VIEWS: 51 PAGES: 94 POSTED ON: 8/2/2009
Lessons Learned in Building a Highly Scalable MySQL Database Presented by, MySQL AB® & O’Reilly Media, Inc. 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 The Hive Started in 1998 (under a different name) Operations in 4 countries Developer-centric organization 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. MySQL Installations at The Hive 70+ MySQL Server Installations Linux Operating System Mostly Commodity Hardware Used for production, batch processing, replication and development The Next Generation “What we need now is a brilliant idea …” Scalable Architecture High Volume of users and access traffic Exponential growth of data access and storage Use of relatively commodity hardware 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) 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 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. 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 swap Each 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) Software Architecture We benchmarked 4 filesystems (Ext3, JFS, ReiserFS and XFS). We chose MySQL as the RDBMS for the following reasons: Open Source Speed Data Storage Capabilities Database Design Fault Tolerance Security MySQL Database Design MySQL supports several storage engines. We chose MyISAM, for the following reasons: Each table data is stored in a MYD file Each table indices are stored in a MYI file Each table file requires less disk space compared to other storage engines. Memory key buffer for the indexed data Downsides of MyISAM Good for high volume of writes and/or read, but not both in high concurrency. 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 data Use of NOT NULL fields where possible Use the most appropriate, efficient field type for a given field Architecture Why that architecture ? 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. Benchmarks Benchmark Configuration Benchmark Hardware Architecture Benchmark Topics 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. 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 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 swap 2 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 swap Each system has identical SATA storage devices. The AMD nodes contain also two identical SSD drives that are raided via an add-on raid controller. Benchmark Hardware Architecture Each system always uses two dedicated storage devices for database benchmarks. Benchmark Topics Data Layout Organize each user’s data in a separate table Partition a table to contain the data associated to a set of users, using MySQL horizontal partitioning. Database Organization 20 and 100 databases with 2 million tables 50 databases with 1 million tables MySQL Storage Engines: MyISAM, InnoDB, Maria, Falcon, PBXT Benchmark Topics Filesystem Organization One data storage device for filesystem. Each filesystem contains 1 million tables Ext3, JFS, ReiserFS and XFS Data Storage Device Technology SATA SSD Data Layout Scenarios Two data layout scenarios are evaluated: Organize each user’s data in a separate table Partition 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. 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 tables Data Selection Table Creation 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. Data Insert, Update and Query 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. 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. 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 Technology MyISAM Data Layout Scenarios Data layout scenarios: Organize each user’s data in a separate table SATA Device Technology 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 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) MyISAM using WB MyISAM using WT MyISAM using JFS MyISAM with XFS MyISAM with ReiserFS MyISAM with Ext3 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. 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. MyISAM and SATA Benchmarks Inserts and Updates on 1 million tables (50 databases) Inserts and Updates on 2 million tables (20 and 100 databases) 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. MyISAM and Ext3 and WB MyISAM and JFS and WB MyISAM and XFS and WB MyISAM and ReiserFS and WB 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. 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. 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 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. InnoDB on JFS 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 InnoDB on JFS with WB 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). 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. 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. 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. 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. Data Layout Scenarios Data layout scenarios: Organize each user’s data in a separate table SSD Device Technology MyISAM and SSD Benchmarks 1 million tables using 50 databases 2 million tables using 100 and 20 databases Table Creation (WB) MyISAM and Ext3 and WB MyISAM and JFS and WB MyISAM and ReiserFS and WB MyISAM and XFS and WB 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. MyISAM and SSD Benchmarks 1 million tables 2 million tables using 100 and 20 databases Table Creation: Ext3 and WB (100) MyISAM and Ext3 and WB (100) MyISAM and Ext3 and WB (100) MyISAM and Ext3 and WB (20) MyISAM and Ext3 and WB (20) 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. 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. MyISAM SSD versus MyISAM SATA The graph below compares DML statements with MyISAM using 100 databases with 1 million tables on SATA and SSD Data Layout Scenarios Initially two data layout scenarios are evaluated: Organize each individual user’s data in a separate table Partition 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 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. 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. 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. Table Creation with WB 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. Cons Higher level of management compared to the first data layout scenario and still a high number of files to be managed Complexity and overhead when using partitioning with a number of partitions set to 1024 Some limitations on the possibility of using sub partitioning to distribute segments of data for a specified user. 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). 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. Optimizations MySQL Optimizations Other Optimizations 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 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. 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 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 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 Other Optimizations File System tuning File System “mount” options tuning OS Linux Scheduler tuning OS Linux Kernel Swap tuning Tuning for Write-Heavy Loads Q&A Thanks for your attention! We would like to thank The Hive for making this talk possible.
"Lessons Learned in Building a Highly Scalable MySQL Database Presentation"