Use CacheDependencies() at start of ApproximateKeyAnchors() (#11230)

Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/11230

Test Plan:
- setup command: `$ ./db_bench -benchmarks=fillrandom,compact -compression_type=none -num=1000000 -write_buffer_size=4194304 -target_file_size_base=4194304 -use_direct_io_for_flush_and_compaction=true -partition_index_and_filters=true -bloom_bits=10 -metadata_block_size=1024`
- measure small read count bucketed by size: `$ strace -fye pread64 ./db_bench.ctrl -use_existing_db=true -benchmarks=compact -compaction_readahead_size=4194304 -compression_type=none -num=1000000 -write_buffer_size=4194304 -target_file_size_base=4194304 -use_direct_io_for_flush_and_compaction=true -partition_index_and_filters=true -bloom_bits=10 -metadata_block_size=1024  -subcompactions=4 -cache_size=1048576000  2>&1 >/dev/null | awk '/= [0-9]+$/{print "[", int($NF / 1024), "KB,", int(1 + $NF / 1024), "KB)"}' | sort -n -k 2 | uniq -c | head -3`
- before:
```
   1119 [ 0 KB, 1 KB)
      1 [ 6 KB, 7 KB)
      2 [ 7 KB, 8 KB)
```
- after:
```
    242 [ 0 KB, 1 KB)
      1 [ 6 KB, 7 KB)
      2 [ 7 KB, 8 KB)
```

Reviewed By: pdillinger

Differential Revision: D43388507

Pulled By: ajkr

fbshipit-source-id: a02413c9f615b00784700646825a9870ee10f3a7

diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index f79636b0e47722ed6214d9b65cc7ef7f3b931010..eff4df56d65fb596dbbad902cfac16bb2d278425 100644 (file)
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -1929,6 +1929,16 @@ Status BlockBasedTable::ApproximateKeyAnchors(const ReadOptions& read_options,
   // likely not to be a problem. We are compacting the whole file, so all
   // keys will be read out anyway. An extra read to index block might be
   // a small share of the overhead. We can try to optimize if needed.
+  //
+  // `CacheDependencies()` brings all the blocks into cache using one I/O. That
+  // way the full index scan usually finds the index data it is looking for in
+  // cache rather than doing an I/O for each "dependency" (partition).
+  Status s =
+      rep_->index_reader->CacheDependencies(read_options, false /* pin */);
+  if (!s.ok()) {
+    return s;
+  }
+
   IndexBlockIter iiter_on_stack;
   auto iiter = NewIndexIterator(
       read_options, /*disable_prefix_seek=*/false, &iiter_on_stack,

diff --git a/table/block_based/partitioned_index_reader.cc b/table/block_based/partitioned_index_reader.cc
index 705223c90acf64ae2a9830e927ee928813dc44eb..dbe8b2bd48f37ce1a6d00110bab0ef0cb117a766 100644 (file)
--- a/table/block_based/partitioned_index_reader.cc
+++ b/table/block_based/partitioned_index_reader.cc
@@ -113,6 +113,11 @@ InternalIteratorBase<IndexValue>* PartitionIndexReader::NewIterator(
 }
 Status PartitionIndexReader::CacheDependencies(const ReadOptions& ro,
                                                bool pin) {
+  if (!partition_map_.empty()) {
+    // The dependencies are already cached since `partition_map_` is filled in
+    // an all-or-nothing manner.
+    return Status::OK();
+  }
   // Before read partitions, prefetch them to avoid lots of IOs
   BlockCacheLookupContext lookup_context{TableReaderCaller::kPrefetch};
   const BlockBasedTable::Rep* rep = table()->rep_;