Lessons Learned in Building a Highly Scalable MySQL Database

94
Presented by, MySQL AB® & O’Reilly Media, Inc. Lessons Learned in Building a Highly Scalable MySQL Database Mariella Di Giacomo Daniel Chote Geoff Harrison The Hive

description

Lessons Learned in Building a Highly Scalable MySQL Database. Mariella Di Giacomo Daniel Chote Geoff Harrison The Hive. Outline. Introduction/Background Motivations, Project Requirements Architectural Design Proposed Solutions Performance Analysis Results MySQL and Other Optimizations. - PowerPoint PPT Presentation

Transcript of Lessons Learned in Building a Highly Scalable MySQL Database

Page 1: Lessons Learned in Building a Highly Scalable MySQL Database

Presented by,

MySQL AB® & O’Reilly Media, Inc.

Lessons Learned in Building a Highly Scalable

MySQL Database

Mariella Di Giacomo

Daniel Chote

Geoff Harrison

The Hive

Page 2: Lessons Learned in Building a Highly Scalable MySQL Database

Outline

Introduction/Background Motivations, Project Requirements Architectural Design Proposed Solutions Performance Analysis Results MySQL and Other Optimizations

Page 3: Lessons Learned in Building a Highly Scalable MySQL Database

The Hive

Started in 1998 (under a different name) Operations in 4 countries Developer-centric organization

Page 4: Lessons Learned in Building a Highly Scalable MySQL Database

The Hive

The Hive’s collective vision is coding and collaborating from anywhere.

The Hive has a number of large scale projects that include video conferencing, VOIP, IM, and virtual whiteboard all connected through smart self-managing project tools.

Page 5: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Installations at The Hive

70+ MySQL Server Installations Linux Operating System Mostly Commodity Hardware Used for production, batch processing,

replication and development

Page 6: Lessons Learned in Building a Highly Scalable MySQL Database

The Next Generation

“What we need now is a brilliant idea …”

Page 7: Lessons Learned in Building a Highly Scalable MySQL Database

Scalable Architecture

High Volume of users and access traffic Exponential growth of data access and storage Use of relatively commodity hardware

Page 8: Lessons Learned in Building a Highly Scalable MySQL Database

Architectural Requirements

Store data related to millions of users Store some of the data in an XML format Store and Retrieve the data through an Open

Source Relational Database Management System (RDBMS)

Page 9: Lessons Learned in Building a Highly Scalable MySQL Database

Architectural Requirements

Build a scalable architecture capable of handling content updates and data addition at fastest rate reasonably possible

A system that has as little service disruption as possible

Robustness, Flexibility, Scalability and Speed

Page 10: Lessons Learned in Building a Highly Scalable MySQL Database

A Scalable Architecture

Scalable, robust, fast and flexible.Capable of handling data related to millions of users.Based on a data layout that maps all the data for each user to its individual table, organizing the users across multiple systems, databases and filesystems. Consists of homogeneous systems.Each system consists of two physical computers that synchronize the data through Master Master MySQL Replication.

Page 11: Lessons Learned in Building a Highly Scalable MySQL Database

Hardware Architecture

The hardware architecture consists of:Pairs of AMD nodes (Opteron). Each node has 2 CPUs 2GHz with 2cores; it has 24 GB of RAM + 8 GB of swapEach computer system has two types of data storage devices: Serial Advanced Technology Attachment (SATA) and Solid State Drive (SSD). Each computer System is composed of Debian Linux Operating System (OS)

Page 12: Lessons Learned in Building a Highly Scalable MySQL Database

Software Architecture We benchmarked 4 filesystems (Ext3, JFS, ReiserFS

and XFS). We chose MySQL as the RDBMS for the following

reasons:Open SourceSpeedData Storage CapabilitiesDatabase Design Fault ToleranceSecurity

Page 13: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Database Design MySQL supports several storage engines. We chose MyISAM, for the following reasons:

Each table data is stored in a MYD fileEach table indices are stored in a MYI fileEach table file requires less disk space compared to other storage engines.Memory key buffer for the indexed data

Downsides of MyISAMGood for high volume of writes and/or read, but not both in high concurrency.

Page 14: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Database Design

Our design takes advantage of the fact that there will not be high level of concurrency on the same table, while high level of concurrency will be present at database and server level.

Table Structure Design:Reduce and or combine indices on some fields, to take fewer resources when adding and modifying dataUse of NOT NULL fields where possibleUse the most appropriate, efficient field type for a given field

Page 15: Lessons Learned in Building a Highly Scalable MySQL Database

Architecture

Page 16: Lessons Learned in Building a Highly Scalable MySQL Database

Why that architecture ?

Page 17: Lessons Learned in Building a Highly Scalable MySQL Database

Why that architecture ?

We had an idea ….. We validated the idea through benchmark

results and analysis We analyzed several MySQL Storage engines

(MyISAM, InnoDB, Maria, PBXT, Falcon) on 4 types of filesystems (Ext3, JFS, XFS, ReiserFS) using two system architectures (AMD and Intel), two types of data device technologies and integrated the benchmarks with some specific low-level optimizations.

Page 18: Lessons Learned in Building a Highly Scalable MySQL Database

Benchmarks

Benchmark Configuration Benchmark Hardware Architecture Benchmark Topics

Page 19: Lessons Learned in Building a Highly Scalable MySQL Database

Benchmark Configuration

The benchmark program is written in Perl. The same program is used for table creation, data insertion, update and selection.

The benchmark program spawns 100 parallel instances (clients) that execute Data Definition Language (DDL) and Data Manipulation Language (DML) statements against the MySQL server.

The statement response time has been taken on the server side and on the client side.

Page 20: Lessons Learned in Building a Highly Scalable MySQL Database

Benchmark Configuration

All benchmarks have been run several times for the same configuration.

All the computer systems benchmarked have been set on an isolated network to lower noise as much as possible.

All computer systems are composed entirely of Debian Linux (Lenny) running 2.6.23 and 2.6.24

All systems have been compiled with latest MySQL 5.1.x

Page 21: Lessons Learned in Building a Highly Scalable MySQL Database

Benchmark Hardware Architecture

The hardware architecture used for the benchmarks consists of:Pairs of AMD nodes (Opteron). Each node has 2 CPUs 2GHz with 2cores; 24 GB of RAM + 8 GB of swap2 Intel Xeon nodes. Each node has 2 CPUs 3GHz with 2cores and hyper-threading enabled (4 virtual cores each CPU); 24GB of RAM + 8 GB of swapEach system has identical SATA storage devices. The AMD nodes contain also two identical SSD drives that are raided via an add-on raid controller.

Page 22: Lessons Learned in Building a Highly Scalable MySQL Database

Benchmark Hardware ArchitectureEach system always uses two dedicated storage devices for database benchmarks.

Page 23: Lessons Learned in Building a Highly Scalable MySQL Database

Benchmark Topics

Data LayoutOrganize each user’s data in a separate tablePartition a table to contain the data associated to a set of users, using MySQL horizontal partitioning.

Database Organization20 and 100 databases with 2 million tables50 databases with 1 million tables

MySQL Storage Engines: MyISAM, InnoDB, Maria, Falcon, PBXT

Page 24: Lessons Learned in Building a Highly Scalable MySQL Database

Benchmark Topics

Filesystem OrganizationOne data storage device for filesystem. Each filesystem contains 1 million tablesExt3, JFS, ReiserFS and XFS

Data Storage Device TechnologySATASSD

Page 25: Lessons Learned in Building a Highly Scalable MySQL Database

Data Layout Scenarios

Two data layout scenarios are evaluated:Organize each user’s data in a separate tablePartition a table to contain the data associated to a set of users, using MySQL horizontal partitioning. In this case the total number of tables required is the total number of users divided by the maximum number of partitions allowed on a partitioned table.

Page 26: Lessons Learned in Building a Highly Scalable MySQL Database

Data Layout Scenarios

Each Data Layout scenario has been benchmarked for:Dynamic Creation of 1 million to 2 million tables on one system. In the case of 1 million tables they would be physically located on only one filesystem (one hard drive). In the case of 2 million tables they would be spread across 100 or 20 databases. Each dedicated filesystem would be mapped to a physical storage device.Data Insertion and Update on the created tablesData Selection

Page 27: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation

Page 28: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation Configuration

Table creation has been executed using a stored procedure.

Each benchmark has been executed on an AMD and Intel system

Each benchmark has been evaluated with 4 distinct filesystem types.

The table creation benchmarks have been evaluated with write cache enabled, write back (WB) and write cache disabled, write through (WT) disk policies.

Page 29: Lessons Learned in Building a Highly Scalable MySQL Database

Data Insert, Update and Query

Page 30: Lessons Learned in Building a Highly Scalable MySQL Database

Inserts and Updates Configuration

The benchmark configuration for the data insert and update is equivalent to the one used for table creation.

Each benchmark has been executed only on AMD systems

Each benchmark has been evaluated with 4 distinct filesystem types.

The benchmarks have been executed using write cache enabled and write cache disabled.

Page 31: Lessons Learned in Building a Highly Scalable MySQL Database

Inserts and Updates Configuration

While table creation would perform better using write cache disabled, initial benchmarks with DML operations using write cache disabled would perform worst compared to the equivalent with write cache enabled.

Page 32: Lessons Learned in Building a Highly Scalable MySQL Database

Data Layout Scenarios

Two data layout scenarios are evaluated:Organize each user’s data in a separate table

SATA Device Technology•MyISAM•InnoDB•Falcon•Maria•PBXT

SSD Device TechnologyMyISAM

Page 33: Lessons Learned in Building a Highly Scalable MySQL Database

Data Layout Scenarios

Data layout scenarios:Organize each user’s data in a separate table

SATA Device Technology

Page 34: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SATA Benchmarks

Benchmarks shown in the coming graphs have been executed using the MyISAM Storage Engine

The MyISAM key_buffer is set to 4GB In each benchmark the system containing

MySQL server has two storage devices for the databases

Page 35: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation

These graphs show the benchmark results derived from processing 2 million tables.

Table creation of 2 million tables was spread across 20 databases. The databases were split across two filesystems. Each filesystem was on a unique SATA drive. (The drives were set up as single volumes through an add-on raid controller)

Page 36: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM using WB

Page 37: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM using WT

Page 38: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM using JFS

Page 39: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM with XFS

Page 40: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM with ReiserFS

Page 41: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM with Ext3

Page 42: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation with MyISAM

The benchmark results showed: When creating 2 million tables, approximately

36GB of space is used The highest processing rate was obtained using

the JFS filesystem on a volume (SATA via raid controller) with write cache disabled

The MyISAM key buffer did not play a major role The number of volumes would influence the

processing rate.

Page 43: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation with MyISAM

Between 1 million tables on one filesystem versus 2 million tables on two filesystems, we did not see any major impact in number of tables created per second (provided that we would not reduce the number of storage devices).

1 million tables would generate a total of 3 million files, 2 million tables would double the number.

Page 44: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SATA Benchmarks

Inserts and Updates on 1 million tables (50 databases)

Inserts and Updates on 2 million tables (20 and 100 databases)

Page 45: Lessons Learned in Building a Highly Scalable MySQL Database

Inserts and Updates

The next graphs show the results of DML operations on 1 million tables created using MyISAM storage engine with 4 types of filesystems and using write cache enabled.

While table creation would perform better using write cache disabled, initial benchmarks with DML operations using write cache disabled result in generating response times 10% higher compared to the equivalent with write cache enabled.

Page 46: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and Ext3 and WB

Page 47: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and JFS and WB

Page 48: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and XFS and WB

Page 49: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and ReiserFS and WB

Page 50: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SATA Benchmarks

Inserts and Updates on 1 million tables Inserts and Updates on 2 million tables

Benchmarks on DML operations on 2 million tables were stopped after a few tests due to their high response time, caused by heavy I/O operations on the volumes.

Page 51: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SATA Results

The benchmarks run using MySQL MyISAM storage engine on SATA device technology showed, after several optimizations, that the average response time of a DML operation would be higher than one second on an environment with 100 concurrent clients.

Page 52: Lessons Learned in Building a Highly Scalable MySQL Database

InnoDB and SATA Benchmarks

Table creation with 1 million tables

(50 databases) Table creation with 2 million tables

(20 databases) Inserts and Updates on 1 million tables Inserts and Updates on 2 million tables

Page 53: Lessons Learned in Building a Highly Scalable MySQL Database

Table creation

These next graphs show the benchmark results derived from processing 2 million InnoDB tables.

The benchmarks have shown:There is a memory overhead with the table dictionary (approximately 6GB of memory for all of the table dictionaries).There is an overhead on the disk. Approximately 4 times the space (in comparison to MyISAM) is used to store a table, whether using single or multiple table spaces, or a single file per table.

Page 54: Lessons Learned in Building a Highly Scalable MySQL Database

InnoDB on JFS

Page 55: Lessons Learned in Building a Highly Scalable MySQL Database

Inserts and Updates with InnoDB

The next graph shows the results obtained with 1 million tables using InnoDB

The number of databases is 50 The filesystem considered is JFS The data storage devices are SATA The InnoDB buffer is set to 6GB, but not limited

to it

Page 56: Lessons Learned in Building a Highly Scalable MySQL Database

InnoDB on JFS with WB

Page 57: Lessons Learned in Building a Highly Scalable MySQL Database

InnoDB and SATA Results

The benchmarks were evaluated only with one filesystem type, JFS. They were not extended to other types because of the InnoDB overheads (memory required for table dictionary and disk space).

Page 58: Lessons Learned in Building a Highly Scalable MySQL Database

Falcon and SATA Benchmarks

We evaluated a few benchmarks related to table creation and DML operations using Falcon as MySQL Storage engine.

The benchmarks were executed using 1 million tables.

Falcon storage engine was discarded because it is a relatively new storage engine and is not yet robust as MyISAM or InnoDB.

Page 59: Lessons Learned in Building a Highly Scalable MySQL Database

Maria and SATA Benchmarks

We evaluated a few benchmarks related to table creation and DML operations using Maria Storage engine.

The benchmarks were executed using 1 million tables.

Maria storage engine was discarded because it is a new storage engine and is not yet optimized as MyISAM.

Page 60: Lessons Learned in Building a Highly Scalable MySQL Database

Maria and SATA Benchmarks

One of the interesting features of Maria is the ability to be generate crash safe tables. Maria is a crash safe version of MyISAM.

It has all the main functionality of MyISAM engine, but includes recovery support, full logging, etc.

Page 61: Lessons Learned in Building a Highly Scalable MySQL Database

PBXT and SATA Benchmarks

We evaluated a few benchmarks related to table creation using PBXT Storage engine.

The benchmarks were executed using 1 million tables.

PBXT storage engine was discarded because it showed some limitations on the number of tables that could be allocated.

Page 62: Lessons Learned in Building a Highly Scalable MySQL Database

Data Layout Scenarios

Data layout scenarios:Organize each user’s data in a separate table

SSD Device Technology

Page 63: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SSD Benchmarks

1 million tables using 50 databases 2 million tables using 100 and 20 databases

Page 64: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation (WB)

Page 65: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and Ext3 and WB

Page 66: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and JFS and WB

Page 67: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and ReiserFS and WB

Page 68: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and XFS and WB

Page 69: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SSD Benchmarks

After having run the benchmarks for 1 million tables, we noticed that Ext3 provided the best performance. We expect this is due to the fact that it does less IO optimization when compared to other filesystems.

Consequently the following benchmarks of 2 million tables have been executed only using the Ext3 filesystem and two SSD storage devices.

Page 70: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SSD Benchmarks

1 million tables 2 million tables using 100 and 20 databases

Page 71: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation: Ext3 and WB (100)

Page 72: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and Ext3 and WB (100)

Page 73: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and Ext3 and WB (100)

Page 74: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and Ext3 and WB (20)

Page 75: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and Ext3 and WB (20)

Page 76: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SSD Results

We have evaluated 50 databases with 1 million of tables on a single SSD device, 20 and 100 databases with 2000000 of tables spread on two SSD devices.

The benchmarks showed that Ext3 with write cache enabled would perform better while doing data inserts, updates and selects with 100 concurrent clients.

Page 77: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM and SSD Results

The benchmarks showed also a quite interesting improvement in response time in the chosen architecture.

The noop I/O scheduler performed better compared to the others available on Linux I/O.

Page 78: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM SSD versus MyISAM SATA

The graph below compares DML statements with MyISAM using 100 databases with 1 million tables on SATA and SSD

Page 79: Lessons Learned in Building a Highly Scalable MySQL Database

Data Layout Scenarios

Initially two data layout scenarios are evaluated:Organize each individual user’s data in a separate tablePartition a table to contain the data associated to a set of users, using MySQL horizontal partitioning. In this case the total number of tables required is the total number of users divided by the maximum number of partitions allowed on a partitioned table.

SATA Device Technology•MyISAM•InnoDB

Page 80: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL and Horizontal Partitioning

The next slides discuss the results of 2 million tables with horizontal partitioning and both the MyISAM and InnoDB storage engines.

The type of partitioning used was “Range” partitioning based on the primary key, unsigned integer field.

The number of partitions investigated was 10 and 1024.

Page 81: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM with Horizontal Partitioning

The next graph shows the creation of 2 million tables using the MyISAM storage engine and horizontal partitioning

The initial number of partitions was set to 10. This implies that for each table will be created 2 extra files per partition, with a total of 22 files per table. In the case of 2 million users this would generate a total of 4.4 million files and 200,000 tables.

Page 82: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM with Horizontal Partitioning

We investigated also the scenario with a number of partitions of 1024.

With 1024 partitions per table, one table would generate 2050 files, so in the case of 2 millions users approximately 4 millions of files and approximately 1954 tables.

The graph shows the results of table creation using a number of partitions of 10.

Page 83: Lessons Learned in Building a Highly Scalable MySQL Database

Table Creation with WB

Page 84: Lessons Learned in Building a Highly Scalable MySQL Database

MyISAM with Horizontal Partitioning

The drawback of using MyISAM versus MyISAM with partitioning was:Pros

The results for using a number of 10 partitions were slightly lower compared to those of the first data layout scenario.

ConsHigher level of management compared to the first data layout scenario and still a high number of files to be managedComplexity and overhead when using partitioning with a number of partitions set to 1024Some limitations on the possibility of using sub partitioning to distribute segments of data for a specified user.

Page 85: Lessons Learned in Building a Highly Scalable MySQL Database

InnoDB with Horizontal Partitioning

InnoDB with partitioning was considered but was excluded for the following reasons:

Memory overhead due to table dictionary Disk Space overhead compared to MyISAM as

discussed earlier (4 times that of MyISAM).

Page 86: Lessons Learned in Building a Highly Scalable MySQL Database

Optimizations

Why Optimize ?

Get more Performance with same Hardware. As your data grows, Performance may degrade.

What Optimize ?

Hardware Architecture, Operating System Components and MySQL.

Where Optimize ?

Monitor your system and applications and watch for possible bottlenecks.

Page 87: Lessons Learned in Building a Highly Scalable MySQL Database

Optimizations

MySQL Optimizations

Other Optimizations

Page 88: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Optimization

Server Compilation (query cache disabled, fast mutexes enabled, etc.)

Server Configuration and Tuning (memory buffers, number of open tables, number of open files, concurrency, etc.)

Table structure (Database Design) and allocation

Query handling

Page 89: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Server Compilation

Compilation using appropriate optimizations for Linux (Debian) and the specific hardware architecture made improvements of 10% of speed over equivalent standard MySQL server binary.

Page 90: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Configuration

MySQL assumes little about your hardware system configuration.

Optimal size of the MyISAM key_buffer used to allocate indices in memory.

Number of open tables at a given point (it plays an important role when dealing with millions of tables).

Max number of temporary tables

Page 91: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Configuration

Number of connections per second, number of threads created per second, max number of connections

Thread concurrency Thread memory allocation Long query time

Page 92: Lessons Learned in Building a Highly Scalable MySQL Database

MySQL Query Handling

Use EXPLAIN SELECT to improve queries Use Indexes Simplify some of the WHERE clauses Use UNION of SELECT instead of only one

SELECT with several conditions in the WHERE clause

Disabled query cache, provided that the queries would not take advantage of that

Avoid Table Scans

Page 93: Lessons Learned in Building a Highly Scalable MySQL Database

Other Optimizations

File System tuning File System “mount” options tuning OS Linux Scheduler tuning OS Linux Kernel Swap tuning Tuning for Write-Heavy Loads

Page 94: Lessons Learned in Building a Highly Scalable MySQL Database

Q&A

Thanks for your attention! We would like to thank The Hive for making this

talk possible.