Merge branch 'master' of github.com:facebook/rocksdb

CMakeLists.txt

@@ -58,7 +58,7 @@ add_custom_command(OUTPUT ${BUILD_VERSION_CC}
 add_custom_target(GenerateBuildVersion DEPENDS ${BUILD_VERSION_CC})
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /Zi /nologo  /EHsc /GS /Gd /GR /GF /fp:precise /Zc:wchar_t /Zc:forScope /errorReport:queue")
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /FC /d2Zi+ /W3 /WX /wd4100 /wd4101 /wd4127 /wd4189 /wd4200 /wd4244 /wd4267 /wd4296 /wd4305 /wd4307 /wd4309 /wd4512 /wd4701 /wd4702 /wd4800 /wd4804 /wd4996")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /FC /d2Zi+ /W3 /WX /wd4127 /wd4200 /wd4244 /wd4267 /wd4305 /wd4307 /wd4309 /wd4512 /wd4701 /wd4702 /wd4800 /wd4804 /wd4996")
 
 # Used to run CI build and tests so we can run faster
 set(OPTIMIZE_DEBUG_DEFAULT 0)        # Debug build is unoptimized by default use -DOPTDBG=1 to optimize

db/c.cc

@@ -1288,6 +1288,11 @@ void rocksdb_block_based_options_set_cache_index_and_filter_blocks(
   options->rep.cache_index_and_filter_blocks = v;
 }
 
+void rocksdb_block_based_options_set_skip_table_builder_flush(
+    rocksdb_block_based_table_options_t* options, unsigned char v) {
+  options->rep.skip_table_builder_flush = v;
+}
+
 void rocksdb_options_set_block_based_table_factory(
     rocksdb_options_t *opt,
     rocksdb_block_based_table_options_t* table_options) {

db/db_bench.cc

@@ -376,6 +376,12 @@ DEFINE_int32(compaction_readahead_size, 0, "Compaction readahead size");
 DEFINE_int32(random_access_max_buffer_size, 1024 * 1024,
              "Maximum windows randomaccess buffer size");
 
+DEFINE_int32(writable_file_max_buffer_size, 1024 * 1024, 
+             "Maximum write buffer for Writeable File");
+
+DEFINE_int32(skip_table_builder_flush, false, "Skip flushing block in "
+             "table builder ");
+
 DEFINE_int32(bloom_bits, -1, "Bloom filter bits per key. Negative means"
              " use default settings.");
 DEFINE_int32(memtable_bloom_bits, 0, "Bloom filter bits per key for memtable. "
@@ -2299,6 +2305,7 @@ class Benchmark {
         FLAGS_new_table_reader_for_compaction_inputs;
     options.compaction_readahead_size = FLAGS_compaction_readahead_size;
     options.random_access_max_buffer_size = FLAGS_random_access_max_buffer_size;
+    options.writable_file_max_buffer_size = FLAGS_writable_file_max_buffer_size;
     options.statistics = dbstats;
     if (FLAGS_enable_io_prio) {
       FLAGS_env->LowerThreadPoolIOPriority(Env::LOW);
@@ -2441,6 +2448,8 @@ class Benchmark {
       block_based_options.block_size = FLAGS_block_size;
       block_based_options.block_restart_interval = FLAGS_block_restart_interval;
       block_based_options.filter_policy = filter_policy_;
+      block_based_options.skip_table_builder_flush = 
+          FLAGS_skip_table_builder_flush;
       block_based_options.format_version = 2;
       options.table_factory.reset(
           NewBlockBasedTableFactory(block_based_options));

db/memtablerep_bench.cc

@@ -132,8 +132,6 @@ DEFINE_int64(seed, 0,
              "Seed base for random number generators. "
              "When 0 it is deterministic.");
 
-static rocksdb::Env* FLAGS_env = rocksdb::Env::Default();
-
 namespace rocksdb {
 
 namespace {

include/rocksdb/c.h

@@ -450,6 +450,9 @@ rocksdb_block_based_options_set_hash_index_allow_collision(
 extern ROCKSDB_LIBRARY_API void
 rocksdb_block_based_options_set_cache_index_and_filter_blocks(
     rocksdb_block_based_table_options_t*, unsigned char);
+extern ROCKSDB_LIBRARY_API void
+rocksdb_block_based_options_set_skip_table_builder_flush(
+    rocksdb_block_based_table_options_t* options, unsigned char);
 extern ROCKSDB_LIBRARY_API void rocksdb_options_set_block_based_table_factory(
     rocksdb_options_t* opt, rocksdb_block_based_table_options_t* table_options);
 

include/rocksdb/env.h

@@ -94,6 +94,9 @@ struct EnvOptions {
   // See DBOPtions doc
   size_t random_access_max_buffer_size;
 
+  // See DBOptions doc
+  size_t writable_file_max_buffer_size = 1024 * 1024;
+
   // If not nullptr, write rate limiting is enabled for flush and compaction
   RateLimiter* rate_limiter = nullptr;
 };

include/rocksdb/options.h

@@ -1089,6 +1089,14 @@ struct DBOptions {
   // Default: 1 Mb
   size_t random_access_max_buffer_size;
 
+  // This is the maximum buffer size that is used by WritableFileWriter.
+  // On Windows, we need to maintain an aligned buffer for writes. 
+  // We allow the buffer to grow until it's size hits the limit. 
+  //
+  // Default: 1024 * 1024 (1 MB)
+  size_t writable_file_max_buffer_size;
+
+
   // Use adaptive mutex, which spins in the user space before resorting
   // to kernel. This could reduce context switch when the mutex is not
   // heavily contended. However, if the mutex is hot, we could end up

include/rocksdb/table.h

@@ -128,6 +128,20 @@ struct BlockBasedTableOptions {
   // This must generally be true for gets to be efficient.
   bool whole_key_filtering = true;
 
+  // If true, block will not be explictly flushed to disk during building
+  // a SstTable. Instead, buffer in WritableFileWriter will take 
+  // care of the flushing when it is full. 
+  //
+  // On Windows, this option helps a lot when unbuffered I/O 
+  // (allow_os_buffer = false) is used, since it avoids small 
+  // unbuffered disk write. 
+  //
+  // User may also adjust writable_file_max_buffer_size to optimize disk I/O 
+  // size. 
+  //
+  // Default: false
+  bool skip_table_builder_flush = false;
+  
   // We currently have three versions:
   // 0 -- This version is currently written out by all RocksDB's versions by
   // default.  Can be read by really old RocksDB's. Doesn't support changing

table/block_based_table_builder.cc

@@ -592,7 +592,7 @@ void BlockBasedTableBuilder::Flush() {
   if (!ok()) return;
   if (r->data_block.empty()) return;
   WriteBlock(&r->data_block, &r->pending_handle);
-  if (ok()) {
+  if (ok() && !r->table_options.skip_table_builder_flush) {
     r->status = r->file->Flush();
   }
   if (r->filter_block != nullptr) {

table/block_based_table_factory.cc

@@ -152,6 +152,10 @@ std::string BlockBasedTableFactory::GetPrintableTableOptions() const {
   ret.append(buffer);
   snprintf(buffer, kBufferSize, "  whole_key_filtering: %d\n",
            table_options_.whole_key_filtering);
+  ret.append(buffer);
+  snprintf(buffer, kBufferSize, "  skip_table_builder_flush: %d\n",
+           table_options_.skip_table_builder_flush);
+  ret.append(buffer);
   snprintf(buffer, kBufferSize, "  format_version: %d\n",
            table_options_.format_version);
   ret.append(buffer);

util/env.cc

@@ -296,6 +296,8 @@ void AssignEnvOptions(EnvOptions* env_options, const DBOptions& options) {
   env_options->random_access_max_buffer_size =
       options.random_access_max_buffer_size;
   env_options->rate_limiter = options.rate_limiter.get();
+  env_options->writable_file_max_buffer_size = 
+                  options.writable_file_max_buffer_size;
   env_options->allow_fallocate = options.allow_fallocate;
 }
 

util/file_reader_writer.cc

@@ -21,10 +21,6 @@
 
 namespace rocksdb {
 
-namespace {
-  const size_t c_OneMb = (1 << 20);
-}
-
 Status SequentialFileReader::Read(size_t n, Slice* result, char* scratch) {
   Status s = file_->Read(n, result, scratch);
   IOSTATS_ADD(bytes_read, result->size());
@@ -76,9 +72,9 @@ Status WritableFileWriter::Append(const Slice& data) {
       }
     }
 
-    if (buf_.Capacity() < c_OneMb) {
+    if (buf_.Capacity() < max_buffer_size_) {
       size_t desiredCapacity = buf_.Capacity() * 2;
-      desiredCapacity = std::min(desiredCapacity, c_OneMb);
+      desiredCapacity = std::min(desiredCapacity, max_buffer_size_);
       buf_.AllocateNewBuffer(desiredCapacity);
     }
     assert(buf_.CurrentSize() == 0);
@@ -102,9 +98,9 @@ Status WritableFileWriter::Append(const Slice& data) {
         // We double the buffer here because
         // Flush calls do not keep up with the incoming bytes
         // This is the only place when buffer is changed with unbuffered I/O
-        if (buf_.Capacity() < c_OneMb) {
+        if (buf_.Capacity() < max_buffer_size_) {
           size_t desiredCapacity = buf_.Capacity() * 2;
-          desiredCapacity = std::min(desiredCapacity, c_OneMb);
+          desiredCapacity = std::min(desiredCapacity, max_buffer_size_);
           buf_.AllocateNewBuffer(desiredCapacity);
         }
       }
@@ -156,7 +152,6 @@ Status WritableFileWriter::Close() {
   return s;
 }
 
-
 // write out the cached data to the OS cache
 Status WritableFileWriter::Flush() {
   Status s;

util/file_reader_writer.h

@@ -93,6 +93,7 @@ class WritableFileWriter {
  private:
   std::unique_ptr<WritableFile> writable_file_;
   AlignedBuffer           buf_;
+  size_t                  max_buffer_size_;
   // Actually written data size can be used for truncate
   // not counting padding data
   uint64_t                filesize_;
@@ -113,6 +114,7 @@ class WritableFileWriter {
                      const EnvOptions& options)
       : writable_file_(std::move(file)),
         buf_(),
+        max_buffer_size_(options.writable_file_max_buffer_size),
         filesize_(0),
         next_write_offset_(0),
         pending_sync_(false),

util/options.cc

@@ -251,6 +251,7 @@ DBOptions::DBOptions()
       new_table_reader_for_compaction_inputs(false),
       compaction_readahead_size(0),
       random_access_max_buffer_size(1024 * 1024),
+      writable_file_max_buffer_size(1024 * 1024),
       use_adaptive_mutex(false),
       bytes_per_sync(0),
       wal_bytes_per_sync(0),
@@ -313,6 +314,7 @@ DBOptions::DBOptions(const Options& options)
           options.new_table_reader_for_compaction_inputs),
       compaction_readahead_size(options.compaction_readahead_size),
       random_access_max_buffer_size(options.random_access_max_buffer_size),
+      writable_file_max_buffer_size(options.writable_file_max_buffer_size),
       use_adaptive_mutex(options.use_adaptive_mutex),
       bytes_per_sync(options.bytes_per_sync),
       wal_bytes_per_sync(options.wal_bytes_per_sync),
@@ -412,6 +414,10 @@ void DBOptions::Dump(Logger* log) const {
         "               Options.random_access_max_buffer_size: %" ROCKSDB_PRIszt
         "d",
         random_access_max_buffer_size);
+    Header(log,
+         "              Options.writable_file_max_buffer_size: %" ROCKSDB_PRIszt
+         "d",
+         writable_file_max_buffer_size);
     Header(log, "                      Options.use_adaptive_mutex: %d",
         use_adaptive_mutex);
     Header(log, "                            Options.rate_limiter: %p",

util/options_helper.cc

@@ -891,7 +891,7 @@ bool ParseColumnFamilyOption(const std::string& name,
           reinterpret_cast<char*>(new_options) + opt_info.offset, opt_info.type,
           value);
     }
-  } catch (std::exception& e) {
+  } catch (const std::exception&) {
     return false;
   }
   return true;
@@ -1058,7 +1058,7 @@ bool ParseDBOption(const std::string& name, const std::string& org_value,
             opt_info.type, value);
       }
     }
-  } catch (const std::exception& e) {
+  } catch (const std::exception&) {
     return false;
   }
   return true;

util/options_helper.h

@@ -184,6 +184,9 @@ static std::unordered_map<std::string, OptionTypeInfo> db_options_type_info = {
     {"random_access_max_buffer_size",
      {offsetof(struct DBOptions, random_access_max_buffer_size),
       OptionType::kSizeT, OptionVerificationType::kNormal}},
+    {"writable_file_max_buffer_size",
+     {offsetof(struct DBOptions, writable_file_max_buffer_size),  
+      OptionType::kSizeT, OptionVerificationType::kNormal}},
     {"use_adaptive_mutex",
      {offsetof(struct DBOptions, use_adaptive_mutex), OptionType::kBoolean,
       OptionVerificationType::kNormal}},
@@ -460,6 +463,9 @@ static std::unordered_map<std::string,
     {"whole_key_filtering",
      {offsetof(struct BlockBasedTableOptions, whole_key_filtering),
       OptionType::kBoolean, OptionVerificationType::kNormal}},
+    {"skip_table_builder_flush",
+     {offsetof(struct BlockBasedTableOptions, skip_table_builder_flush),  
+      OptionType::kBoolean, OptionVerificationType::kNormal}},
     {"format_version",
      {offsetof(struct BlockBasedTableOptions, format_version),
       OptionType::kUInt32T, OptionVerificationType::kNormal}}};

util/options_test.cc

@@ -341,6 +341,7 @@ TEST_F(OptionsTest, GetOptionsFromMapTest) {
       {"new_table_reader_for_compaction_inputs", "true"},
       {"compaction_readahead_size", "100"},
       {"random_access_max_buffer_size", "3145728"},
+      {"writable_file_max_buffer_size", "314159"},
       {"bytes_per_sync", "47"},
       {"wal_bytes_per_sync", "48"},
   };
@@ -452,6 +453,7 @@ TEST_F(OptionsTest, GetOptionsFromMapTest) {
   ASSERT_EQ(new_db_opt.new_table_reader_for_compaction_inputs, true);
   ASSERT_EQ(new_db_opt.compaction_readahead_size, 100);
   ASSERT_EQ(new_db_opt.random_access_max_buffer_size, 3145728);
+  ASSERT_EQ(new_db_opt.writable_file_max_buffer_size, 314159);
   ASSERT_EQ(new_db_opt.bytes_per_sync, static_cast<uint64_t>(47));
   ASSERT_EQ(new_db_opt.wal_bytes_per_sync, static_cast<uint64_t>(48));
 }
@@ -621,7 +623,8 @@ TEST_F(OptionsTest, GetBlockBasedTableOptionsFromString) {
             "checksum=kxxHash;hash_index_allow_collision=1;no_block_cache=1;"
             "block_cache=1M;block_cache_compressed=1k;block_size=1024;"
             "block_size_deviation=8;block_restart_interval=4;"
-            "filter_policy=bloomfilter:4:true;whole_key_filtering=1",
+            "filter_policy=bloomfilter:4:true;whole_key_filtering=1;"
+            "skip_table_builder_flush=1",
             &new_opt));
   ASSERT_TRUE(new_opt.cache_index_and_filter_blocks);
   ASSERT_EQ(new_opt.index_type, BlockBasedTableOptions::kHashSearch);
@@ -636,6 +639,7 @@ TEST_F(OptionsTest, GetBlockBasedTableOptionsFromString) {
   ASSERT_EQ(new_opt.block_size_deviation, 8);
   ASSERT_EQ(new_opt.block_restart_interval, 4);
   ASSERT_TRUE(new_opt.filter_policy != nullptr);
+  ASSERT_TRUE(new_opt.skip_table_builder_flush);
 
   // unknown option
   ASSERT_NOK(GetBlockBasedTableOptionsFromString(table_opt,