Commit Graph

13 Commits

Author SHA1 Message Date
Avi Kivity
1f7dca0225 Merge 'Fix bad performance for densely populated partition index pages' from Tomasz Grabiec
This applies to small partition workload where index pages have high partition count, and the index doesn't fit in cache. It was observed that the count can be in the order of hundreds. In such a workload pages undergo constant population, LSA compaction, and LSA eviction, which has severe impact on CPU utilization.

Refs https://scylladb.atlassian.net/browse/SCYLLADB-620

This PR reduces the impact by several changes:

  - reducing memory footprint in the partition index. Assuming partition key size is 16 bytes, the cost dropped from 96 bytes to 36 bytes per partition.

  - flattening the object graph and amortizing storage. Storing entries directly in the vector. Storing all key values in a single managed_bytes. Making index_entry a trivial struct.

  - index entries and key storage are now trivially moveable, and batched inside vector storage
    so LSA migration can use memcpy(), which amortizes the cost per key. This reduces the cost of LSA segment compaction.

 - LSA eviction is now pretty much constant time for the whole page
   regardless of the number of entries, because elements are trivial and batched inside vectors.
   Page eviction cost dropped from 50 us to 1 us.

Performance evaluated with:

   scylla perf-simple-query -c1 -m200M --partitions=1000000

Before:

```
7774.96 tps (166.0 allocs/op, 521.7 logallocs/op,  54.0 tasks/op,  802428 insns/op,  430457 cycles/op,        0 errors)
7511.08 tps (166.1 allocs/op, 527.2 logallocs/op,  54.0 tasks/op,  804185 insns/op,  430752 cycles/op,        0 errors)
7740.44 tps (166.3 allocs/op, 526.2 logallocs/op,  54.2 tasks/op,  805347 insns/op,  432117 cycles/op,        0 errors)
7818.72 tps (165.2 allocs/op, 517.6 logallocs/op,  53.7 tasks/op,  794965 insns/op,  427751 cycles/op,        0 errors)
7865.49 tps (165.1 allocs/op, 513.3 logallocs/op,  53.6 tasks/op,  788898 insns/op,  425171 cycles/op,        0 errors)
```

After (+318%):

```
32492.40 tps (130.7 allocs/op,  12.8 logallocs/op,  36.1 tasks/op,  109236 insns/op,  103203 cycles/op,        0 errors)
32591.99 tps (130.4 allocs/op,  12.8 logallocs/op,  36.0 tasks/op,  108947 insns/op,  102889 cycles/op,        0 errors)
32514.52 tps (130.6 allocs/op,  12.8 logallocs/op,  36.0 tasks/op,  109118 insns/op,  103219 cycles/op,        0 errors)
32491.14 tps (130.6 allocs/op,  12.8 logallocs/op,  36.0 tasks/op,  109349 insns/op,  103272 cycles/op,        0 errors)
32582.90 tps (130.5 allocs/op,  12.8 logallocs/op,  36.0 tasks/op,  109269 insns/op,  102872 cycles/op,        0 errors)
32479.43 tps (130.6 allocs/op,  12.8 logallocs/op,  36.0 tasks/op,  109313 insns/op,  103242 cycles/op,        0 errors)
32418.48 tps (130.7 allocs/op,  12.8 logallocs/op,  36.1 tasks/op,  109201 insns/op,  103301 cycles/op,        0 errors)
31394.14 tps (130.7 allocs/op,  12.8 logallocs/op,  36.1 tasks/op,  109267 insns/op,  103301 cycles/op,        0 errors)
32298.55 tps (130.7 allocs/op,  12.8 logallocs/op,  36.1 tasks/op,  109323 insns/op,  103551 cycles/op,        0 errors)
```

When the workload is miss-only, with both row cache and index cache disabled (no cache maintenance cost):

  perf-simple-query -c1 -m200M --duration 6000 --partitions=100000 --enable-index-cache=0 --enable-cache=0

Before:

```
9124.57 tps (146.2 allocs/op, 789.0 logallocs/op,  45.3 tasks/op,  889320 insns/op,  357937 cycles/op,        0 errors)
9437.23 tps (146.1 allocs/op, 789.3 logallocs/op,  45.3 tasks/op,  889613 insns/op,  357782 cycles/op,        0 errors)
9455.65 tps (146.0 allocs/op, 787.4 logallocs/op,  45.2 tasks/op,  887606 insns/op,  357167 cycles/op,        0 errors)
9451.22 tps (146.0 allocs/op, 787.4 logallocs/op,  45.3 tasks/op,  887627 insns/op,  357357 cycles/op,        0 errors)
9429.50 tps (146.0 allocs/op, 787.4 logallocs/op,  45.3 tasks/op,  887761 insns/op,  358148 cycles/op,        0 errors)
9430.29 tps (146.1 allocs/op, 788.2 logallocs/op,  45.3 tasks/op,  888501 insns/op,  357679 cycles/op,        0 errors)
9454.08 tps (146.0 allocs/op, 787.3 logallocs/op,  45.3 tasks/op,  887545 insns/op,  357132 cycles/op,        0 errors)
```

After (+55%):

```
14484.84 tps (150.7 allocs/op,   6.5 logallocs/op,  44.7 tasks/op,  396164 insns/op,  229490 cycles/op,        0 errors)
14526.21 tps (150.8 allocs/op,   6.5 logallocs/op,  44.8 tasks/op,  396401 insns/op,  228824 cycles/op,        0 errors)
14567.53 tps (150.7 allocs/op,   6.5 logallocs/op,  44.7 tasks/op,  396319 insns/op,  228701 cycles/op,        0 errors)
14545.63 tps (150.6 allocs/op,   6.5 logallocs/op,  44.7 tasks/op,  395889 insns/op,  228493 cycles/op,        0 errors)
14626.06 tps (150.5 allocs/op,   6.5 logallocs/op,  44.7 tasks/op,  395254 insns/op,  227891 cycles/op,        0 errors)
14593.74 tps (150.5 allocs/op,   6.5 logallocs/op,  44.7 tasks/op,  395480 insns/op,  227993 cycles/op,        0 errors)
14538.10 tps (150.8 allocs/op,   6.5 logallocs/op,  44.8 tasks/op,  397035 insns/op,  228831 cycles/op,        0 errors)
14527.18 tps (150.8 allocs/op,   6.5 logallocs/op,  44.8 tasks/op,  396992 insns/op,  228839 cycles/op,        0 errors)
```

Same as above, but with summary ratio increased from 0.0005 to 0.005 (smaller pages):

Before:

```
33906.70 tps (146.1 allocs/op,  83.6 logallocs/op,  45.1 tasks/op,  170553 insns/op,   98104 cycles/op,        0 errors)
32696.16 tps (146.0 allocs/op,  83.5 logallocs/op,  45.1 tasks/op,  170369 insns/op,   98405 cycles/op,        0 errors)
33889.05 tps (146.1 allocs/op,  83.6 logallocs/op,  45.1 tasks/op,  170551 insns/op,   98135 cycles/op,        0 errors)
33893.24 tps (146.1 allocs/op,  83.5 logallocs/op,  45.1 tasks/op,  170488 insns/op,   98168 cycles/op,        0 errors)
33836.73 tps (146.1 allocs/op,  83.6 logallocs/op,  45.1 tasks/op,  170528 insns/op,   98226 cycles/op,        0 errors)
33897.61 tps (146.0 allocs/op,  83.5 logallocs/op,  45.1 tasks/op,  170428 insns/op,   98081 cycles/op,        0 errors)
33834.73 tps (146.1 allocs/op,  83.5 logallocs/op,  45.1 tasks/op,  170438 insns/op,   98178 cycles/op,        0 errors)
33776.31 tps (146.3 allocs/op,  83.9 logallocs/op,  45.2 tasks/op,  170958 insns/op,   98418 cycles/op,        0 errors)
33808.08 tps (146.3 allocs/op,  83.9 logallocs/op,  45.2 tasks/op,  170940 insns/op,   98388 cycles/op,        0 errors)
```

After (+18%):

```
40081.51 tps (148.2 allocs/op,   4.4 logallocs/op,  45.0 tasks/op,  121047 insns/op,   82231 cycles/op,        0 errors)
40005.85 tps (148.6 allocs/op,   4.4 logallocs/op,  45.2 tasks/op,  121327 insns/op,   82545 cycles/op,        0 errors)
39816.75 tps (148.3 allocs/op,   4.4 logallocs/op,  45.1 tasks/op,  121067 insns/op,   82419 cycles/op,        0 errors)
39953.11 tps (148.1 allocs/op,   4.4 logallocs/op,  45.0 tasks/op,  121027 insns/op,   82258 cycles/op,        0 errors)
40073.96 tps (148.2 allocs/op,   4.4 logallocs/op,  45.0 tasks/op,  121006 insns/op,   82313 cycles/op,        0 errors)
39882.25 tps (148.2 allocs/op,   4.4 logallocs/op,  45.0 tasks/op,  120925 insns/op,   82320 cycles/op,        0 errors)
39916.08 tps (148.3 allocs/op,   4.4 logallocs/op,  45.1 tasks/op,  121054 insns/op,   82393 cycles/op,        0 errors)
39786.30 tps (148.2 allocs/op,   4.4 logallocs/op,  45.0 tasks/op,  121027 insns/op,   82465 cycles/op,        0 errors)
38662.45 tps (148.3 allocs/op,   4.4 logallocs/op,  45.0 tasks/op,  121108 insns/op,   82312 cycles/op,        0 errors)
39849.42 tps (148.3 allocs/op,   4.4 logallocs/op,  45.1 tasks/op,  121098 insns/op,   82447 cycles/op,        0 errors)
```

Closes scylladb/scylladb#28603

* github.com:scylladb/scylladb:
  sstables: mx: index_reader: Optimize parsing for no promoted index case
  vint: Use std::countl_zero()
  test: sstable_partition_index_cache_test: Validate scenario of pages with sparse promoted index placement
  sstables: mx: index_reader: Amoritze partition key storage
  managed_bytes: Hoist write_fragmented() to common header
  utils: managed_vector: Use std::uninitialized_move() to move objects
  sstables: mx: index_reader: Keep promoted_index info next to index_entry
  sstables: mx: index_reader: Extract partition_index_page::clear_gently()
  sstables: mx: index_reader: Shave-off 16 bytes from index_entry by using raw_token
  sstables: mx: index_reader: Reduce allocation_section overhead during index page parsing by batching allocation
  sstables: mx: index_reader: Keep index_entry directly in the vector
  dht: Introduce raw_token
  test: perf_simple_query: Add 'sstable-format' command-line option
  test: perf_simple_query: Add 'sstable-summary-ratio' command-line option
  test: perf-simple-query: Add option to disable index cache
  test: cql_test_env: Respect enable-index-cache config

(cherry picked from commit 5e7fb08bf3)

Closes scylladb/scylladb#29136

Closes scylladb/scylladb#29140
2026-03-20 10:58:26 +02:00
Dawid Mędrek
a151944fa6 treewide: Replace __builtin_expect with (un)likely
C++20 introduced two new attributes--likely and unlikely--that
function as a built-in replacement for __builtin_expect implemented
in various compilers. Since it makes code easier to read and it's
an integral part of the language, there's no reason to not use it
instead.

Closes scylladb/scylladb#24786
2025-07-03 13:34:04 +03:00
Avi Kivity
f3eade2f62 treewide: relicense to ScyllaDB-Source-Available-1.0
Drop the AGPL license in favor of a source-available license.
See the blog post [1] for details.

[1] https://www.scylladb.com/2024/12/18/why-were-moving-to-a-source-available-license/
2024-12-18 17:45:13 +02:00
Kefu Chai
0ae81446ef ./: not include unused headers
these unused includes were identified by clangd. see
https://clangd.llvm.org/guides/include-cleaner#unused-include-warning
for more details on the "Unused include" warning.

Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>

Closes scylladb/scylladb#16766
2024-01-17 16:30:14 +02:00
Avi Kivity
5937b1fa23 treewide: remove empty comments in top-of-files
After fcb8d040 ("treewide: use Software Package Data Exchange
(SPDX) license identifiers"), many dual-licensed files were
left with empty comments on top. Remove them to avoid visual
noise.

Closes #10562
2022-05-13 07:11:58 +02:00
Avi Kivity
fcb8d040e8 treewide: use Software Package Data Exchange (SPDX) license identifiers
Instead of lengthy blurbs, switch to single-line, machine-readable
standardized (https://spdx.dev) license identifiers. The Linux kernel
switched long ago, so there is strong precedent.

Three cases are handled: AGPL-only, Apache-only, and dual licensed.
For the latter case, I chose (AGPL-3.0-or-later and Apache-2.0),
reasoning that our changes are extensive enough to apply our license.

The changes we applied mechanically with a script, except to
licenses/README.md.

Closes #9937
2022-01-18 12:15:18 +01:00
Avi Kivity
a55b434a2b treewide: extent copyright statements to present day 2021-06-06 19:18:49 +03:00
Paweł Dziepak
552fc0c6b9 vint: optimise deserialisation routine
At the moment, vint deserialisation is using a naive approach, reading
each byte separately. In practice, vints are going to most often appears
inside larger buffers. That means we can read 8-bytes at a time end then
figure out unneded parts and mask them out. This way we avoid a loop and
do less memory loads which are much more expensive than arithmetic
operations (even if they hit the cache).
2019-03-14 13:37:06 +00:00
Paweł Dziepak
57de2c26b3 vint: drop deserialize_type structure
Deserialisation function returns a structure containing both the value
and its length in the input buffer. In the vast majority of the cases
the caller will already know the length and having this structure will
make it harder for the compiler to emit good code, especially if the
function is not inlined.

In practice I've seen the structure causing register pressure problems
that lead to spilling variables to memory.
2019-03-14 13:37:06 +00:00
Duarte Nunes
fa2b0384d2 Replace std::experimental types with C++17 std version.
Replace stdx::optional and stdx::string_view with the C++ std
counterparts.

Some instances of boost::variant were also replaced with std::variant,
namely those that called seastar::visit.

Scylla now requires GCC 8 to compile.

Signed-off-by: Duarte Nunes <duarte@scylladb.com>
Message-Id: <20190108111141.5369-1-duarte@scylladb.com>
2019-01-08 13:16:36 +02:00
Piotr Jastrzebski
b602dea726 Add signed_vint::serialized_size_from_first_byte
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
2018-05-09 11:41:00 +02:00
Piotr Jastrzebski
e423529077 core: add unsigned_vint::serialized_size_from_first_byte
This method takes first byte and determins how many bytes
are used to represent an unsigned variant integer.

Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
2018-04-16 20:12:03 +02:00
Jesse Haber-Kucharsky
91dab1d998 CQL native protocol: Add support for vint serialization
Version 5 of the native protocol for CQL [1] adds the `vint` and `unsigned vint`
types.

An unsigned integer encoded as a `vint` has a variable size based on the
magnitude of the value. The first byte indicates the total number of bytes.

For signed integers, a "zig-zag" encoding scheme ensures that small negative
values are encoded as short-length `vint`s (0 -> 0, -1 -> 1, 1 -> 2, 2 -> 3, -2
-> 4, etc).

[1] https://github.com/apache/cassandra/blob/trunk/doc/native_protocol_v5.spec
2017-08-10 14:11:30 -04:00