NextGraph
/
rocksdb
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Eliminate redundant cache lookup with range deletion

Browse Source 
Summary:
When we introduced range deletion block, TableCache::Get() and TableCache::NewIterator() each did two table cache lookups, one for range deletion block iterator and another for getting the table reader to which the Get()/NewIterator() is delegated. This extra cache lookup was very CPU-intensive (about 10% overhead in a read-heavy benchmark). We can avoid it by reusing the Cache::Handle created for range deletion block iterator to get the file reader.
Closes https://github.com/facebook/rocksdb/pull/1537

Differential Revision: D4201167

Pulled By: ajkr

fbshipit-source-id: d33ffd8
main
Andrew Kryczka 8 years ago committed by Facebook Github Bot
parent 182b940e70
commit 734e4acafb
3 changed files with 77 additions and 121 deletions
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Show Stats Download Patch File Download Diff File
  1. 21
      db/db_compaction_test.cc
  2. 167
      db/table_cache.cc
  3. 10
      db/table_cache.h

21
db/db_compaction_test.cc
Unescape View File

@ -294,8 +294,7 @@ TEST_F(DBCompactionTest, TestTableReaderForCompaction) {
num_new_table_reader = 0; num_new_table_reader = 0;
ASSERT_EQ(Key(k), Get(Key(k))); ASSERT_EQ(Key(k), Get(Key(k)));
// lookup iterator from table cache and no need to create a new one. // lookup iterator from table cache and no need to create a new one.
// a second table cache iterator is created for range tombstones ASSERT_EQ(num_table_cache_lookup, 1);
ASSERT_EQ(num_table_cache_lookup, 2);
ASSERT_EQ(num_new_table_reader, 0); ASSERT_EQ(num_new_table_reader, 0);
} }
} }
@ -306,9 +305,9 @@ TEST_F(DBCompactionTest, TestTableReaderForCompaction) {
dbfull()->TEST_WaitForCompact(); dbfull()->TEST_WaitForCompact();
// Preloading iterator issues one table cache lookup and creates // Preloading iterator issues one table cache lookup and creates
// a new table reader. One file is created for flush and one for compaction. // a new table reader. One file is created for flush and one for compaction.
// Compaction inputs make no table cache look-up for data iterators and one // Compaction inputs make no table cache look-up for data/range deletion
// look-up per compaction input file (three). // iterators
ASSERT_EQ(num_table_cache_lookup, 5); ASSERT_EQ(num_table_cache_lookup, 2);
// Create new iterator for: // Create new iterator for:
// (1) 1 for verifying flush results // (1) 1 for verifying flush results
// (2) 3 for compaction input files // (2) 3 for compaction input files
@ -318,8 +317,7 @@ TEST_F(DBCompactionTest, TestTableReaderForCompaction) {
num_table_cache_lookup = 0; num_table_cache_lookup = 0;
num_new_table_reader = 0; num_new_table_reader = 0;
ASSERT_EQ(Key(1), Get(Key(1))); ASSERT_EQ(Key(1), Get(Key(1)));
// a second table cache iterator is created for range tombstones ASSERT_EQ(num_table_cache_lookup, 1);
ASSERT_EQ(num_table_cache_lookup, 2);
ASSERT_EQ(num_new_table_reader, 0); ASSERT_EQ(num_new_table_reader, 0);
num_table_cache_lookup = 0; num_table_cache_lookup = 0;
@ -329,17 +327,16 @@ TEST_F(DBCompactionTest, TestTableReaderForCompaction) {
cro.target_level = 2; cro.target_level = 2;
cro.bottommost_level_compaction = BottommostLevelCompaction::kForce; cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
db_->CompactRange(cro, nullptr, nullptr); db_->CompactRange(cro, nullptr, nullptr);
// Only verifying compaction outputs issues two table cache lookup // Only verifying compaction outputs issues one table cache lookup
// (one for data block, one for range deletion block). // for both data block and range deletion block).
ASSERT_EQ(num_table_cache_lookup, 2); ASSERT_EQ(num_table_cache_lookup, 1);
// One for compaction input, one for verifying compaction results. // One for compaction input, one for verifying compaction results.
ASSERT_EQ(num_new_table_reader, 2); ASSERT_EQ(num_new_table_reader, 2);
num_table_cache_lookup = 0; num_table_cache_lookup = 0;
num_new_table_reader = 0; num_new_table_reader = 0;
ASSERT_EQ(Key(1), Get(Key(1))); ASSERT_EQ(Key(1), Get(Key(1)));
// a second table cache iterator is created for range tombstones ASSERT_EQ(num_table_cache_lookup, 1);
ASSERT_EQ(num_table_cache_lookup, 2);
ASSERT_EQ(num_new_table_reader, 0); ASSERT_EQ(num_new_table_reader, 0);
rocksdb::SyncPoint::GetInstance()->ClearAllCallBacks(); rocksdb::SyncPoint::GetInstance()->ClearAllCallBacks();

167
db/table_cache.cc
Unescape View File

@ -175,17 +175,6 @@ InternalIterator* TableCache::NewIterator(
PERF_TIMER_GUARD(new_table_iterator_nanos); PERF_TIMER_GUARD(new_table_iterator_nanos);
Status s; Status s;
if (range_del_agg != nullptr && !options.ignore_range_deletions) {
std::unique_ptr<InternalIterator> range_del_iter(NewRangeDeletionIterator(
options, icomparator, fd, file_read_hist, skip_filters, level));
if (range_del_iter != nullptr) {
s = range_del_iter->status();
}
if (s.ok()) {
s = range_del_agg->AddTombstones(std::move(range_del_iter));
}
}
bool create_new_table_reader = false; bool create_new_table_reader = false;
TableReader* table_reader = nullptr; TableReader* table_reader = nullptr;
Cache::Handle* handle = nullptr; Cache::Handle* handle = nullptr;
@ -226,14 +215,15 @@ InternalIterator* TableCache::NewIterator(
} }
} }
} }
InternalIterator* result = nullptr;
if (s.ok()) { if (s.ok()) {
InternalIterator* result = result = table_reader->NewIterator(options, arena, skip_filters);
table_reader->NewIterator(options, arena, skip_filters);
if (create_new_table_reader) { if (create_new_table_reader) {
assert(handle == nullptr); assert(handle == nullptr);
result->RegisterCleanup(&DeleteTableReader, table_reader, nullptr); result->RegisterCleanup(&DeleteTableReader, table_reader, nullptr);
} else if (handle != nullptr) { } else if (handle != nullptr) {
result->RegisterCleanup(&UnrefEntry, cache_, handle); result->RegisterCleanup(&UnrefEntry, cache_, handle);
handle = nullptr; // prevent from releasing below
} }
if (for_compaction) { if (for_compaction) {
@ -242,45 +232,26 @@ InternalIterator* TableCache::NewIterator(
if (table_reader_ptr != nullptr) { if (table_reader_ptr != nullptr) {
*table_reader_ptr = table_reader; *table_reader_ptr = table_reader;
} }
return result;
} }
if (handle != nullptr) { if (s.ok() && range_del_agg != nullptr && !options.ignore_range_deletions) {
ReleaseHandle(handle); std::unique_ptr<InternalIterator> range_del_iter(
} table_reader->NewRangeTombstoneIterator(options));
return NewErrorInternalIterator(s, arena); if (range_del_iter != nullptr) {
} s = range_del_iter->status();
}
InternalIterator* TableCache::NewRangeDeletionIterator(
const ReadOptions& options, const InternalKeyComparator& icmp,
const FileDescriptor& fd, HistogramImpl* file_read_hist, bool skip_filters,
int level) {
if (options.ignore_range_deletions) {
return nullptr;
}
Status s;
TableReader* table_reader = fd.table_reader;
Cache::Handle* cache_handle = nullptr;
if (table_reader == nullptr) {
s = FindTable(env_options_, icmp, fd, &cache_handle,
options.read_tier == kBlockCacheTier /* no_io */,
true /* record_read_stats */, file_read_hist, skip_filters,
level);
if (s.ok()) { if (s.ok()) {
table_reader = GetTableReaderFromHandle(cache_handle); s = range_del_agg->AddTombstones(std::move(range_del_iter));
} }
} }
if (s.ok()) {
auto* result = table_reader->NewRangeTombstoneIterator(options); if (handle != nullptr) {
if (cache_handle != nullptr) { ReleaseHandle(handle);
if (result == nullptr) {
ReleaseHandle(cache_handle);
} else {
result->RegisterCleanup(&UnrefEntry, cache_, cache_handle);
}
}
return result;
} }
return NewErrorInternalIterator(s); if (!s.ok()) {
assert(result == nullptr);
result = NewErrorInternalIterator(s, arena);
}
return result;
} }
Status TableCache::Get(const ReadOptions& options, Status TableCache::Get(const ReadOptions& options,
@ -288,67 +259,52 @@ Status TableCache::Get(const ReadOptions& options,
const FileDescriptor& fd, const Slice& k, const FileDescriptor& fd, const Slice& k,
GetContext* get_context, HistogramImpl* file_read_hist, GetContext* get_context, HistogramImpl* file_read_hist,
bool skip_filters, int level) { bool skip_filters, int level) {
Status s;
if (get_context->range_del_agg() != nullptr &&
!options.ignore_range_deletions) {
std::unique_ptr<InternalIterator> range_del_iter(NewRangeDeletionIterator(
options, internal_comparator, fd, file_read_hist, skip_filters, level));
if (range_del_iter != nullptr) {
s = range_del_iter->status();
}
if (s.ok()) {
s = get_context->range_del_agg()->AddTombstones(
std::move(range_del_iter));
}
}
TableReader* t = fd.table_reader;
Cache::Handle* handle = nullptr;
std::string* row_cache_entry = nullptr; std::string* row_cache_entry = nullptr;
bool done = false; bool done = false;
#ifndef ROCKSDB_LITE #ifndef ROCKSDB_LITE
IterKey row_cache_key; IterKey row_cache_key;
std::string row_cache_entry_buffer; std::string row_cache_entry_buffer;
if (s.ok()) { // Check row cache if enabled. Since row cache does not currently store
// Check row cache if enabled. Since row cache does not currently store // sequence numbers, we cannot use it if we need to fetch the sequence.
// sequence numbers, we cannot use it if we need to fetch the sequence. if (ioptions_.row_cache && !get_context->NeedToReadSequence()) {
if (ioptions_.row_cache && !get_context->NeedToReadSequence()) { uint64_t fd_number = fd.GetNumber();
uint64_t fd_number = fd.GetNumber(); auto user_key = ExtractUserKey(k);
auto user_key = ExtractUserKey(k); // We use the user key as cache key instead of the internal key,
// We use the user key as cache key instead of the internal key, // otherwise the whole cache would be invalidated every time the
// otherwise the whole cache would be invalidated every time the // sequence key increases. However, to support caching snapshot
// sequence key increases. However, to support caching snapshot // reads, we append the sequence number (incremented by 1 to
// reads, we append the sequence number (incremented by 1 to // distinguish from 0) only in this case.
// distinguish from 0) only in this case. uint64_t seq_no =
uint64_t seq_no = options.snapshot == nullptr ? 0 : 1 + GetInternalKeySeqno(k);
options.snapshot == nullptr ? 0 : 1 + GetInternalKeySeqno(k);
// Compute row cache key.
// Compute row cache key. row_cache_key.TrimAppend(row_cache_key.Size(), row_cache_id_.data(),
row_cache_key.TrimAppend(row_cache_key.Size(), row_cache_id_.data(), row_cache_id_.size());
row_cache_id_.size()); AppendVarint64(&row_cache_key, fd_number);
AppendVarint64(&row_cache_key, fd_number); AppendVarint64(&row_cache_key, seq_no);
AppendVarint64(&row_cache_key, seq_no); row_cache_key.TrimAppend(row_cache_key.Size(), user_key.data(),
row_cache_key.TrimAppend(row_cache_key.Size(), user_key.data(), user_key.size());
user_key.size());
if (auto row_handle =
if (auto row_handle = ioptions_.row_cache->Lookup(row_cache_key.GetKey())) {
ioptions_.row_cache->Lookup(row_cache_key.GetKey())) { auto found_row_cache_entry = static_cast<const std::string*>(
auto found_row_cache_entry = static_cast<const std::string*>( ioptions_.row_cache->Value(row_handle));
ioptions_.row_cache->Value(row_handle)); replayGetContextLog(*found_row_cache_entry, user_key, get_context);
replayGetContextLog(*found_row_cache_entry, user_key, get_context); ioptions_.row_cache->Release(row_handle);
ioptions_.row_cache->Release(row_handle); RecordTick(ioptions_.statistics, ROW_CACHE_HIT);
RecordTick(ioptions_.statistics, ROW_CACHE_HIT); done = true;
done = true; } else {
} else { // Not found, setting up the replay log.
// Not found, setting up the replay log. RecordTick(ioptions_.statistics, ROW_CACHE_MISS);
RecordTick(ioptions_.statistics, ROW_CACHE_MISS); row_cache_entry = &row_cache_entry_buffer;
row_cache_entry = &row_cache_entry_buffer;
}
} }
} }
#endif // ROCKSDB_LITE #endif // ROCKSDB_LITE
Status s;
TableReader* t = fd.table_reader;
Cache::Handle* handle = nullptr;
if (!done && s.ok()) { if (!done && s.ok()) {
if (!t) { if (t == nullptr) {
s = FindTable(env_options_, internal_comparator, fd, &handle, s = FindTable(env_options_, internal_comparator, fd, &handle,
options.read_tier == kBlockCacheTier /* no_io */, options.read_tier == kBlockCacheTier /* no_io */,
true /* record_read_stats */, file_read_hist, skip_filters, true /* record_read_stats */, file_read_hist, skip_filters,
@ -368,6 +324,19 @@ Status TableCache::Get(const ReadOptions& options,
done = true; done = true;
} }
} }
if (!done && s.ok() && get_context->range_del_agg() != nullptr &&
!options.ignore_range_deletions) {
std::unique_ptr<InternalIterator> range_del_iter(
t->NewRangeTombstoneIterator(options));
if (range_del_iter != nullptr) {
s = range_del_iter->status();
}
if (s.ok()) {
s = get_context->range_del_agg()->AddTombstones(
std::move(range_del_iter));
}
}
#ifndef ROCKSDB_LITE #ifndef ROCKSDB_LITE
// Put the replay log in row cache only if something was found. // Put the replay log in row cache only if something was found.
if (!done && s.ok() && row_cache_entry && !row_cache_entry->empty()) { if (!done && s.ok() && row_cache_entry && !row_cache_entry->empty()) {

10
db/table_cache.h
Unescape View File

@ -58,16 +58,6 @@ class TableCache {
HistogramImpl* file_read_hist = nullptr, bool for_compaction = false, HistogramImpl* file_read_hist = nullptr, bool for_compaction = false,
Arena* arena = nullptr, bool skip_filters = false, int level = -1); Arena* arena = nullptr, bool skip_filters = false, int level = -1);
// Return an iterator over the range deletion meta-block for the specified
// file number.
// @param skip_filters Disables loading/accessing the filter block
// @param level The level this table is at, -1 for "not set / don't know"
InternalIterator* NewRangeDeletionIterator(const ReadOptions& options,
const InternalKeyComparator& icmp,
const FileDescriptor& fd,
HistogramImpl* file_read_hist,
bool skip_filters, int level);
// If a seek to internal key "k" in specified file finds an entry, // If a seek to internal key "k" in specified file finds an entry,
// call (*handle_result)(arg, found_key, found_value) repeatedly until // call (*handle_result)(arg, found_key, found_value) repeatedly until
// it returns false. // it returns false.

Write Preview
Loading…
Cancel
Save