Merge branch 'performance' of github.com:facebook/rocksdb into performance

db/db_impl.cc

@@ -1238,13 +1238,17 @@ void DBImpl::TEST_CompactRange(int level, const Slice* begin,const Slice* end) {
   manual.level = level;
   manual.done = false;
   manual.in_progress = false;
-  if (begin == nullptr) {
+  // For universal compaction, we enforce every manual compaction to compact
+  // all files.
+  if (begin == nullptr ||
+      options_.compaction_style == kCompactionStyleUniversal) {
     manual.begin = nullptr;
   } else {
     begin_storage = InternalKey(*begin, kMaxSequenceNumber, kValueTypeForSeek);
     manual.begin = &begin_storage;
   }
-  if (end == nullptr) {
+  if (end == nullptr ||
+      options_.compaction_style == kCompactionStyleUniversal) {
     manual.end = nullptr;
   } else {
     end_storage = InternalKey(*end, 0, static_cast<ValueType>(0));
@@ -1498,6 +1502,18 @@ Status DBImpl::BackgroundCompaction(bool* madeProgress,
     if (!status.ok()) {
       m->done = true;
     }
+    // For universal compaction:
+    //   Because universal compaction always happens at level 0, so one
+    //   compaction will pick up all overlapped files. No files will be
+    //   filtered out due to size limit and left for a successive compaction.
+    //   So we can safely conclude the current compaction.
+    //
+    //   Also note that, if we don't stop here, then the current compaction
+    //   writes a new file back to level 0, which will be used in successive
+    //   compaction. Hence the manual compaction will never finish.
+    if (options_.compaction_style == kCompactionStyleUniversal) {
+      m->done = true;
+    }
     if (!m->done) {
       // We only compacted part of the requested range.  Update *m
       // to the range that is left to be compacted.
@@ -1745,14 +1761,7 @@ Status DBImpl::DoCompactionWork(CompactionState* compact) {
   }
 
   // Is this compaction producing files at the bottommost level?
-  bool bottommost_level = true;
-  for (int i = compact->compaction->level() + 2;
-       i < versions_->NumberLevels(); i++) {
-    if (versions_->NumLevelFiles(i) > 0) {
-      bottommost_level = false;
-      break;
-    }
-  }
+  bool bottommost_level = compact->compaction->BottomMostLevel();
 
   // Allocate the output file numbers before we release the lock
   AllocateCompactionOutputFileNumbers(compact);
@@ -2088,7 +2097,7 @@ Status DBImpl::DoCompactionWork(CompactionState* compact) {
       versions_->LevelSummary(&tmp),
       (stats.bytes_readn + stats.bytes_readnp1 + stats.bytes_written) /
           (double) stats.micros,
-      compact->compaction->level() + 1,
+      compact->compaction->output_level(),
       stats.files_in_leveln, stats.files_in_levelnp1, stats.files_out_levelnp1,
       stats.bytes_readn / 1048576.0,
       stats.bytes_readnp1 / 1048576.0,

db/version_set.cc

@@ -2266,6 +2266,13 @@ Compaction* VersionSet::PickCompactionUniversal(int level, double score) {
     newerfile = f;
   }
 
+  // Is the earliest file part of this compaction?
+  int last_index = file_by_time[file_by_time.size()-1];
+  FileMetaData* last_file = current_->files_[level][last_index];
+  if (c->inputs_[0][c->inputs_[0].size()-1] == last_file) {
+    c->bottommost_level_ = true;
+  }
+
   // update statistics
   if (options_->statistics != nullptr) {
     options_->statistics->measureTime(NUM_FILES_IN_SINGLE_COMPACTION,
@@ -2403,7 +2410,7 @@ Compaction* VersionSet::PickCompaction() {
     if (level != 0 || compactions_in_progress_[0].empty()) {
       if(!ParentRangeInCompaction(&f->smallest, &f->largest, level,
                                   &parent_index)) {
-        c = new Compaction(level, level, MaxFileSizeForLevel(level+1),
+        c = new Compaction(level, level+1, MaxFileSizeForLevel(level+1),
                 MaxGrandParentOverlapBytes(level), NumberLevels(), true);
         c->inputs_[0].push_back(f);
         c->parent_index_ = parent_index;
@@ -2444,13 +2451,15 @@ Compaction* VersionSet::PickCompaction() {
     assert(!c->inputs_[0].empty());
   }
 
-
   // Setup "level+1" files (inputs_[1])
   SetupOtherInputs(c);
 
   // mark all the files that are being compacted
   c->MarkFilesBeingCompacted(true);
 
+  // Is this compaction creating a file at the bottommost level
+  c->SetupBottomMostLevel(false);
+
   // remember this currently undergoing compaction
   compactions_in_progress_[level].insert(c);
 
@@ -2624,6 +2633,13 @@ Compaction* VersionSet::CompactRange(
     const InternalKey* begin,
     const InternalKey* end) {
   std::vector<FileMetaData*> inputs;
+
+  // All files are 'overlapping' in universal style compaction.
+  // We have to compact the entire range in one shot.
+  if (options_->compaction_style == kCompactionStyleUniversal) {
+    begin = nullptr;
+    end = nullptr;
+  }
   current_->GetOverlappingInputs(level, begin, end, &inputs);
   if (inputs.empty()) {
     return nullptr;
@@ -2667,6 +2683,9 @@ Compaction* VersionSet::CompactRange(
   // upon other files because manual compactions are processed when
   // the system has a max of 1 background compaction thread.
   c->MarkFilesBeingCompacted(true);
+
+  // Is this compaction creating a file at the bottommost level
+  c->SetupBottomMostLevel(true);
   return c;
 }
 
@@ -2686,6 +2705,7 @@ Compaction::Compaction(int level, int out_level, uint64_t target_file_size,
       base_index_(-1),
       parent_index_(-1),
       score_(0),
+      bottommost_level_(false),
       level_ptrs_(std::vector<size_t>(number_levels)) {
   edit_ = new VersionEdit(number_levels_);
   for (int i = 0; i < number_levels_; i++) {
@@ -2718,6 +2738,10 @@ void Compaction::AddInputDeletions(VersionEdit* edit) {
 }
 
 bool Compaction::IsBaseLevelForKey(const Slice& user_key) {
+  if (input_version_->vset_->options_->compaction_style ==
+      kCompactionStyleUniversal) {
+    return bottommost_level_;
+  }
   // Maybe use binary search to find right entry instead of linear search?
   const Comparator* user_cmp = input_version_->vset_->icmp_.user_comparator();
   for (int lvl = level_ + 2; lvl < number_levels_; lvl++) {
@@ -2776,6 +2800,31 @@ void Compaction::MarkFilesBeingCompacted(bool value) {
   }
 }
 
+// Is this compaction producing files at the bottommost level?
+void Compaction::SetupBottomMostLevel(bool isManual) {
+  if (input_version_->vset_->options_->compaction_style  ==
+         kCompactionStyleUniversal) {
+    // If universal compaction style is used and manual
+    // compaction is occuring, then we are guaranteed that
+    // all files will be picked in a single compaction
+    // run. We can safely set bottommost_level_ = true.
+    // If it is not manual compaction, then bottommost_level_
+    // is already set when the Compaction was created.
+    if (isManual) {
+      bottommost_level_ = true;
+    }
+    return;
+  }
+  bottommost_level_ = true;
+  int num_levels = input_version_->vset_->NumberLevels();
+  for (int i = level() + 2; i < num_levels; i++) {
+    if (input_version_->vset_->NumLevelFiles(i) > 0) {
+      bottommost_level_ = false;
+      break;
+    }
+  }
+}
+
 void Compaction::ReleaseInputs() {
   if (input_version_ != nullptr) {
     input_version_->Unref();

db/version_set.h

@@ -557,6 +557,9 @@ class Compaction {
   // Return the score that was used to pick this compaction run.
   double score() const { return score_; }
 
+  // Is this compaction creating a file in the bottom most level?
+  bool BottomMostLevel() { return bottommost_level_; }
+
  private:
   friend class Version;
   friend class VersionSet;
@@ -589,7 +592,8 @@ class Compaction {
   int parent_index_; // index of some file with same range in files_[level_+1]
   double score_;     // score that was used to pick this compaction.
 
-  // State for implementing IsBaseLevelForKey
+  // Is this compaction creating a file in the bottom most level?
+  bool bottommost_level_;
 
   // level_ptrs_ holds indices into input_version_->levels_: our state
   // is that we are positioned at one of the file ranges for each
@@ -600,6 +604,9 @@ class Compaction {
   // mark (or clear) all files that are being compacted
   void MarkFilesBeingCompacted(bool);
 
+  // Initialize whether compaction producing files at the bottommost level
+  void SetupBottomMostLevel(bool isManual);
+
   // In case of compaction error, reset the nextIndex that is used
   // to pick up the next file to be compacted from files_by_size_
   void ResetNextCompactionIndex();

util/options.cc

@@ -185,7 +185,7 @@ Options::Dump(Logger* log) const
         max_background_compactions);
     Log(log,"                      Options.hard_rate_limit: %.2f",
         hard_rate_limit);
-    Log(log,"      Options.rate_limit_delay_max_milliseconds: %d",
+    Log(log,"      Options.rate_limit_delay_max_milliseconds: %u",
         rate_limit_delay_max_milliseconds);
     Log(log,"               Options.disable_auto_compactions: %d",
         disable_auto_compactions);
@@ -205,7 +205,7 @@ Options::Dump(Logger* log) const
         is_fd_close_on_exec);
     Log(log,"              Options.skip_log_error_on_recovery: %d",
         skip_log_error_on_recovery);
-    Log(log,"                   Options.stats_dump_period_sec: %d",
+    Log(log,"                   Options.stats_dump_period_sec: %u",
         stats_dump_period_sec);
     Log(log,"                    Options.block_size_deviation: %d",
         block_size_deviation);
@@ -221,11 +221,11 @@ Options::Dump(Logger* log) const
         filter_deletes);
     Log(log,"                        Options.compaction_style: %d",
         compaction_style);
-    Log(log,"        Options.compaction_options_universal.size_ratio: %d",
+    Log(log,"        Options.compaction_options_universal.size_ratio: %u",
         compaction_options_universal.size_ratio);
-    Log(log,"   Options.compaction_options_universal.min_merge_width: %d",
+    Log(log,"   Options.compaction_options_universal.min_merge_width: %u",
         compaction_options_universal.min_merge_width);
-    Log(log,"   Options.compaction_options_universal.max_merge_width: %d",
+    Log(log,"   Options.compaction_options_universal.max_merge_width: %u",
         compaction_options_universal.max_merge_width);
 }   // Options::Dump