LinkBench is a database benchmark developed to evaluate database performance for workloads similar to those of Facebook's production MySQL deployment. LinkBench is highly configurable and extensible. It can be reconfigured to simulate a variety of workloads and plugins can be written for benchmarking additional database systems.
LinkBench is released under the Apache License, Version 2.0.
One way of modeling social network data is as a social graph, where entities or nodes such as people, posts, comments and pages are connected by links which model different relationships between the nodes. Different types of links can represent friendship between two users, a user liking another object, ownership of a post, or any relationship you like. These nodes and links carry metadata such as their type, timestamps and version numbers, along with arbitrary payload data.
Facebook represents much of its data in this way, with the data stored in MySQL databases. The goal of LinkBench is to emulate the social graph database workload and provide a realistic benchmark for database performance on social workloads. LinkBench's data model is based on the social graph, and LinkBench has the ability to generate a large synthetic social graph with key properties similar to the real graph. The workload of database operations is based on Facebook's production workload, and is also generated in such a way that key properties of the workload match the production workload.
++====================================++ || LinkBench Driver || ++====================================++ || +---------------------------+ || || | Graph | Workload | || Open connections +=======+ || | Generator | Generator | || /------------------> | Graph | || +---------------------------+ || /-------------------> | Store | || | |<======---------------------> | Shard | || | Graph Store Adapter |<======---------------------> | | || | (e.g. MySQL adapter) |<======---------------------> | e.g. | || +---------------------------+ || \-------------------> | MySQL | || || \------------------> | Server| || ~~~~~~~~~~~~ ~~~~~~~~~~~~ || +=======+ || ~~~~~~~~~~~~ ~~~~~~~~~~~~ || || ~~~~~~~~~~~~ ~~~~~~~~~~~~ || || ~~~~~~~~~~~~ ~~~~~~~~~~~~ || || Requester Threads || ++====================================++
The main software component of LinkBench is the driver, which acts as the client to the database being benchmarked. LinkBench is designed to support benchmarking of any database system that can support all of the require graph operations through a Graph Store Adapter.
The LinkBench benchmark typically proceeds in two phases.
The first is the load phase, where an initial graph is generated using the graph generator and loaded into the graph store in bulk. On a large benchmark run, this graph might have a billion nodes, and occupy over a terabyte on disk. The generated graph is designed to have similar properties to the Facebook social graph. For example, the number of links out from each node follows a power-law distribution, where most nodes have at most a few links, but a few nodes have many more links.
The second is the request phase, where the actual benchmarking occurs. In the request phase, the benchmark driver spawns many request threads, which make concurrent requests to the database. The workload generator is used by each request thread to generate a series of database operations that mimics the Facebook production workload in many aspects. For example, the mix of different varieties of read and write operations is the same, and the access patterns create a similar pattern of hot (frequently access) and cold nodes in the graph. At the end of the request phase LinkBench will report a range of statistics such as latency and throughput.
In this README we'll walk you through compiling LinkBench and running a MySQL benchmark.
These instructions assume you are using a UNIX-like system such as a Linux distribution or Mac OS X.
Java: You will need a Java 7+ runtime environment. LinkBench by default uses the version of Java on your path. You can override this by setting the JAVA_HOME environment variable to the directory of the desired Java runtime version. You will also need a Java JDK to compile from source.
Maven: To build LinkBench, you will need the Apache Maven build tool. If you do not have it already, it is available from http://maven.apache.org .
MySQL Connector: To benchmark MySQL with LinkBench, you need MySQL Connector/J, A version of the MySQL connector is bundled with LinkBench. If you wish to use a more recent version, replace the mysql jar under lib/. See http://dev.mysql.com/downloads/connector/j/
MySQL Server: To benchmark MySQL you will need a running MySQL server with free disk space.
First get the source code
git clone [email protected]:facebook/linkbench.git
Then enter the directory and build LinkBench
cd linkbench
mvn clean package
In order to skip slower tests (some run quite long), type
mvn clean package -P fast-test
To skip all tests
mvn clean package -DskipTests
If the build is successful, you should get a message like this at the end of the output:
BUILD SUCCESSFUL
Total time: 3 seconds
If the build fails while downloading required files, you may need to configure Maven, for example to use a proxy. Example Maven proxy configuration is shown here: http://maven.apache.org/guides/mini/guide-proxies.html
Now you can run the LinkBench command line tool:
./bin/linkbench
Running it without arguments will show a brief help message:
Did not select benchmark mode
usage: linkbench [-c <file>] [-csvstats <file>] [-csvstream <file>] [-D
<property=value>] [-L <file>] [-l] [-r]
-c <file> Linkbench config file
-csvstats,--csvstats <file> CSV stats output
-csvstream,--csvstream <file> CSV streaming stats output
-D <property=value> Override a config setting
-L <file> Log to this file
-l Execute loading stage of benchmark
-r Execute request stage of benchmark
In this section we will document the process of setting up a new MySQL database and running a benchmark with LinkBench.
We need to create a new database and tables on the MySQL server.
We'll create a new database called linkdb and
the needed tables to store graph nodes, links and link counts.
Run the following commands in the MySQL console:
create database linkdb;
use linkdb;
CREATE TABLE `linktable` (
`id1` bigint(20) unsigned NOT NULL DEFAULT '0',
`id2` bigint(20) unsigned NOT NULL DEFAULT '0',
`link_type` bigint(20) unsigned NOT NULL DEFAULT '0',
`visibility` tinyint(3) NOT NULL DEFAULT '0',
`data` varchar(255) NOT NULL DEFAULT '',
`time` bigint(20) unsigned NOT NULL DEFAULT '0',
`version` int(11) unsigned NOT NULL DEFAULT '0',
PRIMARY KEY (link_type, `id1`,`id2`),
KEY `id1_type` (`id1`,`link_type`,`visibility`,`time`,`id2`,`version`,`data`)
) ENGINE=InnoDB DEFAULT CHARSET=latin1 PARTITION BY key(id1) PARTITIONS 16;
CREATE TABLE `counttable` (
`id` bigint(20) unsigned NOT NULL DEFAULT '0',
`link_type` bigint(20) unsigned NOT NULL DEFAULT '0',
`count` int(10) unsigned NOT NULL DEFAULT '0',
`time` bigint(20) unsigned NOT NULL DEFAULT '0',
`version` bigint(20) unsigned NOT NULL DEFAULT '0',
PRIMARY KEY (`id`,`link_type`)
) ENGINE=InnoDB DEFAULT CHARSET=latin1;
CREATE TABLE `nodetable` (
`id` bigint(20) unsigned NOT NULL AUTO_INCREMENT,
`type` int(10) unsigned NOT NULL,
`version` bigint(20) unsigned NOT NULL,
`time` int(10) unsigned NOT NULL,
`data` mediumtext NOT NULL,
PRIMARY KEY(`id`)
) ENGINE=InnoDB DEFAULT CHARSET=latin1;
You may want to set up a special database user account for benchmarking:
-- Note: replace 'linkbench'@'localhost' with 'linkbench'@'%' to allow remote connections
CREATE USER 'linkbench'@'localhost' IDENTIFIED BY 'mypassword';
-- Grant all privileges on linkdb to this user
GRANT ALL ON linkdb TO 'linkbench'@'localhost'
If you want to obtain representative benchmark results, we highly recommend that you invest some time configuring and tuning MySQL. MySQL performance tuning can be complex and a comprehensive guide is beyond the scope of this readme, but here are a few basic guidelines:
- Read the Optimization section of the MySQL user manual.
- Make sure you have a sensible size setting for the InnoDB buffer pool size, so as to reduce disk I/O.
- Table partitioning (as shown above) can eliminate some bottlenecks that occur with LinkBench where the linktable is heavily accessed.
LinkBench requires several configuration files that specify the benchmark setup, the parameters of the graph to be generated, etc. Before benchmarking you will want to make a copy of the example config file:
cp config/LinkConfigMysql.properties config/MyConfig.properties
Open MyConfig.properties. At a minimum you will need to fill in the settings under MySQL Connection Information to match the server, user and database you set up earlier. E.g.
# MySQL connection information
host = localhost
user = linkbench
password = your_password
port = 3306
dbid = linkdb
You can read through the settings in this file. There are a lot of settings that control the benchmark itself, and the output of the LinkBench command link tool. Notice that MyConfig.properties references another file in this line:
workload_file = config/FBWorkload.properties
This workload file defines how the social graph should be generated and what mix of operations should make up the benchmark. The included workload file has been tuned to match our production workload in query mix. If you want to change the scale of the benchmark (the default graph is quite small for benchmarking purposes), you should look at the maxid1 setting. This controls the number of nodes in the initial graph created in the load phase: increase it to get a larger database.
# start node id (inclusive)
startid1 = 1
# end node id for initial load (exclusive)
# With default config and MySQL/InnoDB, 1M ids ~= 1GB
maxid1 = 10000001
First we need to do an initial load of data using our new config file:
./bin/linkbench -c config/MyConfig.properties -l
This will take a while to load, and you should get frequent progress updates. Once loading is finished you should see a notification like:
LOAD PHASE COMPLETED. Loaded 10000000 nodes (Expected 10000000).
Loaded 47423071 links (4.74 links per node). Took 620.4 seconds.
Links/second = 76435
At the end LinkBench reports a range of statistics on load time that are of limited interest at this stage.
You can significantly speed up the LinkBench load phase by making these temporary changes in the MySQL command shell before loading:
alter table linktable drop key `id1_type`;
set global innodb_flush_log_at_trx_commit = 2;
set global sync_binlog = 0;
After loading you should revert the changes:
set global innodb_flush_log_at_trx_commit = 1;
set global sync_binlog = 1;
alter table linktable add key `id1_type`
(`id1`,`link_type`,`visibility`,`time`,`version`,`data`);
Now you can do some benchmarking. Run the request phase using the below command:
./bin/linkbench -c config/MyConfig.properties -r
LinkBench will log progress to the console, along with statistics. Once all requests have been sent, or the time limit has elapsed, LinkBench will notify you of completion:
REQUEST PHASE COMPLETED. 25000000 requests done in 2266 seconds.
Requests/second = 11029
You can also inspect the latency statistics. For example, the following line tells us the mean latency for link range scan operations, along with latency ranges for median (p50), 99th percentile (p99) and so on.
GET_LINKS_LIST count = 12678653 p25 = [0.7,0.8]ms p50 = [1,2]ms
p75 = [1,2]ms p95 = [10,11]ms p99 = [15,16]ms
max = 2064.476ms mean = 2.427ms
Here are some further examples of how to use the LinkBench command link utility.
You can override any properties from the configuration file from the command line with -D key=value. For example, this runs the benchmark with a 10 minute warmup before collecting statistics:
./bin/linkbench -c config/MyConfig.properties -D warmup_time=600 -r
This runs the benchmark with more detailed logging, and all output going to the file linkbench.log:
./bin/linkbench -c config/MyConfig.properties -D debuglevel=DEBUG -L linkbench.log -r
LinkBench supports output of statistics in csv format for easier analysis. There are two categories of statistic: the final summary and per-thread statistics output periodically through the benchmark. -csvstats controls the former and -csvstream the latter:
./bin/linkbench -c config/MyConfig.properties -csvstats final-stats.csv -csvstreams streaming-stats.csv -r
Benchmarks are often controversial and are challenging to do well. Here are some guidelines for avoiding common pitfalls with LinkBench.
To remove confounding factors in database setup, there are several steps you can take to obtain better results:
- Warm up the databases before collecting statistics. LinkBench has a warmup_time setting that sends requests for a period before starting to collect statistics.
- Run benchmarks for long periods of time (hours rather than minutes) to reduce impact of random variation and to allow the database to reach a steady state.
- If at all possible, get expert help tuning the database for your hardware and workload.
- Benchmarks where the database fits mostly or entirely in RAM are interesting but aren't comparable to benchmarks where the database is much larger than RAM. Typically for MySQL benchmarks our databases are 10-15x larger than the buffer pool.
- Databases should be benchmarked in comparable configurations. We always run LinkBench with durable writes (i.e. so that after an operation returns, the data is written to persistent storage and can be recovered in the event of a system crash). Similarly, our LinkBench MySQL implementation provides serializable consistency of operations. Weaker durability or consistency properties should be disclosed alongside benchmark results.
Different systems can behave different when heavily or lightly loaded. The default benchmark settings simulate a heavily loaded database, with 100 concurrent request threads each sending requests as quickly as they can. Some database systems perform better than others with many concurrent clients or heavy load, so performance under heavy load does not give a complete picture of performance. Typically databases are not fully loaded all of the time, so latency of requests under moderate load is also an important measure of database performance.
To get a better understanding of database performance under varying load it can be helpful to:
- Modify the requesters parameter to test database performance with varying numbers of clients.
- Modify the requestrate config setting so that requests are throttled. Request latency vs. throughput curves help with understanding the full performance profile of a database system.
If you are doing a benchmark exercise, it is often a good idea to collect additional information about system resource utilization, particularly for CPU and I/O. This can aid a lot in understanding and comparing benchmark results beyond headline performance numbers. It is easiest to make use of collected data if you can match up timestamps to your benchmark logs, so the examples here will append timestamps to each line of output.
vmstat reports useful summary information on CPU and memory:
vmstat 1 | gawk '{now=strftime("%Y-%m-%d %T "); print now $0}' > linkbench.run.1/vmstat.out
iostat reports some useful I/O statistics:
iostat -d -x 1 | gawk '{now=strftime("%Y-%m-%d %T "); print now $0}' > linkbench.run.1/iostat.out
You can customize LinkBench in several ways.
We have already introduced you to the LinkBench configuration files. All settings in these files are documented and a great deal can be changed simply through these configuration files. For example:
- You can experiment with read-intensive or write-intensive workloads by modifying the mix of operations.
- You can alter the mix of hot/cold rows by modifying the shape parameter for ZipfDistribution. If you set it close to 1, there will be only a few very hot nodes in the database, or if you set if close to 0, accesses will be spread evenly across all nodes.
It is possible to further customize the data and workload by providing new implementations of some key classes:
- ProbabilityDistribution: which can be used to control the distribution of out-edges in the graph, or the access patterns for requests.
- DataGenerator: which can be used to generate data in different ways for requests.
You can write plugins to benchmark additional database systems
simply by writing a Java class implementing a small set of graph operations.
Any classes implementing the com.facebook.LinkBench.LinkStore
and com.facebook.LinkBench.NodeStore interfaces can be loaded
through the linkstore and nodestore configuration file keys.
There are several steps you will have to go through to add a new plugin . First you need to choose you will represent LinkBench nodes and links. Several factors play a role in the design, but speed of range scans and atomicity of updates are particularly important. The MySQL schema from earlier in this README serves as a reference implementation.
Next you need to create a new Java class, such as public class MyStore extends GraphStore, and implement all of the required methods of
LinkStore and NodeStore. Two reference implementations are provided:
LinkStoreMysql, a fully-fledged implementation, and MemoryLinkStore,
a toy in-memory implementation.
LinkBench provides some tests to validate your implementation that you
can use during development. If you extend any of the test classes
LinkStoreTestBase, NodeStoreTestBase and GraphStoreTestBase with
the required methods that set up your database, then a range of tests
will be run against it. These tests are sanity checks rather than
comprehensive verification of your implementation. In particular,
they do not try to verify the atomicity, consistency or durability
properties of the implementation.
Database-specific tests are not run by default. You can enable them with Maven profiles. For example, to run the MySQL tests you can run:
mvn test -P mysql-test
The MySQL related unit tests are run against a test database that needs setting up before running the unit tests. The default settings for this test database are hardcoded in src/test/java/com/facebook/LinkBench/MySqlTestConfig.java. The default settings uses localhost:3306 to connect to the database and uses username "linkbench" and password "linkbench". The unit test code creates all the required tables, so the developer needs to setup a MySql database called "linkbench_unittestdb" to which the linkbench user has permissions to create and drop tables.
If you implement a plugin for a new database, please consider contributing it back to the main LinkBench distribution with a pull request.