Make BlockBasedTable::kMaxAutoReadAheadSize configurable (#7951)

Summary:
RocksDB does auto-readahead for iterators on noticing more
than two reads for a table file. The readahead starts at 8KB and doubles on every
additional read upto BlockBasedTable::kMaxAutoReadAheadSize which is
256*1024.
This PR adds a new option BlockBasedTableOptions::max_auto_readahead_size which
replaces BlockBasedTable::kMaxAutoReadAheadSize and the new option can be
configured.
If max_auto_readahead_size is set 0 then no implicit auto prefetching will
be done. If max_auto_readahead_size provided is less than
8KB (which is initial readahead size used by rocksdb in case of
auto-readahead), readahead size will remain same as max_auto_readahead_size.

Pull Request resolved: #7951

Test Plan: Add new unit test case.

Reviewed By: anand1976

Differential Revision: D26568085

Pulled By: akankshamahajan15

fbshipit-source-id: b6543520fc74e97d859f2002328d4c5254d417af

HISTORY.md

@@ -1,5 +1,7 @@
 # Rocksdb Change Log
 ## Unreleased
+### Public API change
+* Add a new option BlockBasedTableOptions::max_auto_readahead_size. RocksDB does auto-readahead for iterators on noticing more than two reads for a table file if user doesn't provide readahead_size. The readahead starts at 8KB and doubles on every additional read upto max_auto_readahead_size and now max_auto_readahead_size can be configured dynamically as well. Found that 256 KB readahead size provides the best performance, based on experiments, for auto readahead. Experiment data is in PR #3282. If value is set 0 then no automatic prefetching will be done by rocksdb. Also changing the value will only affect files opened after the change.
 
 ## 6.18.0 (02/19/2021)
 ### Behavior Changes

file/prefetch_test.cc

@@ -175,9 +175,154 @@ TEST_P(PrefetchTest, Basic) {
   Close();
 }
 
+#ifndef ROCKSDB_LITE
+TEST_P(PrefetchTest, ConfigureAutoMaxReadaheadSize) {
+  // First param is if the mockFS support_prefetch or not
+  bool support_prefetch =
+      std::get<0>(GetParam()) &&
+      test::IsPrefetchSupported(env_->GetFileSystem(), dbname_);
+
+  // Second param is if directIO is enabled or not
+  bool use_direct_io = std::get<1>(GetParam());
+
+  std::shared_ptr<MockFS> fs =
+      std::make_shared<MockFS>(env_->GetFileSystem(), support_prefetch);
+  std::unique_ptr<Env> env(new CompositeEnvWrapper(env_, fs));
+
+  Options options = CurrentOptions();
+  options.write_buffer_size = 1024;
+  options.create_if_missing = true;
+  options.compression = kNoCompression;
+  options.env = env.get();
+  options.disable_auto_compactions = true;
+  if (use_direct_io) {
+    options.use_direct_reads = true;
+    options.use_direct_io_for_flush_and_compaction = true;
+  }
+  BlockBasedTableOptions table_options;
+  table_options.no_block_cache = true;
+  table_options.cache_index_and_filter_blocks = false;
+  table_options.metadata_block_size = 1024;
+  table_options.index_type =
+      BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
+  table_options.max_auto_readahead_size = 0;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+
+  int buff_prefetch_count = 0;
+  SyncPoint::GetInstance()->SetCallBack("FilePrefetchBuffer::Prefetch:Start",
+                                        [&](void*) { buff_prefetch_count++; });
+
+  // DB open will create table readers unless we reduce the table cache
+  // capacity. SanitizeOptions will set max_open_files to minimum of 20. Table
+  // cache is allocated with max_open_files - 10 as capacity. So override
+  // max_open_files to 10 so table cache capacity will become 0. This will
+  // prevent file open during DB open and force the file to be opened during
+  // Iteration.
+  SyncPoint::GetInstance()->SetCallBack(
+      "SanitizeOptions::AfterChangeMaxOpenFiles", [&](void* arg) {
+        int* max_open_files = (int*)arg;
+        *max_open_files = 11;
+      });
+
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  Status s = TryReopen(options);
+
+  if (use_direct_io && (s.IsNotSupported() || s.IsInvalidArgument())) {
+    // If direct IO is not supported, skip the test
+    return;
+  } else {
+    ASSERT_OK(s);
+  }
+
+  Random rnd(309);
+  int key_count = 0;
+  const int num_keys_per_level = 100;
+  // Level 0 : Keys in range [0, 99], Level 1:[100, 199], Level 2:[200, 299].
+  for (int level = 2; level >= 0; level--) {
+    key_count = level * num_keys_per_level;
+    for (int i = 0; i < num_keys_per_level; ++i) {
+      ASSERT_OK(Put(Key(key_count++), rnd.RandomString(500)));
+    }
+    ASSERT_OK(Flush());
+    MoveFilesToLevel(level);
+  }
+  Close();
+  std::vector<int> buff_prefectch_level_count = {0, 0, 0};
+  TryReopen(options);
+  {
+    auto iter = std::unique_ptr<Iterator>(db_->NewIterator(ReadOptions()));
+    fs->ClearPrefetchCount();
+    buff_prefetch_count = 0;
+
+    for (int level = 2; level >= 0; level--) {
+      key_count = level * num_keys_per_level;
+      switch (level) {
+        case 0:
+          // max_auto_readahead_size is set 0 so data and index blocks are not
+          // prefetched.
+          ASSERT_OK(db_->SetOptions(
+              {{"block_based_table_factory", "{max_auto_readahead_size=0;}"}}));
+          break;
+        case 1:
+          // max_auto_readahead_size is set less than
+          // BlockBasedTable::kInitAutoReadaheadSize. So readahead_size remains
+          // equal to max_auto_readahead_size.
+          ASSERT_OK(db_->SetOptions({{"block_based_table_factory",
+                                      "{max_auto_readahead_size=4096;}"}}));
+          break;
+        case 2:
+          ASSERT_OK(db_->SetOptions({{"block_based_table_factory",
+                                      "{max_auto_readahead_size=65536;}"}}));
+          break;
+        default:
+          assert(false);
+      }
+
+      for (int i = 0; i < num_keys_per_level; ++i) {
+        iter->Seek(Key(key_count++));
+        iter->Next();
+      }
+
+      buff_prefectch_level_count[level] = buff_prefetch_count;
+      if (support_prefetch && !use_direct_io) {
+        if (level == 0) {
+          ASSERT_FALSE(fs->IsPrefetchCalled());
+        } else {
+          ASSERT_TRUE(fs->IsPrefetchCalled());
+        }
+        fs->ClearPrefetchCount();
+      } else {
+        ASSERT_FALSE(fs->IsPrefetchCalled());
+        if (level == 0) {
+          ASSERT_EQ(buff_prefetch_count, 0);
+        } else {
+          ASSERT_GT(buff_prefetch_count, 0);
+        }
+        buff_prefetch_count = 0;
+      }
+    }
+  }
+
+  if (!support_prefetch) {
+    ASSERT_GT(buff_prefectch_level_count[1], buff_prefectch_level_count[2]);
+  }
+
+  SyncPoint::GetInstance()->DisableProcessing();
+  SyncPoint::GetInstance()->ClearAllCallBacks();
+  Close();
+}
+
 INSTANTIATE_TEST_CASE_P(PrefetchTest, PrefetchTest,
                         ::testing::Combine(::testing::Bool(),
                                            ::testing::Bool()));
+#endif  // !ROCKSDB_LITE
+
+class PrefetchTest1 : public DBTestBase,
+                      public ::testing::WithParamInterface<bool> {
+ public:
+  PrefetchTest1() : DBTestBase("/prefetch_test1", true) {}
+};
 
 }  // namespace ROCKSDB_NAMESPACE
 

include/rocksdb/table.h

@@ -435,6 +435,33 @@ struct BlockBasedTableOptions {
 
   IndexShorteningMode index_shortening =
       IndexShorteningMode::kShortenSeparators;
+
+  // RocksDB does auto-readahead for iterators on noticing more than two reads
+  // for a table file if user doesn't provide readahead_size. The readahead
+  // starts at 8KB and doubles on every additional read upto
+  // max_auto_readahead_size and max_auto_readahead_size can be configured.
+  //
+  // Special Value: 0 - If max_auto_readahead_size is set 0 then no implicit
+  // auto prefetching will be done. If max_auto_readahead_size provided is less
+  // than 8KB (which is initial readahead size used by rocksdb in case of
+  // auto-readahead), readahead size will remain same as
+  // max_auto_readahead_size.
+  //
+  // Value should be provided along with KB i.e. 256 * 1024 as it will prefetch
+  // the blocks.
+  //
+  // Found that 256 KB readahead size provides the best performance, based on
+  // experiments, for auto readahead. Experiment data is in PR #3282.
+  //
+  // This parameter can be changed dynamically by
+  // DB::SetOptions({{"block_based_table_factory",
+  //                  "{max_auto_readahead_size=0;}"}}));
+  //
+  // Changing the value dynamically will only affect files opened after the
+  // change.
+  //
+  // Default: 256 KB (256 * 1024).
+  size_t max_auto_readahead_size = 256 * 1024;
 };
 
 // Table Properties that are specific to block-based table properties.

options/options_settable_test.cc

@@ -179,7 +179,8 @@ TEST_F(OptionsSettableTest, BlockBasedTableOptionsAllFieldsSettable) {
       "hash_index_allow_collision=false;"
       "verify_compression=true;read_amp_bytes_per_bit=0;"
       "enable_index_compression=false;"
-      "block_align=true",
+      "block_align=true;"
+      "max_auto_readahead_size=0",
       new_bbto));
 
   ASSERT_EQ(unset_bytes_base,

table/block_based/block_based_table_factory.cc

@@ -415,6 +415,10 @@ static std::unordered_map<std::string, OptionTypeInfo>
             auto* cache = reinterpret_cast<std::shared_ptr<Cache>*>(addr);
             return Cache::CreateFromString(opts, value, cache);
           }}},
+        {"max_auto_readahead_size",
+         {offsetof(struct BlockBasedTableOptions, max_auto_readahead_size),
+          OptionType::kSizeT, OptionVerificationType::kNormal,
+          OptionTypeFlags::kMutable}},
 #endif  // ROCKSDB_LITE
 };
 
@@ -687,6 +691,9 @@ std::string BlockBasedTableFactory::GetPrintableOptions() const {
   snprintf(buffer, kBufferSize, "  block_align: %d\n",
            table_options_.block_align);
   ret.append(buffer);
+  snprintf(buffer, kBufferSize,
+           "  max_auto_readahead_size: %" ROCKSDB_PRIszt "\n",
+           table_options_.max_auto_readahead_size);
   return ret;
 }
 

table/block_based/block_based_table_reader.cc

@@ -67,11 +67,6 @@ extern const uint64_t kBlockBasedTableMagicNumber;
 extern const std::string kHashIndexPrefixesBlock;
 extern const std::string kHashIndexPrefixesMetadataBlock;
 
-
-// Found that 256 KB readahead size provides the best performance, based on
-// experiments, for auto readahead. Experiment data is in PR #3282.
-const size_t BlockBasedTable::kMaxAutoReadaheadSize = 256 * 1024;
-
 BlockBasedTable::~BlockBasedTable() {
   delete rep_;
 }
@@ -2921,7 +2916,7 @@ Status BlockBasedTable::VerifyChecksumInBlocks(
   // increasing of the buffer size.
   size_t readahead_size = (read_options.readahead_size != 0)
                               ? read_options.readahead_size
-                              : kMaxAutoReadaheadSize;
+                              : rep_->table_options.max_auto_readahead_size;
   // FilePrefetchBuffer doesn't work in mmap mode and readahead is not
   // needed there.
   FilePrefetchBuffer prefetch_buffer(

table/block_based/block_based_table_reader.h

@@ -65,9 +65,6 @@ class BlockBasedTable : public TableReader {
 
   // All the below fields control iterator readahead
   static const size_t kInitAutoReadaheadSize = 8 * 1024;
-  // Found that 256 KB readahead size provides the best performance, based on
-  // experiments, for auto readahead. Experiment data is in PR #3282.
-  static const size_t kMaxAutoReadaheadSize;
   static const int kMinNumFileReadsToStartAutoReadahead = 2;
 
   // Attempt to open the table that is stored in bytes [0..file_size)

table/block_based/block_prefetcher.cc

@@ -35,10 +35,23 @@ void BlockPrefetcher::PrefetchIfNeeded(const BlockBasedTable::Rep* rep,
     return;
   }
 
+  size_t max_auto_readahead_size = rep->table_options.max_auto_readahead_size;
+  size_t initial_auto_readahead_size = BlockBasedTable::kInitAutoReadaheadSize;
+
+  // If max_auto_readahead_size is set to be 0 by user, no data will be
+  // prefetched.
+  if (max_auto_readahead_size == 0) {
+    return;
+  }
+
+  if (initial_auto_readahead_size > max_auto_readahead_size) {
+    initial_auto_readahead_size = max_auto_readahead_size;
+  }
+
   if (rep->file->use_direct_io()) {
-    rep->CreateFilePrefetchBufferIfNotExists(
-        BlockBasedTable::kInitAutoReadaheadSize,
-        BlockBasedTable::kMaxAutoReadaheadSize, &prefetch_buffer_);
+    rep->CreateFilePrefetchBufferIfNotExists(initial_auto_readahead_size,
+                                             max_auto_readahead_size,
+                                             &prefetch_buffer_);
     return;
   }
 
@@ -47,20 +60,24 @@ void BlockPrefetcher::PrefetchIfNeeded(const BlockBasedTable::Rep* rep,
     return;
   }
 
+  if (readahead_size_ > max_auto_readahead_size) {
+    readahead_size_ = max_auto_readahead_size;
+  }
+
   // If prefetch is not supported, fall back to use internal prefetch buffer.
   // Discarding other return status of Prefetch calls intentionally, as
   // we can fallback to reading from disk if Prefetch fails.
   Status s = rep->file->Prefetch(handle.offset(), readahead_size_);
   if (s.IsNotSupported()) {
-    rep->CreateFilePrefetchBufferIfNotExists(
-        BlockBasedTable::kInitAutoReadaheadSize,
-        BlockBasedTable::kMaxAutoReadaheadSize, &prefetch_buffer_);
+    rep->CreateFilePrefetchBufferIfNotExists(initial_auto_readahead_size,
+                                             max_auto_readahead_size,
+                                             &prefetch_buffer_);
     return;
   }
   readahead_limit_ = static_cast<size_t>(handle.offset() + readahead_size_);
+
   // Keep exponentially increasing readahead size until
-  // kMaxAutoReadaheadSize.
-  readahead_size_ =
-      std::min(BlockBasedTable::kMaxAutoReadaheadSize, readahead_size_ * 2);
+  // max_auto_readahead_size.
+  readahead_size_ = std::min(max_auto_readahead_size, readahead_size_ * 2);
 }
 }  // namespace ROCKSDB_NAMESPACE