idx = 0;
+  int32_t lower_idx = 0;
+  while (upper_idx < upper_size && lower_idx < lower_size) {
+    int cmp = cmp_op(upper_files[upper_idx], lower_files[lower_idx]);
+
+    if (cmp == 0) {
+      set_index(&(*index)[upper_idx], lower_idx);
+      ++upper_idx;
+      ++lower_idx;
+    } else if (cmp > 0) {
+      // Lower level's file (largest) is smaller, a key won't hit in that
+      // file. Move to next lower file
+      ++lower_idx;
+    } else {
+      // Lower level's file becomes larger, update the index, and
+      // move to the next upper file
+      set_index(&(*index)[upper_idx], lower_idx);
+      ++upper_idx;
+    }
+  }
+
+  while (upper_idx < upper_size) {
+    // Lower files are exhausted, that means the remaining upper files are
+    // greater than any lower files. Set the index to be the lower level size.
+    set_index(&(*index)[upper_idx], lower_size);
+    ++upper_idx;
+  }
+}
+
+void FileIndexer::CalculateRB(const std::vector<FileMetaData*>& upper_files,
+    const std::vector<FileMetaData*>& lower_files,
+    std::vector<IndexUnit>* index,
+    std::function<int(const FileMetaData*, const FileMetaData*)> cmp_op,
+    std::function<void(IndexUnit*, int32_t)> set_index) {
+  const int32_t upper_size = upper_files.size();
+  const int32_t lower_size = lower_files.size();
+  int32_t upper_idx = upper_size - 1;
+  int32_t lower_idx = lower_size - 1;
+  while (upper_idx >= 0 && lower_idx >= 0) {
+    int cmp = cmp_op(upper_files[upper_idx], lower_files[lower_idx]);
+
+    if (cmp == 0) {
+      set_index(&(*index)[upper_idx], lower_idx);
+      --upper_idx;
+      --lower_idx;
+    } else if (cmp < 0) {
+      // Lower level's file (smallest) is larger, a key won't hit in that
+      // file. Move to next lower file.
+      --lower_idx;
+    } else {
+      // Lower level's file becomes smaller, update the index, and move to
+      // the next the upper file
+      set_index(&(*index)[upper_idx], lower_idx);
+      --upper_idx;
+    }
+  }
+  while (upper_idx >= 0) {
+    // Lower files are exhausted, that means the remaining upper files are
+    // smaller than any lower files. Set it to -1.
+    set_index(&(*index)[upper_idx], -1);
+    --upper_idx;
+  }
+}
+
+}  // namespace rocksdb

db/file_indexer.h

@@ -0,0 +1,129 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#pragma once
+#include <cstdint>
+#include <functional>
+#include <limits>
+#include <vector>
+
+namespace rocksdb {
+
+class Comparator;
+class FileMetaData;
+
+// The file tree structure in Version is prebuilt and the range of each file
+// is known. On Version::Get(), it uses binary search to find a potential file
+// and then check if a target key can be found in the file by comparing the key
+// to each file's smallest and largest key. The results of these comparisions
+// can be reused beyond checking if a key falls into a file's range.
+// With some pre-calculated knowledge, each key comparision that has been done
+// can serve as a hint to narrow down further searches: if a key compared to
+// be smaller than a file's smallest or largest, that comparison can be used
+// to find out the right bound of next binary search. Similarly, if a key
+// compared to be larger than a file's smallest or largest, it can be utilized
+// to find out the left bound of next binary search.
+// With these hints: it can greatly reduce the range of binary search,
+// especially for bottom levels, given that one file most likely overlaps with
+// only N files from level below (where N is max_bytes_for_level_multiplier).
+// So on level L, we will only look at ~N files instead of N^L files on the
+// naive approach.
+class FileIndexer {
+ public:
+  FileIndexer(const uint32_t num_levels, const Comparator* ucmp);
+
+  uint32_t NumLevelIndex();
+
+  uint32_t LevelIndexSize(uint32_t level);
+
+  // Return a file index range in the next level to search for a key based on
+  // smallest and largest key comparision for the current file specified by
+  // level and file_index. When *left_index < *right_index, both index should
+  // be valid and fit in the vector size.
+  void GetNextLevelIndex(
+    const uint32_t level, const uint32_t file_index, const int cmp_smallest,
+    const int cmp_largest, int32_t* left_bound, int32_t* right_bound);
+
+  void ClearIndex();
+
+  void UpdateIndex(std::vector<FileMetaData*>* const files);
+
+  enum {
+    kLevelMaxIndex = std::numeric_limits<int32_t>::max()
+  };
+
+ private:
+  const uint32_t num_levels_;
+  const Comparator* ucmp_;
+
+  struct IndexUnit {
+    IndexUnit()
+      : smallest_lb(0), largest_lb(0), smallest_rb(-1), largest_rb(-1) {}
+    // During file search, a key is compared against smallest and largest
+    // from a FileMetaData. It can have 3 possible outcomes:
+    // (1) key is smaller than smallest, implying it is also smaller than
+    //     larger. Precalculated index based on "smallest < smallest" can
+    //     be used to provide right bound.
+    // (2) key is in between smallest and largest.
+    //     Precalculated index based on "smallest > greatest" can be used to
+    //     provide left bound.
+    //     Precalculated index based on "largest < smallest" can be used to
+    //     provide right bound.
+    // (3) key is larger than largest, implying it is also larger than smallest.
+    //     Precalculated index based on "largest > largest" can be used to
+    //     provide left bound.
+    //
+    // As a result, we will need to do:
+    // Compare smallest (<=) and largest keys from upper level file with
+    // smallest key from lower level to get a right bound.
+    // Compare smallest (>=) and largest keys from upper level file with
+    // largest key from lower level to get a left bound.
+    //
+    // Example:
+    //    level 1:              [50 - 60]
+    //    level 2:        [1 - 40], [45 - 55], [58 - 80]
+    // A key 35, compared to be less than 50, 3rd file on level 2 can be
+    // skipped according to rule (1). LB = 0, RB = 1.
+    // A key 53, sits in the middle 50 and 60. 1st file on level 2 can be
+    // skipped according to rule (2)-a, but the 3rd file cannot be skipped
+    // because 60 is greater than 58. LB = 1, RB = 2.
+    // A key 70, compared to be larger than 60. 1st and 2nd file can be skipped
+    // according to rule (3). LB = 2, RB = 2.
+    //
+    // Point to a left most file in a lower level that may contain a key,
+    // which compares greater than smallest of a FileMetaData (upper level)
+    int32_t smallest_lb;
+    // Point to a left most file in a lower level that may contain a key,
+    // which compares greater than largest of a FileMetaData (upper level)
+    int32_t largest_lb;
+    // Point to a right most file in a lower level that may contain a key,
+    // which compares smaller than smallest of a FileMetaData (upper level)
+    int32_t smallest_rb;
+    // Point to a right most file in a lower level that may contain a key,
+    // which compares smaller than largest of a FileMetaData (upper level)
+    int32_t largest_rb;
+  };
+
+  void CalculateLB(const std::vector<FileMetaData*>& upper_files,
+    const std::vector<FileMetaData*>& lower_files,
+    std::vector<IndexUnit>* index,
+    std::function<int(const FileMetaData*, const FileMetaData*)> cmp_op,
+    std::function<void(IndexUnit*, int32_t)> set_index);
+
+  void CalculateRB(const std::vector<FileMetaData*>& upper_files,
+    const std::vector<FileMetaData*>& lower_files,
+    std::vector<IndexUnit>* index,
+    std::function<int(const FileMetaData*, const FileMetaData*)> cmp_op,
+    std::function<void(IndexUnit*, int32_t)> set_index);
+
+  std::vector<std::vector<IndexUnit>> next_level_index_;
+  std::vector<int32_t> level_rb_;
+};
+
+}  // namespace rocksdb

db/file_indexer_test.cc

@@ -0,0 +1,330 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+//
+// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file. See the AUTHORS file for names of contributors.
+
+#include <string>
+#include "db/file_indexer.h"
+#include "db/dbformat.h"
+#include "db/version_edit.h"
+#include "rocksdb/comparator.h"
+#include "util/testharness.h"
+#include "util/testutil.h"
+
+namespace rocksdb {
+
+class IntComparator : public Comparator {
+ public:
+  int Compare(const Slice& a, const Slice& b) const {
+    assert(a.size() == 8);
+    assert(b.size() == 8);
+    return *reinterpret_cast<const int64_t*>(a.data()) -
+      *reinterpret_cast<const int64_t*>(b.data());
+  }
+
+  const char* Name() const {
+    return "IntComparator";
+  }
+
+  void FindShortestSeparator(std::string* start, const Slice& limit) const {}
+
+  void FindShortSuccessor(std::string* key) const {}
+};
+
+
+struct FileIndexerTest {
+ public:
+  FileIndexerTest() :
+    kNumLevels(4), indexer(kNumLevels, &ucmp),
+    files(new std::vector<FileMetaData*>[kNumLevels]) {
+  }
+
+  ~FileIndexerTest() {
+    Reset();
+    delete[] files;
+  }
+
+  void AddFile(int level, int64_t smallest, int64_t largest) {
+    auto* f = new FileMetaData();
+    f->smallest = IntKey(smallest);
+    f->largest = IntKey(largest);
+    files[level].push_back(f);
+  }
+
+  InternalKey IntKey(int64_t v) {
+    return InternalKey(Slice(reinterpret_cast<char*>(&v), 8), 0, kTypeValue);
+  }
+
+  void Reset() {
+    for (int i = 0; i < kNumLevels; ++i) {
+      for (auto* f : files[i]) {
+        delete f;
+      }
+      files[i].clear();
+    }
+    indexer.ClearIndex();
+  }
+
+  void GetNextLevelIndex(const uint32_t level, const uint32_t file_index,
+      const int cmp_smallest, const int cmp_largest, int32_t* left_index,
+      int32_t* right_index) {
+    *left_index = 100;
+    *right_index = 100;
+    indexer.GetNextLevelIndex(level, file_index, cmp_smallest, cmp_largest,
+        left_index, right_index);
+  }
+
+  const uint32_t kNumLevels;
+  IntComparator ucmp;
+  FileIndexer indexer;
+
+  std::vector<FileMetaData*>* files;
+};
+
+TEST(FileIndexerTest, next_level_hint) {
+  for (uint32_t i = 0; i < kNumLevels; ++i) {
+    ASSERT_EQ(0, indexer.LevelIndexSize(i));
+  }
+
+  // Case 1: no overlap, files are on the left of next level files
+  // level 1
+  AddFile(1, 100, 200);
+  AddFile(1, 300, 400);
+  AddFile(1, 500, 600);
+  // level 2
+  AddFile(2, 1500, 1600);
+  AddFile(2, 1601, 1699);
+  AddFile(2, 1700, 1800);
+  // level 3
+  AddFile(3, 2500, 2600);
+  AddFile(3, 2601, 2699);
+  AddFile(3, 2700, 2800);
+  indexer.UpdateIndex(files);
+  int32_t left = 100;
+  int32_t right = 100;
+  for (uint32_t level = 1; level < 3; ++level) {
+    for (uint32_t f = 0; f < 3; ++f) {
+      GetNextLevelIndex(level, f, -1, -1, &left, &right);
+      ASSERT_EQ(0, left);
+      ASSERT_EQ(-1, right);
+      GetNextLevelIndex(level, f, 0, -1, &left, &right);
+      ASSERT_EQ(0, left);
+      ASSERT_EQ(-1, right);
+      GetNextLevelIndex(level, f, 1, -1, &left, &right);
+      ASSERT_EQ(0, left);
+      ASSERT_EQ(-1, right);
+      GetNextLevelIndex(level, f, 1, 0, &left, &right);
+      ASSERT_EQ(0, left);
+      ASSERT_EQ(-1, right);
+      GetNextLevelIndex(level, f, 1, 1, &left, &right);
+      ASSERT_EQ(0, left);
+      ASSERT_EQ(2, right);
+    }
+  }
+
+  // Case 2: no overlap, files are on the right of next level files
+  Reset();
+  for (uint32_t i = 1; i < kNumLevels; ++i) {
+    ASSERT_EQ(0, indexer.LevelIndexSize(i));
+  }
+  // level 1
+  AddFile(1, 2100, 2200);
+  AddFile(1, 2300, 2400);
+  AddFile(1, 2500, 2600);
+  // level 2
+  AddFile(2, 1500, 1600);
+  AddFile(2, 1501, 1699);
+  AddFile(2, 1700, 1800);
+  // level 3
+  AddFile(3, 500, 600);
+  AddFile(3, 501, 699);
+  AddFile(3, 700, 800);
+  indexer.UpdateIndex(files);
+  for (uint32_t level = 1; level < 3; ++level) {
+    for (uint32_t f = 0; f < 3; ++f) {
+      GetNextLevelIndex(level, f, -1, -1, &left, &right);
+      ASSERT_EQ(f == 0 ? 0 : 3, left);
+      ASSERT_EQ(2, right);
+      GetNextLevelIndex(level, f, 0, -1, &left, &right);
+      ASSERT_EQ(3, left);
+      ASSERT_EQ(2, right);
+      GetNextLevelIndex(level, f, 1, -1, &left, &right);
+      ASSERT_EQ(3, left);
+      ASSERT_EQ(2, right);
+      GetNextLevelIndex(level, f, 1, -1, &left, &right);
+      ASSERT_EQ(3, left);
+      ASSERT_EQ(2, right);
+      GetNextLevelIndex(level, f, 1, 0, &left, &right);
+      ASSERT_EQ(3, left);
+      ASSERT_EQ(2, right);
+      GetNextLevelIndex(level, f, 1, 1, &left, &right);
+      ASSERT_EQ(3, left);
+      ASSERT_EQ(2, right);
+    }
+  }
+
+  // Case 3: empty L2
+  Reset();
+  for (uint32_t i = 1; i < kNumLevels; ++i) {
+    ASSERT_EQ(0, indexer.LevelIndexSize(i));
+  }
+  // level 1
+  AddFile(1, 2100, 2200);
+  AddFile(1, 2300, 2400);
+  AddFile(1, 2500, 2600);
+  // level 3
+  AddFile(3, 500, 600);
+  AddFile(3, 501, 699);
+  AddFile(3, 700, 800);
+  indexer.UpdateIndex(files);
+  for (uint32_t f = 0; f < 3; ++f) {
+    GetNextLevelIndex(1, f, -1, -1, &left, &right);
+    ASSERT_EQ(0, left);
+    ASSERT_EQ(-1, right);
+    GetNextLevelIndex(1, f, 0, -1, &left, &right);
+    ASSERT_EQ(0, left);
+    ASSERT_EQ(-1, right);
+    GetNextLevelIndex(1, f, 1, -1, &left, &right);
+    ASSERT_EQ(0, left);
+    ASSERT_EQ(-1, right);
+    GetNextLevelIndex(1, f, 1, -1, &left, &right);
+    ASSERT_EQ(0, left);
+    ASSERT_EQ(-1, right);
+    GetNextLevelIndex(1, f, 1, 0, &left, &right);
+    ASSERT_EQ(0, left);
+    ASSERT_EQ(-1, right);
+    GetNextLevelIndex(1, f, 1, 1, &left, &right);
+    ASSERT_EQ(0, left);
+    ASSERT_EQ(-1, right);
+  }
+
+
+  // Case 4: mixed
+  Reset();
+  for (uint32_t i = 1; i < kNumLevels; ++i) {
+    ASSERT_EQ(0, indexer.LevelIndexSize(i));
+  }
+  // level 1
+  AddFile(1, 100, 200);
+  AddFile(1, 250, 400);
+  AddFile(1, 450, 500);
+  // level 2
+  AddFile(2, 100, 150);  // 0
+  AddFile(2, 200, 250);  // 1
+  AddFile(2, 251, 300);  // 2
+  AddFile(2, 301, 350);  // 3
+  AddFile(2, 500, 600);  // 4
+  // level 3
+  AddFile(3, 0, 50);
+  AddFile(3, 100, 200);
+  AddFile(3, 201, 250);
+  indexer.UpdateIndex(files);
+  // level 1, 0
+  GetNextLevelIndex(1, 0, -1, -1, &left, &right);
+  ASSERT_EQ(0, left);
+  ASSERT_EQ(0, right);
+  GetNextLevelIndex(1, 0, 0, -1, &left, &right);
+  ASSERT_EQ(0, left);
+  ASSERT_EQ(0, right);
+  GetNextLevelIndex(1, 0, 1, -1, &left, &right);
+  ASSERT_EQ(0, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(1, 0, 1, 0, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(1, 0, 1, 1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(4, right);
+  // level 1, 1
+  GetNextLevelIndex(1, 1, -1, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(1, 1, 0, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(1, 1, 1, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(3, right);
+  GetNextLevelIndex(1, 1, 1, 0, &left, &right);
+  ASSERT_EQ(4, left);
+  ASSERT_EQ(3, right);
+  GetNextLevelIndex(1, 1, 1, 1, &left, &right);
+  ASSERT_EQ(4, left);
+  ASSERT_EQ(4, right);
+  // level 1, 2
+  GetNextLevelIndex(1, 2, -1, -1, &left, &right);
+  ASSERT_EQ(4, left);
+  ASSERT_EQ(3, right);
+  GetNextLevelIndex(1, 2, 0, -1, &left, &right);
+  ASSERT_EQ(4, left);
+  ASSERT_EQ(3, right);
+  GetNextLevelIndex(1, 2, 1, -1, &left, &right);
+  ASSERT_EQ(4, left);
+  ASSERT_EQ(4, right);
+  GetNextLevelIndex(1, 2, 1, 0, &left, &right);
+  ASSERT_EQ(4, left);
+  ASSERT_EQ(4, right);
+  GetNextLevelIndex(1, 2, 1, 1, &left, &right);
+  ASSERT_EQ(4, left);
+  ASSERT_EQ(4, right);
+  // level 2, 0
+  GetNextLevelIndex(2, 0, -1, -1, &left, &right);
+  ASSERT_EQ(0, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(2, 0, 0, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(2, 0, 1, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(2, 0, 1, 0, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(2, 0, 1, 1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(2, right);
+  // level 2, 1
+  GetNextLevelIndex(2, 1, -1, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(2, 1, 0, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(1, right);
+  GetNextLevelIndex(2, 1, 1, -1, &left, &right);
+  ASSERT_EQ(1, left);
+  ASSERT_EQ(2, right);
+  GetNextLevelIndex(2, 1, 1, 0, &left, &right);
+  ASSERT_EQ(2, left);
+  ASSERT_EQ(2, right);
+  GetNextLevelIndex(2, 1, 1, 1, &left, &right);
+  ASSERT_EQ(2, left);
+  ASSERT_EQ(2, right);
+  // level 2, [2 - 4], no overlap
+  for (uint32_t f = 2; f <= 4; ++f) {
+    GetNextLevelIndex(2, f, -1, -1, &left, &right);
+    ASSERT_EQ(f == 2 ? 2 : 3, left);
+    ASSERT_EQ(2, right);
+    GetNextLevelIndex(2, f, 0, -1, &left, &right);
+    ASSERT_EQ(3, left);
+    ASSERT_EQ(2, right);
+    GetNextLevelIndex(2, f, 1, -1, &left, &right);
+    ASSERT_EQ(3, left);
+    ASSERT_EQ(2, right);
+    GetNextLevelIndex(2, f, 1, 0, &left, &right);
+    ASSERT_EQ(3, left);
+    ASSERT_EQ(2, right);
+    GetNextLevelIndex(2, f, 1, 1, &left, &right);
+    ASSERT_EQ(3, left);
+    ASSERT_EQ(2, right);
+  }
+}
+
+}  // namespace rocksdb
+
+int main(int argc, char** argv) {
+  return rocksdb::test::RunAllTests();
+}

db/version_set.cc

@@ -71,11 +71,11 @@ Version::~Version() {
   delete[] files_;
 }
 
-int FindFile(const InternalKeyComparator& icmp,
+int FindFileInRange(const InternalKeyComparator& icmp,
-             const std::vector<FileMetaData*>& files,
+    const std::vector<FileMetaData*>& files,
-             const Slice& key) {
+    const Slice& key,
-  uint32_t left = 0;
+    uint32_t left,
-  uint32_t right = files.size();
+    uint32_t right) {
   while (left < right) {
     uint32_t mid = (left + right) / 2;
     const FileMetaData* f = files[mid];
@@ -92,6 +92,12 @@ int FindFile(const InternalKeyComparator& icmp,
   return right;
 }
 
+int FindFile(const InternalKeyComparator& icmp,
+             const std::vector<FileMetaData*>& files,
+             const Slice& key) {
+  return FindFileInRange(icmp, files, key, 0, files.size());
+}
+
 static bool AfterFile(const Comparator* ucmp,
                       const Slice* user_key, const FileMetaData* f) {
   // nullptr user_key occurs before all keys and is therefore never after *f
@@ -507,7 +513,10 @@ Version::Version(ColumnFamilyData* cfd, VersionSet* vset,
       file_to_compact_level_(-1),
       compaction_score_(num_levels_),
       compaction_level_(num_levels_),
-      version_number_(version_number) {}
+      version_number_(version_number),
+      file_indexer_(num_levels_, cfd == nullptr ?  nullptr
+          : cfd->internal_comparator().user_comparator()) {
+}
 
 void Version::Get(const ReadOptions& options,
                   const LookupKey& k,
@@ -538,12 +547,27 @@ void Version::Get(const ReadOptions& options,
   int last_file_read_level = -1;
 
   // We can search level-by-level since entries never hop across
-  // levels.  Therefore we are guaranteed that if we find data
+  // levels. Therefore we are guaranteed that if we find data
   // in an smaller level, later levels are irrelevant (unless we
   // are MergeInProgress).
-  for (int level = 0; level < num_levels_; level++) {
+
-    size_t num_files = files_[level].size();
+  int32_t search_left_bound = 0;
-    if (num_files == 0) continue;
+  int32_t search_right_bound = FileIndexer::kLevelMaxIndex;
+  for (int level = 0; level < num_levels_; ++level) {
+    int num_files = files_[level].size();
+    if (num_files == 0) {
+      // When current level is empty, the search bound generated from upper
+      // level must be [0, -1] or [0, FileIndexer::kLevelMaxIndex] if it is
+      // also empty.
+      assert(search_left_bound == 0);
+      assert(search_right_bound == -1 ||
+             search_right_bound == FileIndexer::kLevelMaxIndex);
+      // Since current level is empty, it will need to search all files in the
+      // next level
+      search_left_bound = 0;
+      search_right_bound = FileIndexer::kLevelMaxIndex;
+      continue;
+    }
 
     // Get the list of files to search in this level
     FileMetaData* const* files = &files_[level][0];
@@ -553,38 +577,65 @@ void Version::Get(const ReadOptions& options,
     // newest to oldest. In the context of merge-operator,
     // this can occur at any level. Otherwise, it only occurs
     // at Level-0 (since Put/Deletes are always compacted into a single entry).
-    uint32_t start_index;
+    int32_t start_index;
     if (level == 0) {
       // On Level-0, we read through all files to check for overlap.
       start_index = 0;
     } else {
-      // On Level-n (n>=1), files are sorted.
+      // On Level-n (n>=1), files are sorted. Binary search to find the earliest
-      // Binary search to find earliest index whose largest key >= ikey.
+      // file whose largest key >= ikey. Search left bound and right bound are
-      // We will also stop when the file no longer overlaps ikey
+      // used to narrow the range.
-      start_index = FindFile(*internal_comparator_, files_[level], ikey);
+      if (search_left_bound == search_right_bound) {
+        start_index = search_left_bound;
+      } else if (search_left_bound < search_right_bound) {
+        if (search_right_bound == FileIndexer::kLevelMaxIndex) {
+          search_right_bound = num_files - 1;
+        }
+        start_index = FindFileInRange(cfd_->internal_comparator(),
+            files_[level], ikey, search_left_bound, search_right_bound);
+      } else {
+        // search_left_bound > search_right_bound, key does not exist in this
+        // level. Since no comparision is done in this level, it will need to
+        // search all files in the next level.
+        search_left_bound = 0;
+        search_right_bound = FileIndexer::kLevelMaxIndex;
+        continue;
+      }
     }
-
     // Traverse each relevant file to find the desired key
 #ifndef NDEBUG
     FileMetaData* prev_file = nullptr;
 #endif
-    for (uint32_t i = start_index; i < num_files; ++i) {
+
+    for (int32_t i = start_index; i < num_files;) {
       FileMetaData* f = files[i];
-      // Skip key range filtering for levle 0 if there are few level 0 files.
+      // Check if key is within a file's range. If search left bound and right
-      if ((level > 0 || num_files > 2) &&
+      // bound point to the same find, we are sure key falls in range.
-          (user_comparator_->Compare(user_key, f->smallest.user_key()) < 0 ||
+      assert(level == 0 || i == start_index ||
-           user_comparator_->Compare(user_key, f->largest.user_key()) > 0)) {
+             user_comparator_->Compare(user_key, f->smallest.user_key()) <= 0);
-        // Only process overlapping files.
+
-        if (level > 0) {
+      int cmp_smallest = user_comparator_->Compare(user_key, f->smallest.user_key());
-          // If on Level-n (n>=1) then the files are sorted.
+      int cmp_largest = -1;
-          // So we can stop looking when we are past the ikey.
+      if (cmp_smallest >= 0) {
+        cmp_largest = user_comparator_->Compare(user_key, f->largest.user_key());
+      }
+
+      // Setup file search bound for the next level based on the comparison
+      // results
+      if (level > 0) {
+        file_indexer_.GetNextLevelIndex(level, i, cmp_smallest, cmp_largest,
+            &search_left_bound, &search_right_bound);
+      }
+      // Key falls out of current file's range
+      if (cmp_smallest < 0 || cmp_largest > 0) {
+        if (level == 0) {
+          ++i;
+          continue;
+        } else {
           break;
         }
-        // TODO: do we want to check file ranges for level0 files at all?
-        // For new SST format where Get() is fast, we might want to consider
-        // to avoid those two comparisons, if it can filter out too few files.
-        continue;
       }
+
 #ifndef NDEBUG
       // Sanity check to make sure that the files are correctly sorted
       if (prev_file) {
@@ -643,6 +694,11 @@ void Version::Get(const ReadOptions& options,
         case kMerge:
           break;
       }
+      if (level > 0 && cmp_largest < 0) {
+        break;
+      } else {
+        ++i;
+      }
     }
   }
 
@@ -1454,6 +1510,8 @@ class VersionSet::Builder {
     }
 
     CheckConsistency(v);
+
+    v->file_indexer_.UpdateIndex(v->files_);
   }
 
   void LoadTableHandlers() {

db/version_set.h

@@ -33,6 +33,7 @@
 #include "db/compaction_picker.h"
 #include "db/column_family.h"
 #include "db/log_reader.h"
+#include "db/file_indexer.h"
 
 namespace rocksdb {
 
@@ -281,6 +282,7 @@ class Version {
   uint64_t version_number_;
 
   Version(ColumnFamilyData* cfd, VersionSet* vset, uint64_t version_number = 0);
+  FileIndexer file_indexer_;
 
   ~Version();
 

tools/db_stress.cc

@@ -59,6 +59,7 @@ static bool ValidateUint32Range(const char* flagname, uint64_t value) {
   }
   return true;
 }
+
 DEFINE_uint64(seed, 2341234, "Seed for PRNG");
 static const bool FLAGS_seed_dummy __attribute__((unused)) =
     google::RegisterFlagValidator(&FLAGS_seed, &ValidateUint32Range);
@@ -377,6 +378,17 @@ static std::string Key(long val) {
   return big_endian_key;
 }
 
+static std::string StringToHex(const std::string& str) {
+  std::string result = "0x";
+  char buf[10];
+  for (size_t i = 0; i < str.length(); i++) {
+    snprintf(buf, 10, "%02X", (unsigned char)str[i]);
+    result += buf;
+  }
+  return result;
+}
+
+
 class StressTest;
 namespace {
 
@@ -953,8 +965,8 @@ class StressTest {
     for (int i = 1; i < 10; i++) {
       if (values[i] != values[0]) {
         fprintf(stderr, "error : inconsistent values for key %s: %s, %s\n",
-                key.ToString().c_str(), values[0].c_str(),
+                key.ToString(true).c_str(), StringToHex(values[0]).c_str(),
-                values[i].c_str());
+                StringToHex(values[i]).c_str());
       // we continue after error rather than exiting so that we can
       // find more errors if any
       }
@@ -1013,9 +1025,9 @@ class StressTest {
       // make sure all values are equivalent
       for (int i = 0; i < 10; i++) {
         if (values[i] != values[0]) {
-          fprintf(stderr, "error : inconsistent values for prefix %s: %s, %s\n",
+          fprintf(stderr, "error : %d, inconsistent values for prefix %s: %s, %s\n",
-                  prefixes[i].c_str(), values[0].c_str(),
+                  i, prefixes[i].c_str(), StringToHex(values[0]).c_str(),
-                  values[i].c_str());
+                  StringToHex(values[i]).c_str());
           // we continue after error rather than exiting so that we can
           // find more errors if any
         }