Optimize aggregation performance of nullable String key when use AggregationMethodSerialized by liuneng1994 · Pull Request #51399 · ClickHouse/ClickHouse

liuneng1994 · 2023-06-26T04:48:41Z

Changelog category (leave one):

Performance Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

optimize aggregation performance of Nullable String key when using AggregationMethodSerialized

liuneng1994 · 2023-06-26T04:51:06Z

Same as #46812. The current implementation can improve TPCH as a whole by 4%.

robot-clickhouse · 2023-06-26T10:40:06Z

This is an automated comment for commit 65aeb05 with description of existing statuses. It's updated for the latest CI running
The full report is available here
The overall status of the commit is 🟢 success

Check name	Description	Status
AST fuzzer	Runs randomly generated queries to catch program errors. The build type is optionally given in parenthesis. If it fails, ask a maintainer for help	🟢 success
CI running	A meta-check that indicates the running CI. Normally, it's in success or pending state. The failed status indicates some problems with the PR	🟢 success
ClickHouse build check	Builds ClickHouse in various configurations for use in further steps. You have to fix the builds that fail. Build logs often has enough information to fix the error, but you might have to reproduce the failure locally. The cmake options can be found in the build log, grepping for cmake. Use these options and follow the general build process	🟢 success
Compatibility check	Checks that clickhouse binary runs on distributions with old libc versions. If it fails, ask a maintainer for help	🟢 success
Docker image for servers	The check to build and optionally push the mentioned image to docker hub	🟢 success
Fast test	Normally this is the first check that is ran for a PR. It builds ClickHouse and runs most of stateless functional tests, omitting some. If it fails, further checks are not started until it is fixed. Look at the report to see which tests fail, then reproduce the failure locally as described here	🟢 success
Flaky tests	Checks if new added or modified tests are flaky by running them repeatedly, in parallel, with more randomization. Functional tests are run 100 times with address sanitizer, and additional randomization of thread scheduling. Integrational tests are run up to 10 times. If at least once a new test has failed, or was too long, this check will be red. We don't allow flaky tests, read the doc	🟢 success
Install packages	Checks that the built packages are installable in a clear environment	🟢 success
Integration tests	The integration tests report. In parenthesis the package type is given, and in square brackets are the optional part/total tests	🟢 success
Mergeable Check	Checks if all other necessary checks are successful	🟢 success
Performance Comparison	Measure changes in query performance. The performance test report is described in detail here. In square brackets are the optional part/total tests	🟢 success
Push to Dockerhub	The check for building and pushing the CI related docker images to docker hub	🟢 success
SQLancer	Fuzzing tests that detect logical bugs with SQLancer tool	🟢 success
Sqllogic	Run clickhouse on the sqllogic test set against sqlite and checks that all statements are passed	🟢 success
Stateful tests	Runs stateful functional tests for ClickHouse binaries built in various configurations -- release, debug, with sanitizers, etc	🟢 success
Stateless tests	Runs stateless functional tests for ClickHouse binaries built in various configurations -- release, debug, with sanitizers, etc	🟢 success
Stress test	Runs stateless functional tests concurrently from several clients to detect concurrency-related errors	🟢 success
Style Check	Runs a set of checks to keep the code style clean. If some of tests failed, see the related log from the report	🟢 success
Unit tests	Runs the unit tests for different release types	🟢 success
Upgrade check	Runs stress tests on server version from last release and then tries to upgrade it to the version from the PR. It checks if the new server can successfully startup without any errors, crashes or sanitizer asserts	🟢 success

liuneng1994 · 2023-06-27T08:20:33Z

The slower case doesn't seem to be related to the changed code, and the results that come out of the two runs are different and seem unstable

nickitat · 2023-06-27T21:27:24Z

src/Columns/ColumnNullable.cpp

am I right that the whole idea of optimisation is to devirtualize serializeValueIntoArena? if so I think it would be better to just static_cast to ColumnString and call serializeValueIntoArena as usual.

I'm not quite sure that the key to this optimization is devirtualize serializeValueIntoArena, and I've tested that this optimization has performance degradation on the number type. So need to determine whether the nested column is a ColumnString

I tested the above modification and there was no impact on performance

I'm not quite sure that the key to this optimization is devirtualize serializeValueIntoArena, and I've tested that this optimization has performance degradation on the number type. So need to determine whether the nested column is a ColumnString

There are some new discoveries that I will try again to apply in the fixed size column

I would assume that calculating this condition for every element may hurt perf (especially for numeric types, because it is the fastest case). better would be to move this condition out of the hot loop

@nickitat After my test, the performance improvement should be due to function inline and reduced allocContinue calls. If want to modify this part of the code to each column class, the amount of change will be very large

I added two variables to ColumnNullable to represent the type of nested column. I thought about moving these two variables to HashMethodSerialized, but this optimization can only be used in aggregation, and other places where serializeIntoArea is used cannot obtain optimization effects. So I think it's more appropriate to place these two variables in ColumnNullable

liuneng1994 · 2023-06-28T07:12:43Z

If use getNestedColumn().valuesHaveFixedSize() to make a judgment, a lock instruction prefix is generated, which reduces performance

DB::ColumnNullable::serializeValueIntoArena(unsigned long, DB::Arena&, char const*&) const:
	pushq  %rbp
	pushq  %r15
	pushq  %r14
	pushq  %r13
	pushq  %r12
	pushq  %rbx
	subq   $0x28, %rsp
	movq   %rcx, %rbp
	movq   %rdx, %r15
	movq   %rsi, %rbx
	movq   %rdi, %r13
	movq   0x10(%rdi), %rdi
	movq   0x18(%r13), %r12
	movq   0x10(%r12), %rax
	movzbl (%rax,%rsi), %r14d
	movq   -0xfe8b163(%rip), %rcx     ; typeinfo for DB::ColumnString + 8
	cmpq   %rcx, -0xfe83ce2(%rip)     ; typeinfo for DB::IColumn + 8
	nopw   (%rax,%rax)
	je     0x177b2c98                 ; <+280> [inlined] DB::PODArray<unsigned long, 4096ul, Allocator<false, false>, 63ul, 64ul>::t_start() const at PODArray.h:324:83
	movq   (%rdi), %rax
	movq   -0x8(%rax), %rdx
	cmpq   %rcx, 0x8(%rdx)
	je     0x177b2c98                 ; <+280> [inlined] DB::PODArray<unsigned long, 4096ul, Allocator<false, false>, 63ul, 64ul>::t_start() const at PODArray.h:324:83
	callq  *0x260(%rax)
	nopl   (%rax)
	testb  %al, %al
	je     0x177b2d11                 ; <+401> at ColumnNullable.cpp
	movq   %r15, 0x10(%rsp)
	movq   0x10(%r13), %r13
	testq  %r13, %r13
	je     0x177b2bfb                 ; <+123> at ColumnNullable.cpp:160:26
	lock
	incl   0x8(%r13)
	movq   (%r13), %r15
	movq   %r13, %rdi
	movq   %rbx, %rsi
	movq   %r15, 0x8(%rsp)
	callq  *0x60(%r15)
	movq   %rax, 0x18(%rsp)
	movq   %r12, 0x20(%rsp)
	movq   %r13, %rdi
	callq  *0x278(%r15)
	movq   %rax, %r12
	incq   %rax
	testb  %r14b, %r14b
	movl   %r14d, %r15d
	movl   $0x1, %r14d
	cmoveq %rax, %r14
	movq   0x10(%rsp), %rdi
	movq   %r14, %rsi
	movq   %rbp, %rdx
	xorl   %ecx, %ecx
	callq  0xd9d3040                  ; DB::Arena::allocContinue(unsigned long, char const*&, unsigned long)
	movq   %rax, %rbp
	movq   0x20(%rsp), %rax
	movq   0x10(%rax), %rax
	movzbl (%rax,%rbx), %eax
	movb   %al, (%rbp)
	nopl   (%rax)
	testb  %r15b, %r15b
	jne    0x177b2c78                 ; <+248> [inlined] boost::detail::atomic_count::operator--() at atomic_count_gcc_atomic.hpp:43:16
	movq   %rbp, %rdi
	incq   %rdi
	movq   0x18(%rsp), %rsi
	movq   %r12, %rdx
	callq  0x1d11e740                 ; ::memcpy(void *__restrict, const void *__restrict, size_t) at memcpy.cpp:4
	lock
	decl   0x8(%r13)
	movq   0x8(%rsp), %r12
	jne    0x177b2d65                 ; <+485> at ColumnNullable.cpp:185:1
	movq   %r13, %rdi
	callq  *0x2b8(%r12)
	jmp    0x177b2d65                 ; <+485> at ColumnNullable.cpp:185:1
	movq   0x10(%rdi), %rax
	movq   0x28(%rdi), %rcx
	movq   %rcx, 0x10(%rsp)
	movq   -0x8(%rax,%rbx,8), %r13
	movq   %r15, %rdi
	movq   (%rax,%rbx,8), %r15
	subq   %r13, %r15
	leaq   0x9(%r15), %rax
	testb  %r14b, %r14b
	movl   $0x1, %esi
	movq   %rax, 0x8(%rsp)
	cmoveq %rax, %rsi
	movq   %rbp, %rdx
	xorl   %ecx, %ecx
	callq  0xd9d3040                  ; DB::Arena::allocContinue(unsigned long, char const*&, unsigned long)
	movq   %rax, %rbp
	movq   0x10(%r12), %rax
	movzbl (%rax,%rbx), %eax
	movb   %al, (%rbp)
	testb  %r14b, %r14b
	je     0x177b2cef                 ; <+367> at ColumnNullable.cpp
	movl   $0x1, %r14d
	jmp    0x177b2d65                 ; <+485> at ColumnNullable.cpp:185:1
	movq   0x10(%rsp), %rsi
	addq   %r13, %rsi
	movq   %r15, 0x1(%rbp)
	movq   %rbp, %rdi
	addq   $0x9, %rdi
	movq   %r15, %rdx
	callq  0x1d11e740                 ; ::memcpy(void *__restrict, const void *__restrict, size_t) at memcpy.cpp:4
	movq   0x8(%rsp), %r14
	jmp    0x177b2d65                 ; <+485> at ColumnNullable.cpp:185:1
	movl   $0x1, %r14d
	movl   $0x1, %esi
	movq   %r15, %rdi
	movq   %rbp, %rdx
	xorl   %ecx, %ecx
	callq  0xd9d3040                  ; DB::Arena::allocContinue(unsigned long, char const*&, unsigned long)
	movq   %rbp, %rcx
	movq   %rax, %rbp
	movq   0x10(%r12), %rax
	movzbl (%rax,%rbx), %eax
	movb   %al, (%rbp)
	movq   0x10(%r12), %rax
	cmpb   $0x0, (%rax,%rbx)
	jne    0x177b2d65                 ; <+485> at ColumnNullable.cpp:185:1
	movq   0x10(%r13), %rdi
	movq   (%rdi), %rax
	movq   %rbx, %rsi
	movq   %r15, %rdx
	callq  *0xf8(%rax)
	movq   %rax, %rbp
	movq   %rdx, %r14
	decq   %rbp
	incq   %r14
	movq   %rbp, %rax
	movq   %r14, %rdx
	addq   $0x28, %rsp
	popq   %rbx
	popq   %r12
	popq   %r13
	popq   %r14
	popq   %r15
	popq   %rbp
	retq
	jmp    0x177b2d7e                 ; <+510> at ColumnNullable.cpp
	jmp    0x177b2d7e                 ; <+510> at ColumnNullable.cpp
	movq   %rax, %rbx
	lock
	decl   0x8(%r13)
	jne    0x177b2d96                 ; <+534> at intrusive_ref_counter.hpp
	movq   %r13, %rdi
	movq   0x8(%rsp), %rax
	callq  *0x2b8(%rax)
	movq   %rbx, %rdi
	callq  0x1d1b09c0                 ; _Unwind_Resume at UnwindLevel1.c:435

liuneng1994 · 2023-06-28T11:18:28Z

I tried both of the above modifications. None worked

baibaichen · 2023-06-30T14:35:34Z

@nickitat any suggestion？

liuneng1994 · 2023-07-04T02:22:34Z

@alexey-milovidov @nickitat Is this change acceptable or are there any other suggestions

nickitat · 2023-07-04T21:46:01Z

@alexey-milovidov @nickitat Is this change acceptable or are there any other suggestions

my view is the following. ideally we want to have batch version of serializeValueIntoArena that will serialize whole block and this way we will not pay for virtual call and have benefits from inlining.
unfortunately we cannot implement it this way, because we have special logic that saves us memory:

ClickHouse/src/Common/HashTable/HashTableKeyHolder.h

Lines 128 to 134 in dea5cbc

    
           inline void ALWAYS_INLINE keyHolderDiscardKey(DB::SerializedKeyHolder & holder) 
        
           { 
        
               [[maybe_unused]] void * new_head = holder.pool.rollback(holder.key.size); 
        
               assert(new_head == holder.key.data); 
        
               holder.key.data = nullptr; 
        
               holder.key.size = 0; 
        
           }

and for serialised aggregation saving memory is really important.
the only alternative I see is to do "devirtualisation" by hands. remember the type of column (e.g. string, fixed_string, vector or smth_else), inside ColumnNullable::serializeValueIntoArena do a static_cast to the actual column type (or fallback if type is smth_else) and call its serializeValueIntoArena. this way we won't have different serialisation algorithms for the same column type in different parts of code but still have non-virtual call and an inlining opportunity.
but it is not guaranteed to give big perf improvement...

baibaichen · 2023-07-05T01:45:43Z

@nickitat IIUC, we should change serializeValueIntoArena's signature so that we always choose nullable version of serializeValueIntoArena at ColumnNullable with static_cast

right?

liuneng1994 · 2023-07-06T05:37:08Z

@nickitat I re-modified a version and added a new parameter null_bit to serializeValueIntoArena to pass the null value of nullable, here is to be able to combine nullable and specific types of serialization algorithms together. For the string type, I use static_cast to avoid calling virtual functions. For other fixed-length types, I still use virtual functions to choose, because there are many fixed-length types, and it is meaningless to make judgments on each of them and static_cast. When the null_bit is NULL, continue with the original serialization algorithm.

There are still some problems with this change, some code can be reused, and generally it can guarantee that there is only one serialization algorithm for a type. If this solution is acceptable in general, I will further improve the code.

If the method of modifying the interface is unacceptable, another way is to add additional functions for specific types (String, Number), and then call the function in nullable. If this method is acceptable, I can also implement it.

nickitat · 2023-07-25T18:42:54Z

current solution looks good to me, let's continue

src/Columns/ColumnNullable.cpp

liuneng1994 · 2023-08-07T10:13:27Z

…lueIntoArena

baibaichen · 2023-08-10T02:30:38Z

@nickitat @alexey-milovidov

Any comments? would we merge this PR?

nickitat · 2023-08-15T16:38:30Z

@nickitat @alexey-milovidov

Any comments? would we merge this PR?

guys, I really apologies for delayed responding. busy with some other tasks. will take another look by the end of the week.

amosbird · 2023-10-19T08:21:38Z

I came across this PR when trying to traceback the mysterious "const UInt8 *" argument in serializeValueIntoArena.

The optimization is likely not related to devirtualization, but to a reduced number of allocations. I've proposed a cleaner implementation in #55809. The complexity is constrained inside ColumnNullable instead of leaking around. @nickitat Could you help take a look?

liuneng1994 marked this pull request as draft June 26, 2023 05:05

liuneng1994 changed the title ~~Optimize aggregation performance of nullable key when use AggregationMethodSerialized~~ Optimize aggregation performance of nullable String key when use AggregationMethodSerialized Jun 26, 2023

liuneng1994 marked this pull request as ready for review June 26, 2023 05:44

liuneng1994 force-pushed the optimize_nullable_aggragate_serialized_method branch from 71a1298 to b9280d5 Compare June 26, 2023 05:48

qoega added the can be tested Allows running workflows for external contributors label Jun 26, 2023

robot-clickhouse added the pr-performance Pull request with some performance improvements label Jun 26, 2023

liuneng1994 force-pushed the optimize_nullable_aggragate_serialized_method branch 2 times, most recently from d35d8b8 to abad35b Compare June 27, 2023 03:08

nickitat self-assigned this Jun 27, 2023

nickitat reviewed Jun 27, 2023

View reviewed changes

liuneng1994 requested a review from nickitat July 3, 2023 01:46

liuneng1994 closed this Jul 5, 2023

liuneng1994 reopened this Jul 5, 2023

liuneng1994 force-pushed the optimize_nullable_aggragate_serialized_method branch 2 times, most recently from ecf8c82 to 94124d3 Compare July 5, 2023 08:33

nickitat reviewed Jul 25, 2023

View reviewed changes

src/Columns/ColumnNullable.cpp Outdated Show resolved Hide resolved

liuneng1994 force-pushed the optimize_nullable_aggragate_serialized_method branch from 7e7ac09 to 0388dc3 Compare August 7, 2023 06:25

liuneng1994 requested a review from nickitat August 7, 2023 10:13

liuneng1994 added 11 commits August 8, 2023 13:38

optimize performance of nullable String And Number column serializeVa…

4f9920c

…lueIntoArena

optimize agg with multiple string key

43e0481

remove numbers optimization. It will decrease performance

035dbda

optimize fixed size column

f96b9b7

optimize conditions

62dffd0

change param name

81f0d17

another version

594b382

add default value for compile

2997fe0

add more test

f33367c

optimize

8a83301

fix include

65aeb05

liuneng1994 force-pushed the optimize_nullable_aggragate_serialized_method branch from af0ca7c to 65aeb05 Compare August 8, 2023 05:38

nickitat approved these changes Aug 16, 2023

View reviewed changes

robot-ch-test-poll4 merged commit 3aa9cb1 into ClickHouse:master Aug 16, 2023

baibaichen mentioned this pull request Aug 17, 2023

[GLUTEN-1632][CH]Daily Update Clickhouse Version (20230817) apache/gluten#2784

Merged

amosbird mentioned this pull request Oct 19, 2023

Improve the performance of serialized aggregation method when involving multiple [nullable] columns #55809

Merged

Conversation

liuneng1994 commented Jun 26, 2023 • edited by nickitat Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changelog category (leave one):

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Uh oh!

liuneng1994 commented Jun 26, 2023

Uh oh!

robot-clickhouse commented Jun 26, 2023 • edited by robot-clickhouse-ci-2 Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

liuneng1994 commented Jun 27, 2023

Uh oh!

nickitat Jun 27, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

liuneng1994 Jun 28, 2023

Choose a reason for hiding this comment

Uh oh!

liuneng1994 Jun 28, 2023

Choose a reason for hiding this comment

Uh oh!

liuneng1994 Jun 28, 2023

Choose a reason for hiding this comment

Uh oh!

nickitat Jun 28, 2023

Choose a reason for hiding this comment

Uh oh!

liuneng1994 Jun 28, 2023

Choose a reason for hiding this comment

Uh oh!

liuneng1994 Jun 29, 2023

Choose a reason for hiding this comment

Uh oh!

liuneng1994 commented Jun 28, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

liuneng1994 commented Jun 28, 2023

Uh oh!

baibaichen commented Jun 30, 2023

Uh oh!

liuneng1994 commented Jul 4, 2023

Uh oh!

nickitat commented Jul 4, 2023

Uh oh!

baibaichen commented Jul 5, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

liuneng1994 commented Jul 6, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nickitat commented Jul 25, 2023

Uh oh!

Uh oh!

liuneng1994 commented Aug 7, 2023

Uh oh!

baibaichen commented Aug 10, 2023

Uh oh!

nickitat commented Aug 15, 2023

Uh oh!

amosbird commented Oct 19, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

liuneng1994 commented Jun 26, 2023 •

edited by nickitat

Loading

robot-clickhouse commented Jun 26, 2023 •

edited by robot-clickhouse-ci-2

Loading

nickitat Jun 27, 2023 •

edited

Loading

liuneng1994 commented Jun 28, 2023 •

edited

Loading

baibaichen commented Jul 5, 2023 •

edited

Loading

liuneng1994 commented Jul 6, 2023 •

edited

Loading