Reorder field based on esan data

Summary:
Running. TEST_TMPDIR=/dev/shm ./buck-out/gen/rocks/tools/rocks_db_bench --benchmarks=readwhilewriting --num=5000000 -benchmark_write_rate_limit=2000000 --threads=32

Collected esan data and reorder field. Accesses to 4th and 6th fields take majority of the access.  Group them. Overall, this struct takes 10%+ of the total accesses in the program. (637773011/6107964986)

==2433831==  class rocksdb::InlineSkipList
==2433831==   size = 48, count = 637773011, ratio = 112412, array access = 0
==2433831==   # 0: offset = 0,   size = 2,       count = 455137, type = i16
==2433831==   # 1: offset = 2,   size = 2,       count = 6,      type = i16
==2433831==   # 2: offset = 4,   size = 4,       count = 182303, type = i32
==2433831==   # 3: offset = 8,   size = 8,       count = 263953900, type = %"class.rocksdb::MemTableRep::KeyComparator"*
==2433831==   # 4: offset = 16,  size = 8,       count = 136409, type = %"class.rocksdb::Allocator"*
==2433831==   # 5: offset = 24,  size = 8,       count = 366628820, type = %"struct.rocksdb::InlineSkipList<const rocksdb::MemTableRep::KeyComparator &>::Node"*
==2433831==   # 6: offset = 32,  size = 4,       count = 6280031, type = %"struct.std::atomic" = type { %"struct.std::__atomic_base" }
==2433831==   # 7: offset = 40,  size = 8,       count = 136405, type = %"struct.rocksdb::InlineSkipList<const rocksdb::MemTableRep::KeyComparator &>::Splice"*
==2433831==EfficiencySanitizer: total struct field access count = 6107964986

Before re-ordering
[trentxintong@devbig460.frc2 ~/fbsource/fbcode]$ fgrep readwhilewriting
without-ro.log
readwhilewriting :       0.036 micros/op 27545605 ops/sec;   26.8 MB/s
(45954 of 5000000 found)
readwhilewriting :       0.036 micros/op 28024240 ops/sec;   27.2 MB/s
(43158 of 5000000 found)
readwhilewriting :       0.037 micros/op 27345145 ops/sec;   27.1 MB/s
(46725 of 5000000 found)
readwhilewriting :       0.037 micros/op 27072588 ops/sec;   27.3 MB/s
(42605 of 5000000 found)
readwhilewriting :       0.034 micros/op 29578781 ops/sec;   28.3 MB/s
(44294 of 5000000 found)
readwhilewriting :       0.035 micros/op 28528304 ops/sec;   27.7 MB/s
(44176 of 5000000 found)
readwhilewriting :       0.037 micros/op 27075497 ops/sec;   26.5 MB/s
(43763 of 5000000 found)
readwhilewriting :       0.036 micros/op 28024117 ops/sec;   27.1 MB/s
(40622 of 5000000 found)
readwhilewriting :       0.037 micros/op 27078709 ops/sec;   27.6 MB/s
(47774 of 5000000 found)
readwhilewriting :       0.034 micros/op 29020689 ops/sec;   28.1 MB/s
(45066 of 5000000 found)
AVERAGE()=27.37 MB/s

After re-ordering
[trentxintong@devbig460.frc2 ~/fbsource/fbcode]$ fgrep readwhilewriting
ro.log
readwhilewriting :       0.036 micros/op 27542409 ops/sec;   27.7 MB/s
(46163 of 5000000 found)
readwhilewriting :       0.036 micros/op 28021148 ops/sec;   28.2 MB/s
(46155 of 5000000 found)
readwhilewriting :       0.036 micros/op 28021035 ops/sec;   27.3 MB/s
(44039 of 5000000 found)
readwhilewriting :       0.036 micros/op 27538659 ops/sec;   27.5 MB/s
(46781 of 5000000 found)
readwhilewriting :       0.036 micros/op 28028604 ops/sec;   27.6 MB/s
(44689 of 5000000 found)
readwhilewriting :       0.036 micros/op 27541452 ops/sec;   27.3 MB/s
(43156 of 5000000 found)
readwhilewriting :       0.034 micros/op 29041338 ops/sec;   28.8 MB/s
(44895 of 5000000 found)
readwhilewriting :       0.036 micros/op 27784974 ops/sec;   26.3 MB/s
(39963 of 5000000 found)
readwhilewriting :       0.036 micros/op 27538892 ops/sec;   28.1 MB/s
(46570 of 5000000 found)
readwhilewriting :       0.038 micros/op 26622473 ops/sec;   27.0 MB/s
(43236 of 5000000 found)
AVERAGE()=27.58 MB/s
Closes https://github.com/facebook/rocksdb/pull/3855

Reviewed By: siying

Differential Revision: D8048781

Pulled By: trentxintong

fbshipit-source-id: bc9807a9845e2a92cb171ce1ecb5a2c8a51f1481
This commit is contained in:
Xin Tong 2018-05-17 17:49:06 -07:00 committed by Facebook Github Bot
parent fa43948cbc
commit aed7abbcca

View File

@ -184,10 +184,9 @@ class InlineSkipList {
const uint16_t kBranching_;
const uint32_t kScaledInverseBranching_;
Allocator* const allocator_; // Allocator used for allocations of nodes
// Immutable after construction
Comparator const compare_;
Allocator* const allocator_; // Allocator used for allocations of nodes
Node* const head_;
// Modified only by Insert(). Read racily by readers, but stale
@ -587,8 +586,8 @@ InlineSkipList<Comparator>::InlineSkipList(const Comparator cmp,
: kMaxHeight_(static_cast<uint16_t>(max_height)),
kBranching_(static_cast<uint16_t>(branching_factor)),
kScaledInverseBranching_((Random::kMaxNext + 1) / kBranching_),
compare_(cmp),
allocator_(allocator),
compare_(cmp),
head_(AllocateNode(0, max_height)),
max_height_(1),
seq_splice_(AllocateSplice()) {