Summary:
As the code changes for block cache tracing are almost complete, I did a benchmark to compare the performance when block cache tracing is enabled/disabled.

 With 1% downsampling ratio, the performance overhead of block cache tracing is negligible. When we trace all block accesses, the throughput drops by 6 folds with 16 threads issuing random reads and all reads are served in block cache.

Setup:
RocksDB:    version 6.2
Date:       Mon Jun 17 17:11:13 2019
CPU:        24 * Intel Core Processor (Skylake)
CPUCache:   16384 KB
Keys:       20 bytes each
Values:     100 bytes each (100 bytes after compression)
Entries:    10000000
Prefix:    20 bytes
Keys per prefix:    0
RawSize:    1144.4 MB (estimated)
FileSize:   1144.4 MB (estimated)
Write rate: 0 bytes/second
Read rate: 0 ops/second
Compression: NoCompression
Compression sampling rate: 0
Memtablerep: skip_list
Perf Level: 1

I ran the readrandom workload for 1 minute. Detailed throughput results:  (ops/second)
Sample rate 0: no block cache tracing.
Sample rate 1: trace all block accesses.
Sample rate 100: trace accesses 1% blocks.
1 thread |   |   |  -- | -- | -- | --
Sample rate | 0 | 1 | 100
1 MB block cache size | 13,094 | 13,166 | 13,341
10 GB block cache size | 202,243 | 188,677 | 229,182

16 threads |   |   |  -- | -- | -- | --
Sample rate | 0 | 1 | 100
1 MB block cache size | 208,761 | 178,700 | 201,872
10 GB block cache size | 2,645,996 | 426,295 | 2,587,605
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5473

Differential Revision: D15869479

Pulled By: HaoyuHuang

fbshipit-source-id: 7ae802abe84811281a6af8649f489887cd7c4618