ClickHouse S3 Engine 数量级调优

原创

jasong

修改于 2022-03-08 21:05:57

8602

修改于 2022-03-08 21:05:57

文章被收录于专栏：ClickHouseClickHouse

本文主要讲解 ClickHouse S3 Engine 的读取写入性能代码及数量级调优

ClickHouse 如何性能调优

一前文

ClickHouse Lamdba

二 perf 调优

1 堆栈来源 trace_log

SELECT
    count(),
    arrayStringConcat(arrayMap(x -> concat(demangle(addressToSymbol(x)), '\n    ', addressToLine(x)), trace), '\n') AS sym
FROM system.trace_log
WHERE query_id = '157d5b6d-fa06-4bed-94f3-01f1d5f04e24'
GROUP BY trace
ORDER BY count() DESC
LIMIT 10

2 ClickHouse 线程

[root@9 ~]# ps H  -o 'tid comm' $(pidof -s clickhouse-server) | awk '{print $2}'  | sort | uniq -c | sort -rn
    230 ChunkParser
    199 QueryPipelineEx
     71 Segmentator
     66 BgSchPool
      9 clickhouse-serv
      3 TCPHandler
      3 SystemLogFlush
      3 Formatter
      2 ExterLdrReload
      2 ConfigReloader
      2 Collector

3 异常堆栈

被调用多次

    /root/ClickHouse/src/IO/BufferBase.h:82
void DB::readIntTextImpl<unsigned int, void, (DB::ReadIntTextCheckOverflow)0>(unsigned int&, DB::ReadBuffer&)
    /root/ClickHouse/src/IO/BufferBase.h:95
void DB::readIntText<(DB::ReadIntTextCheckOverflow)0, unsigned int>(unsigned int&, DB::ReadBuffer&)
    /root/ClickHouse/src/IO/ReadHelpers.h:393
std::__1::enable_if<is_integer_v<unsigned int>, void>::type DB::readText<unsigned int>(unsigned int&, DB::ReadBuffer&)
    /root/ClickHouse/src/IO/ReadHelpers.h:902
void DB::readCSVSimple<unsigned int>(unsigned int&, DB::ReadBuffer&)
    /root/ClickHouse/src/IO/ReadHelpers.h:988
std::__1::enable_if<is_arithmetic_v<unsigned int>, void>::type DB::readCSV<unsigned int>(unsigned int&, DB::ReadBuffer&)
    /root/ClickHouse/src/IO/ReadHelpers.h:994
DB::SerializationNumber<unsigned int>::deserializeTextCSV(DB::IColumn&, DB::ReadBuffer&, DB::FormatSettings const&) const
    /root/ClickHouse/src/DataTypes/Serializations/SerializationNumber.cpp:99
bool DB::SerializationNullable::deserializeTextCSVImpl<bool>(DB::IColumn&, DB::ReadBuffer&, DB::FormatSettings const&, std::__1::shared_ptr<DB::ISerialization const> const&)::'lambda'(DB::IColumn&)::operator()(DB::IColumn&) const
    /root/ClickHouse/src/DataTypes/Serializations/SerializationNullable.cpp:377
bool DB::safeDeserialize<bool, bool DB::SerializationNullable::deserializeTextCSVImpl<bool>(DB::IColumn&, DB::ReadBuffer&, DB::FormatSettings const&, std::__1::shared_ptr<DB::ISerialization const> const&)::'lambda'()&, bool DB::SerializationNullable::deserializeTextCSVImpl<bool>(DB::IColumn&, DB::ReadBuffer&, DB::FormatSettings const&, std::__1::shared_ptr<DB::ISerialization const> const&)::'lambda'(DB::IColumn&)&, (bool*)0>(DB::IColumn&, DB::ISerialization const&, bool DB::SerializationNullable::deserializeTextCSVImpl<bool>(DB::IColumn&, DB::ReadBuffer&, DB::FormatSettings const&, std::__1::shared_ptr<DB::ISerialization const> const&)::'lambda'(DB::IColumn&)&, (bool*)0&&)
    /root/ClickHouse/src/DataTypes/Serializations/SerializationNullable.cpp:193
bool DB::SerializationNullable::deserializeTextCSVImpl<bool>(DB::IColumn&, DB::ReadBuffer&, DB::FormatSettings const&, std::__1::shared_ptr<DB::ISerialization const> const&)
    /root/ClickHouse/src/DataTypes/Serializations/SerializationNullable.cpp:409
DB::CSVRowInputFormat::readField(DB::IColumn&, std::__1::shared_ptr<DB::IDataType const> const&, std::__1::shared_ptr<DB::ISerialization const> const&, bool)
    /root/ClickHouse/src/Processors/Formats/Impl/CSVRowInputFormat.cpp:404
DB::CSVRowInputFormat::readRow(std::__1::vector<COW<DB::IColumn>::mutable_ptr<DB::IColumn>, std::__1::allocator<COW<DB::IColumn>::mutable_ptr<DB::IColumn> > >&, DB::RowReadExtension&)
    /root/ClickHouse/src/Processors/Formats/Impl/CSVRowInputFormat.cpp:236
DB::IRowInputFormat::generate()

4 性能分析

三优化

1 对战调用分析

2 UML 类关系及调用顺序

- 优化方案调整 ReadFromS3 单线程读性能

3 调用堆栈

- ParallelParsingInputFormat.h

- ReadBuffer 解析

四 FileSegmentEngine

1 FileSegmentEngine Lambda

using FileSegmentationEngine = std::function<std::pair<bool, size_t>(
        ReadBuffer & buf,
        DB::Memory<Allocator<false>> & memory,
        size_t min_chunk_bytes)>;

2 ParallelParsingInputFormat.h

/// Function to segment the file. Then "parsers" will parse that segments.
    FormatFactory::FileSegmentationEngine file_segmentation_engine;

3 注册

void registerFileSegmentationEngineCSV(FormatFactory & factory)
{
    auto register_func = [&](const String & format_name, bool with_names, bool with_types)
    {
        size_t min_rows = 1 + int(with_names) + int(with_types);
        factory.registerFileSegmentationEngine(format_name, [min_rows](ReadBuffer & in, DB::Memory<> & memory, size_t min_chunk_size)
        {
            return fileSegmentationEngineCSVImpl(in, memory, min_chunk_size, min_rows);
        });
    };

    registerWithNamesAndTypes("CSV", register_func);
}

static std::pair<bool, size_t> fileSegmentationEngineCSVImpl(ReadBuffer & in, DB::Memory<> & memory, size_t min_chunk_size, size_t min_rows)
{
    char * pos = in.position();
    bool quotes = false;
    bool need_more_data = true;
    size_t number_of_rows = 0;

    while (loadAtPosition(in, memory, pos) && need_more_data)
    {
            else if (*pos == '\n')
            {
                ++number_of_rows;
                
                //min_chunk_size min_rows  成为了过滤的关键
                if (memory.size() + static_cast<size_t>(pos - in.position()) >= min_chunk_size && number_of_rows >= min_rows)
                    need_more_data = false;
                ++pos;
                if (loadAtPosition(in, memory, pos) && *pos == '\r')
                    ++pos;
            }
    }

    saveUpToPosition(in, memory, pos);
    return {loadAtPosition(in, memory, pos), number_of_rows};
}

五验收结果

初期优化后，因为使用DEBUG版本，Buffer 抽象非常高，处理非常的慢

预估单个线程 70MB/S

Debug mode

2022.02.16 10:22:02.465780 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 33, segment size 10485813 
2022.02.16 10:22:02.467560 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 33, segment size 10485813 ,time 1 ,Start Parser!
2022.02.16 10:24:00.756938 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 65505 , time 118291
2022.02.16 10:24:24.289329 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 26287 , time 141823
2022.02.16 10:24:24.289850 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 33, segment size 10485813 ,time 141824 ,Parser Finish!
2022.02.16 10:24:24.290132 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 33 ,segment size 10485813 use time 141824 


2022.02.16 10:25:18.579319 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 15, segment size 10485783 
2022.02.16 10:25:18.580013 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 15, segment size 10485783 ,time 0 ,Start Parser!
2022.02.16 10:25:56.587206 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 65505 , time 38007
2022.02.16 10:26:31.826019 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 26371 , time 73246
2022.02.16 10:26:31.826507 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 15, segment size 10485783 ,time 73247 ,Parser Finish!
2022.02.16 10:26:31.826747 [ 2903129 ] {8f97a908-1ba6-4e25-911b-dbd4ac947bf9} <Trace> ParallelParsingInputFormat: Read S3 object index 15 ,segment size 10485783 use time 73247

INFO Mode

2022.02.16 10:25:44.211472 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 8, segment size 10485834 
2022.02.16 10:25:44.211557 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 8, segment size 10485834 ,time 0 ,Start Parser!
2022.02.16 10:25:44.303112 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 65505 , time 91
2022.02.16 10:25:44.345420 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 26292 , time 133
2022.02.16 10:25:44.345469 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 8, segment size 10485834 ,time 134 ,Parser Finish!
2022.02.16 10:25:44.345490 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 8 ,segment size 10485834 use time 134 
2022.02.16 10:25:44.368619 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 17, segment size 10485789 
2022.02.16 10:25:44.368691 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 17, segment size 10485789 ,time 0 ,Start Parser!
2022.02.16 10:25:44.457678 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 65505 , time 89
2022.02.16 10:25:44.491163 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 26322 , time 122
2022.02.16 10:25:44.491205 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 17, segment size 10485789 ,time 122 ,Parser Finish!
2022.02.16 10:25:44.491227 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 17 ,segment size 10485789 use time 122 
2022.02.16 10:25:44.513499 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 26, segment size 10485847 
2022.02.16 10:25:44.513551 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 26, segment size 10485847 ,time 0 ,Start Parser!
2022.02.16 10:25:44.595114 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 65505 , time 81
2022.02.16 10:25:44.628066 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: chunk.getNumRows() 26318 , time 114
2022.02.16 10:25:44.628102 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 26, segment size 10485847 ,time 114 ,Parser Finish!
2022.02.16 10:25:44.628122 [ 323178 ] {61cbaecf-335d-4deb-baae-51f14cb153e5} <Trace> ParallelParsingInputFormat: Read S3 object index 26 ,segment size 10485847 use time 114

六性能

1 官方测试结果查询

2 优化后测试结果查询

3 优化写入

4 优化后写入

七测试数据

1 Schema

CREATE TABLE lineorder
(
    LO_ORDERKEY             UInt32,
    LO_LINENUMBER           UInt8,
    LO_CUSTKEY              UInt32,
    LO_PARTKEY              UInt32,
    LO_SUPPKEY              UInt32,
    LO_ORDERDATE            Date,
    LO_ORDERPRIORITY        LowCardinality(String),
    LO_SHIPPRIORITY         UInt8,
    LO_QUANTITY             UInt8,
    LO_EXTENDEDPRICE        UInt32,
    LO_ORDTOTALPRICE        UInt32,
    LO_DISCOUNT             UInt8,
    LO_REVENUE              UInt32,
    LO_SUPPLYCOST           UInt32,
    LO_TAX                  UInt8,
    LO_COMMITDATE           Date,
    LO_SHIPMODE             LowCardinality(String)
)
ENGINE = MergeTree PARTITION BY toYear(LO_ORDERDATE) ORDER BY (LO_ORDERDATE, LO_ORDERKEY);

2 单表测试模型

CREATE TABLE default.s3_1
(
    `LO_ORDERKEY` UInt32,
    `LO_LINENUMBER` UInt8,
    `LO_CUSTKEY` UInt32,
    `LO_PARTKEY` UInt32,
    `LO_SUPPKEY` UInt32,
    `LO_ORDERDATE` Date,
    `LO_ORDERPRIORITY` LowCardinality(String),
    `LO_SHIPPRIORITY` UInt8,
    `LO_QUANTITY` UInt8,
    `LO_EXTENDEDPRICE` UInt32,
    `LO_ORDTOTALPRICE` UInt32,
    `LO_DISCOUNT` UInt8,
    `LO_REVENUE` UInt32,
    `LO_SUPPLYCOST` UInt32,
    `LO_TAX` UInt8,
    `LO_COMMITDATE` Date,
    `LO_SHIPMODE` LowCardinality(String)
)
ENGINE = S3('http://xxx/insert01/s3_engine_1.csv', 'xxx', 'xxx', 'CSV')

3 多表测试模型

CREATE TABLE default.s3_5
(
    `LO_ORDERKEY` UInt32,
    `LO_LINENUMBER` UInt8,
    `LO_CUSTKEY` UInt32,
    `LO_PARTKEY` UInt32,
    `LO_SUPPKEY` UInt32,
    `LO_ORDERDATE` Date,
    `LO_ORDERPRIORITY` LowCardinality(String),
    `LO_SHIPPRIORITY` UInt8,
    `LO_QUANTITY` UInt8,
    `LO_EXTENDEDPRICE` UInt32,
    `LO_ORDTOTALPRICE` UInt32,
    `LO_DISCOUNT` UInt8,
    `LO_REVENUE` UInt32,
    `LO_SUPPLYCOST` UInt32,
    `LO_TAX` UInt8,
    `LO_COMMITDATE` Date,
    `LO_SHIPMODE` LowCardinality(String)
)
ENGINE = S3('http://xxx/insert01/s3_engine_{1..5}.csv', 'xxx', 'xxx', 'CSV')

4 腾讯云COS 规格与限制

5 网络代码测试结果

左侧为优化后网络性能基本可以打满 COS 带宽
右侧为优化前网络性能

6 不同数据量下网络带宽测试

希望能给学习ClickHouse 的同学带来帮助！

代码不太适合公布！欢迎大家使用腾讯云 ClickHouse ,感谢！

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

云数据仓库 ClickHouse

数据库

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

云数据仓库 ClickHouse

数据库

登录后参与评论

0 条评论

热度

ClickHouse S3 Engine 数量级调优

ClickHouse S3 Engine 数量级调优

一前文

二 perf 调优

1 堆栈来源 trace_log

2 ClickHouse 线程

3 异常堆栈

4 性能分析

三优化

1 对战调用分析

2 UML 类关系及调用顺序

3 调用堆栈

四 FileSegmentEngine

1 FileSegmentEngine Lambda

2 ParallelParsingInputFormat.h

3 注册

五验收结果

Debug mode

INFO Mode

六性能

1 官方测试结果查询

2 优化后测试结果查询

3 优化写入

4 优化后写入

七测试数据

1 Schema

2 单表测试模型

3 多表测试模型

4 腾讯云COS 规格与限制

5 网络代码测试结果

6 不同数据量下网络带宽测试

社区

活动

资源

关于

腾讯云开发者

热门产品

热门推荐

更多推荐

ClickHouse S3 Engine 数量级调优

ClickHouse S3 Engine 数量级调优

一 前文

二 perf 调优

1 堆栈来源 trace_log

2 ClickHouse 线程

3 异常堆栈

4 性能分析

三 优化

1 对战调用分析

2 UML 类关系及调用顺序

3 调用堆栈

四 FileSegmentEngine

1 FileSegmentEngine Lambda

2 ParallelParsingInputFormat.h

3 注册

五 验收结果

Debug mode

INFO Mode

六 性能

1 官方测试结果 查询

2 优化后测试结果 查询

3 优化 写入

4 优化后写入

七 测试数据

1 Schema

2 单表测试模型

3 多表测试模型

4 腾讯云COS 规格与限制

5 网络代码测试结果

6 不同数据量下 网络带宽测试

社区

活动

资源

关于

腾讯云开发者

热门产品

热门推荐

更多推荐

一前文

三优化

五验收结果

六性能

1 官方测试结果查询

2 优化后测试结果查询

3 优化写入

七测试数据

6 不同数据量下网络带宽测试