Add a utility function to guess optimized options based on constraints

Summary:
Add a function GetOptions(), where based on four parameters users give: read/write amplification threshold, memory budget for mem tables and target DB size, it picks up a compaction style and parameters for them. Background threads are not touched yet.

One limit of this algorithm: since compression rate and key/value size are hard to predict, it's hard to predict level 0 file size from write buffer size. Simply make 1:1 ratio here.

Sample results: https://reviews.facebook.net/P477

Test Plan: Will add some a unit test where some sample scenarios are given and see they pick the results that make sense

Reviewers: yhchiang, dhruba, haobo, igor, ljin

Reviewed By: ljin

Subscribers: leveldb

Differential Revision: https://reviews.facebook.net/D18741

Makefile

@@ -114,9 +114,10 @@ TESTS = \
 	deletefile_test \
 	table_test \
 	thread_local_test \
-        geodb_test \
-        rate_limiter_test \
-	cuckoo_table_builder_test
+	geodb_test \
+  rate_limiter_test \
+	cuckoo_table_builder_test \
+	options_test
 
 TOOLS = \
         sst_dump \
@@ -124,6 +125,7 @@ TOOLS = \
         db_stress \
         ldb \
 	db_repl_stress \
+  options_test \
 	blob_store_bench
 
 PROGRAMS = db_bench signal_test table_reader_bench log_and_apply_bench $(TOOLS)
@@ -414,6 +416,9 @@ geodb_test: utilities/geodb/geodb_test.o $(LIBOBJECTS) $(TESTHARNESS)
 cuckoo_table_builder_test: table/cuckoo_table_builder_test.o $(LIBOBJECTS) $(TESTHARNESS)
 	$(CXX) table/cuckoo_table_builder_test.o  $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)
 
+options_test: util/options_test.o $(LIBOBJECTS) $(TESTHARNESS)
+	$(CXX) util/options_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)
+
 $(MEMENVLIBRARY) : $(MEMENVOBJECTS)
 	rm -f $@
 	$(AR) -rs $@ $(MEMENVOBJECTS)

include/rocksdb/options.h

@@ -1027,7 +1027,6 @@ struct FlushOptions {
   FlushOptions() : wait(true) {}
 };
 
-
 // Create a RateLimiter object, which can be shared among RocksDB instances to
 // control write rate of flush and compaction.
 // @rate_bytes_per_sec: this is the only parameter you want to set most of the
@@ -1051,7 +1050,16 @@ extern RateLimiter* NewRateLimiter(
     int64_t refill_period_us = 100 * 1000,
     int32_t fairness = 10);
 
-
+// Get options based on some guidelines. Now only tune parameter based on
+// flush/compaction and fill default parameters for other parameters.
+// total_write_buffer_limit: budget for memory spent for mem tables
+// read_amplification_threshold: comfortable value of read amplification
+// write_amplification_threshold: comfortable value of write amplification.
+// target_db_size: estimated total DB size.
+extern Options GetOptions(size_t total_write_buffer_limit,
+                          int read_amplification_threshold = 8,
+                          int write_amplification_threshold = 32,
+                          uint64_t target_db_size = 68719476736 /* 64GB */);
 }  // namespace rocksdb
 
 #endif  // STORAGE_ROCKSDB_INCLUDE_OPTIONS_H_

util/options_builder.cc

@@ -0,0 +1,196 @@
+//  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.
+
+#include <math.h>
+#include <algorithm>
+#include "rocksdb/options.h"
+
+namespace rocksdb {
+
+namespace {
+
+// For now, always use 1-0 as level bytes multiplier.
+const int kBytesForLevelMultiplier = 10;
+const size_t kBytesForOneMb = 1024 * 1024;
+
+// Pick compaction style
+CompactionStyle PickCompactionStyle(size_t write_buffer_size,
+                                    int read_amp_threshold,
+                                    int write_amp_threshold,
+                                    uint64_t target_db_size) {
+  // Estimate read amplification and write amplification of two compaction
+  // styles. If there is hard limit to force a choice, make the choice.
+  // Otherwise, calculate a score based on threshold and expected value of
+  // two styles, weighing reads 4X important than writes.
+  int expected_levels = static_cast<int>(ceil(
+      log(target_db_size / write_buffer_size) / log(kBytesForLevelMultiplier)));
+
+  int expected_max_files_universal =
+      static_cast<int>(ceil(log2(target_db_size / write_buffer_size)));
+
+  const int kEstimatedLevel0FilesInLevelStyle = 2;
+  // Estimate write amplification:
+  // (1) 1 for every L0 file
+  // (2) 2 for L1
+  // (3) kBytesForLevelMultiplier for the last level. It's really hard to
+  //     predict.
+  // (3) kBytesForLevelMultiplier for other levels.
+  int expected_write_amp_level = kEstimatedLevel0FilesInLevelStyle + 2
+      + (expected_levels - 2) * kBytesForLevelMultiplier
+      + kBytesForLevelMultiplier;
+  int expected_read_amp_level =
+      kEstimatedLevel0FilesInLevelStyle + expected_levels;
+
+  int max_read_amp_uni = expected_max_files_universal;
+  if (read_amp_threshold <= max_read_amp_uni) {
+    return kCompactionStyleLevel;
+  } else if (write_amp_threshold <= expected_write_amp_level) {
+    return kCompactionStyleUniversal;
+  }
+
+  const double kReadWriteWeight = 4;
+
+  double level_ratio =
+      static_cast<double>(read_amp_threshold) / expected_read_amp_level *
+          kReadWriteWeight +
+      static_cast<double>(write_amp_threshold) / expected_write_amp_level;
+
+  int expected_write_amp_uni = expected_max_files_universal / 2 + 2;
+  int expected_read_amp_uni = expected_max_files_universal / 2 + 1;
+
+  double uni_ratio =
+      static_cast<double>(read_amp_threshold) / expected_read_amp_uni *
+          kReadWriteWeight +
+      static_cast<double>(write_amp_threshold) / expected_write_amp_uni;
+
+  if (level_ratio > uni_ratio) {
+    return kCompactionStyleLevel;
+  } else {
+    return kCompactionStyleUniversal;
+  }
+}
+
+// Pick mem table size
+void PickWriteBufferSize(size_t total_write_buffer_limit, Options* options) {
+  const size_t kMaxWriteBufferSize = 128 * kBytesForOneMb;
+  const size_t kMinWriteBufferSize = 4 * kBytesForOneMb;
+
+  // Try to pick up a buffer size between 4MB and 128MB.
+  // And try to pick 4 as the total number of write buffers.
+  size_t write_buffer_size = total_write_buffer_limit / 4;
+  if (write_buffer_size > kMaxWriteBufferSize) {
+    write_buffer_size = kMaxWriteBufferSize;
+  } else if (write_buffer_size < kMinWriteBufferSize) {
+    write_buffer_size = std::min(static_cast<size_t>(kMinWriteBufferSize),
+                                 total_write_buffer_limit / 2);
+  }
+
+  // Truncate to multiple of 1MB.
+  if (write_buffer_size % kBytesForOneMb != 0) {
+    write_buffer_size =
+        (write_buffer_size / kBytesForOneMb + 1) * kBytesForOneMb;
+  }
+
+  options->write_buffer_size = write_buffer_size;
+  options->max_write_buffer_number =
+      total_write_buffer_limit / write_buffer_size;
+  options->min_write_buffer_number_to_merge = 1;
+}
+
+void OptimizeForUniversal(Options* options) {
+  options->level0_file_num_compaction_trigger = 2;
+  options->level0_slowdown_writes_trigger = 30;
+  options->level0_stop_writes_trigger = 40;
+  options->max_open_files = -1;
+}
+
+// Optimize parameters for level-based compaction
+void OptimizeForLevel(int read_amplification_threshold,
+                      int write_amplification_threshold,
+                      uint64_t target_db_size, Options* options) {
+  int expected_levels_one_level0_file =
+      static_cast<int>(ceil(log(target_db_size / options->write_buffer_size) /
+                            log(kBytesForLevelMultiplier)));
+
+  int level0_stop_writes_trigger =
+      read_amplification_threshold - expected_levels_one_level0_file;
+
+  const size_t kInitialLevel0TotalSize = 128 * kBytesForOneMb;
+  const int kMaxFileNumCompactionTrigger = 4;
+  const int kMinLevel0StopTrigger = 3;
+
+  int file_num_buffer =
+      kInitialLevel0TotalSize / options->write_buffer_size + 1;
+
+  if (level0_stop_writes_trigger > file_num_buffer) {
+    // Have sufficient room for multiple level 0 files
+    // Try enlarge the buffer up to 1GB
+
+    // Try to enlarge the buffer up to 1GB, if still have sufficient headroom.
+    file_num_buffer *=
+        std::pow(2, std::max(0, std::min(3, level0_stop_writes_trigger -
+                                                file_num_buffer - 2)));
+
+    options->level0_stop_writes_trigger = level0_stop_writes_trigger;
+    options->level0_slowdown_writes_trigger = level0_stop_writes_trigger - 2;
+    options->level0_file_num_compaction_trigger =
+        std::min(kMaxFileNumCompactionTrigger, file_num_buffer / 2);
+  } else {
+    options->level0_stop_writes_trigger =
+        std::max(kMinLevel0StopTrigger, file_num_buffer);
+    options->level0_slowdown_writes_trigger =
+        options->level0_stop_writes_trigger - 1;
+    options->level0_file_num_compaction_trigger = 1;
+  }
+
+  // This doesn't consider compaction and overheads of mem tables. But usually
+  // it is in the same order of magnitude.
+  int expected_level0_compaction_size =
+      options->level0_file_num_compaction_trigger * options->write_buffer_size;
+  // Enlarge level1 target file size if level0 compaction size is larger.
+  int max_bytes_for_level_base = 10 * kBytesForOneMb;
+  if (expected_level0_compaction_size > max_bytes_for_level_base) {
+    max_bytes_for_level_base = expected_level0_compaction_size;
+  }
+  options->max_bytes_for_level_base = max_bytes_for_level_base;
+  // Now always set level multiplier to be 10
+  options->max_bytes_for_level_multiplier = kBytesForLevelMultiplier;
+
+  const int kMinFileSize = 2 * kBytesForOneMb;
+  // Allow at least 3-way parallelism for compaction between level 1 and 2.
+  int max_file_size = max_bytes_for_level_base / 3;
+  if (max_file_size < kMinFileSize) {
+    options->target_file_size_base = kMinFileSize;
+  } else {
+    if (max_file_size % kBytesForOneMb != 0) {
+      max_file_size = (max_file_size / kBytesForOneMb + 1) * kBytesForOneMb;
+    }
+    options->target_file_size_base = max_file_size;
+  }
+
+  // TODO: consider to tune num_levels too.
+}
+
+}  // namespace
+
+Options GetOptions(size_t total_write_buffer_limit,
+                   int read_amplification_threshold,
+                   int write_amplification_threshold, uint64_t target_db_size) {
+  Options options;
+  PickWriteBufferSize(total_write_buffer_limit, &options);
+  size_t write_buffer_size = options.write_buffer_size;
+  options.compaction_style =
+      PickCompactionStyle(write_buffer_size, read_amplification_threshold,
+                          write_amplification_threshold, target_db_size);
+  if (options.compaction_style == kCompactionStyleUniversal) {
+    OptimizeForUniversal(&options);
+  } else {
+    OptimizeForLevel(read_amplification_threshold,
+                     write_amplification_threshold, target_db_size, &options);
+  }
+  return options;
+}
+
+}  // namespace rocksdb

util/options_test.cc

@@ -0,0 +1,80 @@
+//  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.
+
+#define __STDC_FORMAT_MACROS
+#include <inttypes.h>
+#include <gflags/gflags.h>
+
+#include "rocksdb/options.h"
+#include "util/testharness.h"
+
+using GFLAGS::ParseCommandLineFlags;
+DEFINE_bool(enable_print, false, "Print options generated to console.");
+
+namespace rocksdb {
+
+class OptionsTest {};
+
+class StderrLogger : public Logger {
+ public:
+  virtual void Logv(const char* format, va_list ap) override {
+    vprintf(format, ap);
+    printf("\n");
+  }
+};
+
+Options PrintAndGetOptions(size_t total_write_buffer_limit,
+                           int read_amplification_threshold,
+                           int write_amplification_threshold,
+                           uint64_t target_db_size = 68719476736) {
+  StderrLogger logger;
+
+  if (FLAGS_enable_print) {
+    printf(
+        "---- total_write_buffer_limit: %zu "
+        "read_amplification_threshold: %d write_amplification_threshold: %d "
+        "target_db_size %" PRIu64 " ----\n",
+        total_write_buffer_limit, read_amplification_threshold,
+        write_amplification_threshold, target_db_size);
+  }
+
+  Options options =
+      GetOptions(total_write_buffer_limit, read_amplification_threshold,
+                 write_amplification_threshold, target_db_size);
+  if (FLAGS_enable_print) {
+    options.Dump(&logger);
+    printf("-------------------------------------\n\n\n");
+  }
+  return options;
+}
+
+TEST(OptionsTest, LooseCondition) {
+  Options options;
+  PrintAndGetOptions(static_cast<size_t>(10) * 1024 * 1024 * 1024, 100, 100);
+
+  // Less mem table memory budget
+  PrintAndGetOptions(32 * 1024 * 1024, 100, 100);
+
+  // Tight read amplification
+  options = PrintAndGetOptions(128 * 1024 * 1024, 8, 100);
+  ASSERT_EQ(options.compaction_style, kCompactionStyleLevel);
+
+  // Tight write amplification
+  options = PrintAndGetOptions(128 * 1024 * 1024, 64, 10);
+  ASSERT_EQ(options.compaction_style, kCompactionStyleUniversal);
+
+  // Both tight amplifications
+  PrintAndGetOptions(128 * 1024 * 1024, 4, 8);
+}
+}  // namespace rocksdb
+
+int main(int argc, char** argv) {
+  ParseCommandLineFlags(&argc, &argv, true);
+  return rocksdb::test::RunAllTests();
+}