"
The "promoted index" is how the sstable format calls the clustering key index within a given partition.
Large partitions with many rows have it. It's embedded in the partition index entry.
Currently, lookups in the promoted index are done by scanning the index linearly so the lookup
is O(N). For large partitions that's inefficient. It consumes both a lot of CPU and I/O.
We could do better and use binary search in the index. This patch series switches the mc-format
index reader to do that. Other formats use the old way.
The "mc" format promoted index has an extra structure at the end of the index called "offset map".
It's a vector of offsets of consecutive promoted index entries. This allows us to access random
entries in the index without reading the whole index.
The location of the offset entry for a given promoted index entry can be derived by knowing where
the offset vector ends in the index file, so the offset map also doesn't have to be read completely
into the memory.
The most tricky part is caching. We need to cache blocks read from the index file to amortize the
cost of binary search:
- if the promoted index fits in the 32 KiB which was read from the index when looking for
the partition entry, we don't want to issue any additional I/O to search the promoted index.
- with large promoted indexes, the last few bisections will fall into the same I/O block and we
want to reuse that block.
- we don't want the cache to grow too big, we don't want to cache the whole promoted index
as the read progresses over the index. Scanning reads may skip multiple times.
This series implements a rather simple approach which meets all the
above requirements and is not worse than the current state of affairs:
- Each index cursor has its own cache of the index file area which corresponds to promoted index
This is managed by the cached_file class.
- Each index cursor has its own cache of parsed blocks. This allows the upper bound estimation to
reuse information obtained during lower bound lookup. This estimation is used to limit
read-aheads in the data file.
- Each cursor drops entries that it walked past so that memory footprint stays O(log N)
- Cached buffers are accounted to read's reader_permit.
Later, we could have a single cache shared by many readers. For that, we need to come up with eviction
policy.
Fixes#4007.
TESTING RESULTS
* Point reads, large promoted index:
Config: rows: 10000000, value size: 2000
Partition size: 20 GB
Index size: 7 MB
Notes:
- Slicing read into the middle of partition (offset=5000000, read=1) is a clear win for the binary search:
time: 1.9ms vs 22.9ms
CPU utilization: 8.9% vs 92.3%
I/O: 21 reqs / 172 KiB vs 29 reqs / 3'520 KiB
It's 12x faster, CPU utilization is 10x times smaller, disk utilization is 20x smaller.
- Slicing at the front (offset=0) is a mixed bag.
time is similar: 1.8ms
CPU utilization is 6.7x smaller for bsearch: 8.5% vs 57.7%
disk bandwidth utilization is smaller for bsearch but uses more IOs: 4 reqs / 320 KiB (scan) vs 17 reqs / 188 KiB (bsearch)
bsearch uses less bandwidth because the series reduces buffer size used for index file I/O.
scan is issuing:
2 * 128 KB (index page)
2 * 32 KB (data file)
bsearch is issuing:
1 * 64 KB (index page)
15 * 4 KB (promoted index)
1 * 64 KB (data file)
The 1 * 64 KB is chosen dynamically by seastar. Sometimes it chooses 2 * 32 KB (with read-ahead).
32 KB is the minimum I/O currently.
Disk utilization could be further improved by changing the way seastar's dynamic I/O adjustments work
so that it uses 1 * 4 KB when it suffices. This is left for the follow-up.
Command:
perf_fast_forward --datasets=large-part-ds1 \
--run-tests=large-partition-slicing-clustering-keys -c1 --test-case-duration=1
Before:
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.001836 172 1 545 9 563 175 4.0 4 320 2 2 0 1 1 0 0 0 57.7% 0
0 32 0.001858 502 32 17220 126 17776 11526 3.2 3 324 2 1 0 1 1 0 0 0 56.4% 0
0 256 0.002833 339 256 90374 427 91757 85931 7.0 7 776 3 1 0 1 1 0 0 0 41.1% 0
0 4096 0.017211 58 4096 237984 2011 241802 233870 66.1 66 8376 59 2 0 1 1 0 0 0 21.4% 0
5000000 1 0.022952 42 1 44 1 45 41 29.2 29 3520 22 2 0 1 1 0 0 0 92.3% 0
5000000 32 0.023052 43 32 1388 14 1414 1331 31.1 32 3588 26 2 0 1 1 0 0 0 91.7% 0
5000000 256 0.024795 41 256 10325 129 10721 9993 43.1 39 4544 29 2 0 1 1 0 0 0 86.4% 0
5000000 4096 0.038856 27 4096 105414 398 106918 103162 95.2 95 12160 78 5 0 1 1 0 0 0 61.4% 0
After (v2):
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.001831 248 1 546 21 581 252 17.6 17 188 2 0 0 1 1 0 0 0 8.5% 0
0 32 0.001910 535 32 16751 626 17770 13896 17.9 19 160 3 0 0 1 1 0 0 0 8.8% 0
0 256 0.003545 266 256 72207 2333 89076 62852 26.9 24 764 7 0 0 1 1 0 0 0 9.7% 0
0 4096 0.016800 56 4096 243812 524 245430 239736 83.6 83 8700 64 0 0 1 1 0 0 0 16.6% 0
5000000 1 0.001968 351 1 508 19 538 380 21.3 21 172 2 0 0 1 1 0 0 0 8.9% 0
5000000 32 0.002273 431 32 14077 436 15503 11551 22.7 22 268 3 0 0 1 1 0 0 0 8.9% 0
5000000 256 0.003889 257 256 65824 2197 81833 57813 34.0 37 652 18 0 0 1 1 0 0 0 11.2% 0
5000000 4096 0.017115 54 4096 239324 834 241310 231993 88.3 88 8844 65 0 0 1 1 0 0 0 16.8% 0
After (v1):
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.001886 259 1 530 4 545 261 18.0 18 376 2 2 0 1 1 0 0 0 9.1% 0
0 32 0.001954 513 32 16381 93 16844 15618 19.0 19 408 3 2 0 1 1 0 0 0 9.3% 0
0 256 0.003266 318 256 78393 1820 81567 61663 30.8 26 1272 7 2 0 1 1 0 0 0 10.4% 0
0 4096 0.017991 57 4096 227666 855 231915 225781 83.1 83 8888 55 5 0 1 1 0 0 0 15.5% 0
5000000 1 0.002353 232 1 425 2 432 232 23.0 23 396 2 2 0 1 1 0 0 0 8.7% 0
5000000 32 0.002573 384 32 12437 47 12571 429 25.0 25 460 4 2 0 1 1 0 0 0 8.5% 0
5000000 256 0.003994 259 256 64101 2904 67924 51427 37.0 35 1484 11 2 0 1 1 0 0 0 10.6% 0
5000000 4096 0.018567 56 4096 220609 448 227395 219029 89.8 89 9036 59 5 0 1 1 0 0 0 15.1% 0
* Point reads, small promoted index (two blocks):
Config: rows: 400, value size: 200
Partition size: 84 KiB
Index size: 65 B
Notes:
- No significant difference in time
- the same disk utilization
- similar CPU utilization
Command:
perf_fast_forward --datasets=large-part-ds1 \
--run-tests=large-partition-slicing-clustering-keys -c1 --test-case-duration=1
Before:
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.000279 470 1 3587 31 3829 478 3.0 3 68 2 1 0 1 1 0 0 0 21.1% 0
0 32 0.000276 3498 32 116038 811 122756 104033 3.0 3 68 2 1 0 1 1 0 0 0 24.0% 0
0 256 0.000412 2554 256 621044 1778 732150 559221 2.0 2 72 2 0 0 1 1 0 0 0 32.6% 0
0 4096 0.000510 1901 400 783883 4078 819058 665616 2.0 2 88 2 0 0 1 1 0 0 0 36.4% 0
200 1 0.000339 2712 1 2951 8 3001 2569 2.0 2 72 2 0 0 1 1 0 0 0 17.8% 0
200 32 0.000352 2586 32 91019 266 92427 83411 2.0 2 72 2 0 0 1 1 0 0 0 20.8% 0
200 256 0.000458 2073 200 436503 1618 453945 385501 2.0 2 88 2 0 0 1 1 0 0 0 29.4% 0
200 4096 0.000458 2097 200 436475 1676 458349 381558 2.0 2 88 2 0 0 1 1 0 0 0 29.0% 0
After (v1):
Testing slicing of large partition using clustering keys:
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.000278 492 1 3598 30 3831 500 3.0 3 68 2 1 0 1 1 0 0 0 19.4% 0
0 32 0.000275 3433 32 116153 753 122915 92559 3.0 3 68 2 1 0 1 1 0 0 0 22.5% 0
0 256 0.000458 2576 256 559437 2978 728075 504375 2.1 2 88 2 0 0 1 1 0 0 0 29.0% 0
0 4096 0.000506 1888 400 790064 3306 822360 623109 2.0 2 88 2 0 0 1 1 0 0 0 36.6% 0
200 1 0.000382 2493 1 2619 10 2675 2268 2.0 2 88 2 0 0 1 1 0 0 0 16.3% 0
200 32 0.000398 2393 32 80422 333 84759 22281 2.0 2 88 2 0 0 1 1 0 0 0 19.0% 0
200 256 0.000459 2096 200 435943 1608 453989 380749 2.0 2 88 2 0 0 1 1 0 0 0 30.5% 0
200 4096 0.000458 2097 200 436410 1651 455779 382485 2.0 2 88 2 0 0 1 1 0 0 0 29.2% 0
* Scan with skips, large index:
Config: rows: 10000000, value size: 2000
Partition size: 20 GB
Index size: 7 MB
Notes:
- Similar time, slightly worse for binary search: 36.1 s (scan) vs 36.4 (bsearch)
- Slightly more I/O for bsearch: 153'932 reqs / 19'703'260 KiB (scan) vs 155'651 reqs / 19'704'088 KiB (bsearch)
Binary search reads more by 828 KB and by 1719 IOs.
It does more I/O to read the the promoted index offset map.
- similar (low) memory footprint. The danger here is that by caching index blocks which we touch as we scan
we would end up caching the whole index. But this is protected against by eviction as demonstrated by the
last "mem" column.
Command:
perf_fast_forward --datasets=large-part-ds1 \
--run-tests=large-partition-skips -c1 --test-case-duration=1
Before:
read skip time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
1 1 36.103451 4 5000000 138491 38 138601 138453 153932.0 153932 19703260 153561 1 0 1 1 0 0 0 31.5% 502690
After (v2):
read skip time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
1 1 37.000145 4 5000000 135135 6 135146 135128 155651.0 155651 19704088 138968 0 0 1 1 0 0 0 34.2% 0
After (v1):
read skip time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
1 1 36.965520 4 5000000 135261 30 135311 135231 155628.0 155628 19704216 139133 1 0 1 1 0 0 0 33.9% 248738
Also in:
git@github.com:tgrabiec/scylla.git sstable-use-index-offset-map-v2
Tests:
- unit (all modes)
- manual using perf_fast_forward
"
* tag 'sstable-use-index-offset-map-v2' of github.com:tgrabiec/scylla:
sstables: Add promoted index cache metrics
position_in_partition: Introduce external_memory_usage()
cached_file, sstables: Add tracing to index binary search and page cache
sstables: Dynamically adjust I/O size for index reads
sstables, tests: Allow disabling binary search in promoted index from perf tests
sstables: mc: Use binary search over the promoted index
utils: Introduce cached_file
sstables: clustered_index: Relax scope of validity of entry_info
sstables: index_entry: Introduce owning promoted_index_block_position
compound_compat: Allow constructing composite from a view
sstables: index_entry: Rename promoted_index_block_position to promoted_index_block_position_view
sstables: mc: Extract parser for promoted index block
sstables: mc: Extract parser for clustering out of the promoted index block parser
sstables: consumer: Extract primitive_consumer
sstables: Abstract the clustering index cursor behavior
sstables: index_reader: Rearrange to reduce branching and optionals
9322c07c71