Rewritten system for scheduling background work

Summary:
When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue.

The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction.

Here are the performance results:

Command:

    ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000  --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333

Before the patch:

     fillrandom   :      26.950 micros/op 37105 ops/sec;    4.1 MB/s

After the patch:

      fillrandom   :      17.404 micros/op 57456 ops/sec;    6.4 MB/s

Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got:

      fillrandom   :       7.590 micros/op 131758 ops/sec;   14.6 MB/s

Test Plan:
make check

two stress tests:

Big number of compactions and flushes:

    ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0  --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000

max_background_flushes=0, to verify that this case also works correctly

    ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0  --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000

Reviewers: ljin, rven, yhchiang, sdong

Reviewed By: sdong

Subscribers: dhruba, leveldb

Differential Revision: https://reviews.facebook.net/D30123

db/column_family.cc

@@ -223,14 +223,11 @@ void SuperVersionUnrefHandle(void* ptr) {
 }
 }  // anonymous namespace
 
-ColumnFamilyData::ColumnFamilyData(uint32_t id, const std::string& name,
-                                   Version* _dummy_versions,
-                                   Cache* _table_cache,
-                                   WriteBuffer* write_buffer,
-                                   const ColumnFamilyOptions& cf_options,
-                                   const DBOptions* db_options,
-                                   const EnvOptions& env_options,
-                                   ColumnFamilySet* column_family_set)
+ColumnFamilyData::ColumnFamilyData(
+    uint32_t id, const std::string& name, Version* _dummy_versions,
+    Cache* _table_cache, WriteBuffer* write_buffer,
+    const ColumnFamilyOptions& cf_options, const DBOptions* db_options,
+    const EnvOptions& env_options, ColumnFamilySet* column_family_set)
     : id_(id),
       name_(name),
       dummy_versions_(_dummy_versions),
@@ -250,7 +247,9 @@ ColumnFamilyData::ColumnFamilyData(uint32_t id, const std::string& name,
       next_(nullptr),
       prev_(nullptr),
       log_number_(0),
-      column_family_set_(column_family_set) {
+      column_family_set_(column_family_set),
+      pending_flush_(false),
+      pending_compaction_(false) {
   Ref();
 
   // if _dummy_versions is nullptr, then this is a dummy column family.
@@ -285,10 +284,14 @@ ColumnFamilyData::ColumnFamilyData(uint32_t id, const std::string& name,
           new LevelCompactionPicker(ioptions_, &internal_comparator_));
     }
 
-    Log(InfoLogLevel::INFO_LEVEL,
-        ioptions_.info_log, "Options for column family \"%s\":\n",
-        name.c_str());
-    options_.Dump(ioptions_.info_log);
+    if (column_family_set_->NumberOfColumnFamilies() < 10) {
+      Log(InfoLogLevel::INFO_LEVEL, ioptions_.info_log,
+          "--------------- Options for column family [%s]:\n", name.c_str());
+      options_.Dump(ioptions_.info_log);
+    } else {
+      Log(InfoLogLevel::INFO_LEVEL, ioptions_.info_log,
+          "\t(skipping printing options)\n");
+    }
   }
 
   RecalculateWriteStallConditions(mutable_cf_options_);
@@ -313,6 +316,11 @@ ColumnFamilyData::~ColumnFamilyData() {
     current_->Unref();
   }
 
+  // It would be wrong if this ColumnFamilyData is in flush_queue_ or
+  // compaction_queue_ and we destroyed it
+  assert(!pending_flush_);
+  assert(!pending_compaction_);
+
   if (super_version_ != nullptr) {
     // Release SuperVersion reference kept in ThreadLocalPtr.
     // This must be done outside of mutex_ since unref handler can lock mutex.
@@ -434,6 +442,10 @@ void ColumnFamilyData::CreateNewMemtable(
   mem_->Ref();
 }
 
+bool ColumnFamilyData::NeedsCompaction() const {
+  return compaction_picker_->NeedsCompaction(current_->storage_info());
+}
+
 Compaction* ColumnFamilyData::PickCompaction(
     const MutableCFOptions& mutable_options, LogBuffer* log_buffer) {
   auto* result = compaction_picker_->PickCompaction(

db/column_family.h

@@ -210,8 +210,11 @@ class ColumnFamilyData {
 
   // See documentation in compaction_picker.h
   // REQUIRES: DB mutex held
+  bool NeedsCompaction() const;
+  // REQUIRES: DB mutex held
   Compaction* PickCompaction(const MutableCFOptions& mutable_options,
                              LogBuffer* log_buffer);
+  // REQUIRES: DB mutex held
   Compaction* CompactRange(
       const MutableCFOptions& mutable_cf_options,
       int input_level, int output_level, uint32_t output_path_id,
@@ -248,6 +251,7 @@ class ColumnFamilyData {
   // if its reference count is zero and needs deletion or nullptr if not
   // As argument takes a pointer to allocated SuperVersion to enable
   // the clients to allocate SuperVersion outside of mutex.
+  // IMPORTANT: Only call this from DBImpl::InstallSuperVersion()
   SuperVersion* InstallSuperVersion(SuperVersion* new_superversion,
                                     port::Mutex* db_mutex,
                                     const MutableCFOptions& mutable_cf_options);
@@ -261,6 +265,12 @@ class ColumnFamilyData {
       bool triggered_flush_slowdown,
       bool triggered_flush_stop);
 
+  // Protected by DB mutex
+  void set_pending_flush(bool value) { pending_flush_ = value; }
+  void set_pending_compaction(bool value) { pending_compaction_ = value; }
+  bool pending_flush() { return pending_flush_; }
+  bool pending_compaction() { return pending_compaction_; }
+
  private:
   friend class ColumnFamilySet;
   ColumnFamilyData(uint32_t id, const std::string& name,
@@ -328,6 +338,13 @@ class ColumnFamilyData {
   ColumnFamilySet* column_family_set_;
 
   std::unique_ptr<WriteControllerToken> write_controller_token_;
+
+  // If true --> this ColumnFamily is currently present in DBImpl::flush_queue_
+  bool pending_flush_;
+
+  // If true --> this ColumnFamily is currently present in
+  // DBImpl::compaction_queue_
+  bool pending_compaction_;
 };
 
 // ColumnFamilySet has interesting thread-safety requirements

db/compaction_picker.cc

@@ -695,15 +695,6 @@ Compaction* LevelCompactionPicker::PickCompaction(
   Compaction* c = nullptr;
   int level = -1;
 
-  // Compute the compactions needed. It is better to do it here
-  // and also in LogAndApply(), otherwise the values could be stale.
-  std::vector<uint64_t> size_being_compacted(NumberLevels() - 1);
-  SizeBeingCompacted(size_being_compacted);
-
-  CompactionOptionsFIFO dummy_compaction_options_fifo;
-  vstorage->ComputeCompactionScore(
-      mutable_cf_options, dummy_compaction_options_fifo, size_being_compacted);
-
   // We prefer compactions triggered by too much data in a level over
   // the compactions triggered by seeks.
   //
@@ -766,6 +757,21 @@ Compaction* LevelCompactionPicker::PickCompaction(
   compactions_in_progress_[level].insert(c);
 
   c->mutable_cf_options_ = mutable_cf_options;
+
+  // Creating a compaction influences the compaction score because the score
+  // takes running compactions into account (by skipping files that are already
+  // being compacted). Since we just changed compaction score, we recalculate it
+  // here
+  {  // this piece of code recomputes compaction score
+    std::vector<uint64_t> size_being_compacted(NumberLevels() - 1);
+    SizeBeingCompacted(size_being_compacted);
+
+    CompactionOptionsFIFO dummy_compaction_options_fifo;
+    vstorage->ComputeCompactionScore(mutable_cf_options,
+                                     dummy_compaction_options_fifo,
+                                     size_being_compacted);
+  }
+
   return c;
 }
 

db/db_impl.h

@@ -362,6 +362,8 @@ class DBImpl : public DB {
   ColumnFamilyData* GetColumnFamilyDataByName(const std::string& cf_name);
 
   void MaybeScheduleFlushOrCompaction();
+  void SchedulePendingFlush(ColumnFamilyData* cfd);
+  void SchedulePendingCompaction(ColumnFamilyData* cfd);
   static void BGWorkCompaction(void* db);
   static void BGWorkFlush(void* db);
   void BackgroundCallCompaction();
@@ -393,6 +395,12 @@ class DBImpl : public DB {
   // hold the data set.
   Status ReFitLevel(ColumnFamilyData* cfd, int level, int target_level = -1);
 
+  // helper functions for adding and removing from flush & compaction queues
+  void AddToCompactionQueue(ColumnFamilyData* cfd);
+  ColumnFamilyData* PopFirstFromCompactionQueue();
+  void AddToFlushQueue(ColumnFamilyData* cfd);
+  ColumnFamilyData* PopFirstFromFlushQueue();
+
   // table_cache_ provides its own synchronization
   std::shared_ptr<Cache> table_cache_;
 
@@ -460,9 +468,32 @@ class DBImpl : public DB {
   // State is protected with db mutex.
   std::list<uint64_t> pending_outputs_;
 
-  // At least one compaction or flush job is pending but not yet scheduled
-  // because of the max background thread limit.
-  bool bg_schedule_needed_;
+  // flush_queue_ and compaction_queue_ hold column families that we need to
+  // flush and compact, respectively.
+  // A column family is inserted into flush_queue_ when it satisfies condition
+  // cfd->imm()->IsFlushPending()
+  // A column family is inserted into compaction_queue_ when it satisfied
+  // condition cfd->NeedsCompaction()
+  // Column families in this list are all Ref()-erenced
+  // TODO(icanadi) Provide some kind of ReferencedColumnFamily class that will
+  // do RAII on ColumnFamilyData
+  // Column families are in this queue when they need to be flushed or
+  // compacted. Consumers of these queues are flush and compaction threads. When
+  // column family is put on this queue, we increase unscheduled_flushes_ and
+  // unscheduled_compactions_. When these variables are bigger than zero, that
+  // means we need to schedule background threads for compaction and thread.
+  // Once the background threads are scheduled, we decrease unscheduled_flushes_
+  // and unscheduled_compactions_. That way we keep track of number of
+  // compaction and flush threads we need to schedule. This scheduling is done
+  // in MaybeScheduleFlushOrCompaction()
+  // invariant(column family present in flush_queue_ <==>
+  // ColumnFamilyData::pending_flush_ == true)
+  std::deque<ColumnFamilyData*> flush_queue_;
+  // invariant(column family present in compaction_queue_ <==>
+  // ColumnFamilyData::pending_compaction_ == true)
+  std::deque<ColumnFamilyData*> compaction_queue_;
+  int unscheduled_flushes_;
+  int unscheduled_compactions_;
 
   // count how many background compactions are running or have been scheduled
   int bg_compaction_scheduled_;
@@ -553,9 +584,17 @@ class DBImpl : public DB {
       ColumnFamilyData* cfd, JobContext* job_context,
       const MutableCFOptions& mutable_cf_options);
 
-  SuperVersion* InstallSuperVersion(
-    ColumnFamilyData* cfd, SuperVersion* new_sv,
-    const MutableCFOptions& mutable_cf_options);
+  // All ColumnFamily state changes go through this function. Here we analyze
+  // the new state and we schedule background work if we detect that the new
+  // state needs flush or compaction.
+  // If dont_schedule_bg_work == true, then caller asks us to not schedule flush
+  // or compaction here, but it also promises to schedule needed background
+  // work. We use this to  scheduling background compactions when we are in the
+  // write thread, which is very performance critical. Caller schedules
+  // background work as soon as it exits the write thread
+  SuperVersion* InstallSuperVersion(ColumnFamilyData* cfd, SuperVersion* new_sv,
+                                    const MutableCFOptions& mutable_cf_options,
+                                    bool dont_schedule_bg_work = false);
 
   // Find Super version and reference it. Based on options, it might return
   // the thread local cached one.