PerformancePerformance ComparisonPolePosition Benchmark Application Profiling Database Profiling Database Performance Tuning Using the Built-In Profiler Fast Database Import Performance ComparisonIn many cases H2 is faster than other (open source and not open source) database engines. Please note this is mostly a single connection benchmark run on one computer. Embedded
Client-Server
Benchmark Results and CommentsH2Version 1.2.137 (2010-06-06) was used for the test. For most operations, the performance of H2 is about the same as for HSQLDB. One situation where H2 is slow is large result sets, because they are buffered to disk if more than a certain number of records are returned. The advantage of buffering is: there is no limit on the result set size. The database open/close time is fixed. HSQLDB
Version 2.0.0 was used for the test.
Cached tables are used in this test ( A disadvantage of HSQLDB is the slow startup / shutdown time (currently not listed) when using bigger databases. The reason is, a backup of the whole data is made whenever the database is opened or closed. Derby
Version 10.6.1.0 was used for the test. Derby is clearly the slowest embedded database in this test.
This seems to be a structural problem, because all operations are really slow.
It will be hard for the developers of Derby to improve the performance to a reasonable level.
A few problems have been identified: leaving autocommit on is a problem for Derby.
If it is switched off during the whole test, the results are about 20% better for Derby.
Derby calls PostgreSQL
Version 8.4.4 was used for the test.
The following options where changed in MySQL
Version 5.1.47-community was used for the test.
MySQL was run with the InnoDB backend.
The setting FirebirdFirebird 1.5 (default installation) was tested, but the results are not published currently. It is possible to run the performance test with the Firebird database, and any information on how to configure Firebird for higher performance are welcome. Why Oracle / MS SQL Server / DB2 are Not ListedThe license of these databases does not allow to publish benchmark results. This doesn't mean that they are fast. They are in fact quite slow, and need a lot of memory. But you will need to test this yourself. SQLite was not tested because the JDBC driver doesn't support transactions. About this BenchmarkHow to RunThis test was as follows: build benchmark Separate Process per DatabaseFor each database, a new process is started, to ensure the previous test does not impact the current test. Number of ConnectionsThis is mostly a single-connection benchmark. BenchB uses multiple connections; the other tests use one connection. Real-World TestsGood benchmarks emulate real-world use cases. This benchmark includes 4 test cases: BenchSimple uses one table and many small updates / deletes. BenchA is similar to the TPC-A test, but single connection / single threaded (see also: www.tpc.org). BenchB is similar to the TPC-B test, using multiple connections (one thread per connection). BenchC is similar to the TPC-C test, but single connection / single threaded. Comparing Embedded with Server DatabasesThis is mainly a benchmark for embedded databases (where the application runs in the same virtual machine as the database engine). However MySQL and PostgreSQL are not Java databases and cannot be embedded into a Java application. For the Java databases, both embedded and server modes are tested. Test PlatformThis test is run on Mac OS X 10.6. No virus scanner was used, and disk indexing was disabled. The VM used is Sun JDK 1.6. Multiple RunsWhen a Java benchmark is run first, the code is not fully compiled and therefore runs slower than when running multiple times. A benchmark should always run the same test multiple times and ignore the first run(s). This benchmark runs three times, but only the last run is measured. Memory UsageIt is not enough to measure the time taken, the memory usage is important as well. Performance can be improved by using a bigger cache, but the amount of memory is limited. HSQLDB tables are kept fully in memory by default; this benchmark uses 'disk based' tables for all databases. Unfortunately, it is not so easy to calculate the memory usage of PostgreSQL and MySQL, because they run in a different process than the test. This benchmark currently does not print memory usage of those databases. Delayed OperationsSome databases delay some operations (for example flushing the buffers) until after the benchmark is run. This benchmark waits between each database tested, and each database runs in a different process (sequentially). Transaction Commit / Durability
Durability means transaction committed to the database will not be lost.
Some databases (for example MySQL) try to enforce this by default by
calling Using Prepared StatementsWherever possible, the test cases use prepared statements. Currently Not Tested: Startup TimeThe startup time of a database engine is important as well for embedded use. This time is not measured currently. Also, not tested is the time used to create a database and open an existing database. Here, one (wrapper) connection is opened at the start, and for each step a new connection is opened and then closed. PolePosition BenchmarkThe PolePosition is an open source benchmark. The algorithms are all quite simple. It was developed / sponsored by db4o.
There are a few problems with the PolePosition test:
Application ProfilingAnalyze First
Before trying to optimize performance, it is important to understand where the problem is (what part of the application is slow).
Blind optimization or optimization based on guesses should be avoided, because usually it is not an efficient strategy.
There are various ways to analyze an application. Sometimes two implementations can be compared using
A simple way to profile an application is to use the built-in profiling tool of java. Example: java -Xrunhprof:cpu=samples,depth=16 com.acme.Test
Unfortunately, it is only possible to profile the application from start to end. Another solution is to create
a number of full thread dumps. To do that, first run A simple profiling tool is included in H2. To use it, the application needs to be changed slightly. Example: import org.h2.util; ... profiler = new Profiler(); profiler.startCollecting(); // application code profiler.stopCollecting(); System.out.println(profiler.getTop(3)); The profiler is built into the H2 Console tool, to analyze databases that open slowly. To use it, run the H2 Console, and then click on 'Test Connection'. Afterwards, click on "Test successful" and you get the most common stack traces, which helps to find out why it took so long to connect. You will only get the stack traces if opening the database took more than a few seconds. Database Profiling
The SET TRACE_LEVEL_FILE 3; DROP TABLE IF EXISTS TEST; CREATE TABLE TEST(ID INT PRIMARY KEY, NAME VARCHAR(255)); @LOOP 1000 INSERT INTO TEST VALUES(?, ?); SET TRACE_LEVEL_FILE 0;
Now convert the
java -cp h2*.jar org.h2.tools.ConvertTraceFile
-traceFile "~/test.trace.db" -script "~/test.sql"
The generated file ----------------------------------------- -- SQL Statement Statistics -- time: total time in milliseconds (accumulated) -- count: how many times the statement ran -- result: total update count or row count ----------------------------------------- -- self accu time count result sql -- 62% 62% 158 1000 1000 INSERT INTO TEST VALUES(?, ?); -- 37% 100% 93 1 0 CREATE TABLE TEST(ID INT PRIMARY KEY, NAME VARCHAR(255)); -- 0% 100% 0 1 0 DROP TABLE IF EXISTS TEST; -- 0% 100% 0 1 0 SET TRACE_LEVEL_FILE 3; Database Performance TuningKeep Connections Open or Use a Connection Pool
If your application opens and closes connections a lot (for example, for each request),
you should consider using a connection pool.
Opening a connection using If you open and close connections a lot but don't want to use a connection pool, consider keeping a 'sentinel' connection open for as long as the application runs, or use delayed database closing. See also Closing a database. Use a Modern JVM
Newer JVMs are faster. Upgrading to the latest version of your JVM can provide a "free" boost to performance.
Switching from the default Client JVM to the Server JVM using the Virus Scanners
Some virus scanners scan files every time they are accessed.
It is very important for performance that database files are not scanned for viruses.
The database engine never interprets the data stored in the files as programs,
that means even if somebody would store a virus in a database file, this would
be harmless (when the virus does not run, it cannot spread).
Some virus scanners allow to exclude files by suffix. Ensure files ending with Using the Trace OptionsIf the performance hot spots are in the database engine, in many cases the performance can be optimized by creating additional indexes, or changing the schema. Sometimes the application does not directly generate the SQL statements, for example if an O/R mapping tool is used. To view the SQL statements and JDBC API calls, you can use the trace options. For more information, see Using the Trace Options. Index Usage
This database uses indexes to improve the performance of
OptimizerThis database uses a cost based optimizer. For simple and queries and queries with medium complexity (less than 7 tables in the join), the expected cost (running time) of all possible plans is calculated, and the plan with the lowest cost is used. For more complex queries, the algorithm first tries all possible combinations for the first few tables, and the remaining tables added using a greedy algorithm (this works well for most joins). Afterwards a genetic algorithm is used to test at most 2000 distinct plans. Only left-deep plans are evaluated. Expression Optimization
After the statement is parsed, all expressions are simplified automatically if possible. Operations
are evaluated only once if all parameters are constant. Functions are also optimized, but only
if the function is constant (always returns the same result for the same parameter values).
If the COUNT(*) Optimization
If the query only counts all rows of a table, then the data is not accessed.
However, this is only possible if no Updating Optimizer Statistics / Column Selectivity
When executing a query, at most one index per joined table can be used.
If the same table is joined multiple times, for each join only one index is used.
Example: for the query
If a table has multiple indexes, sometimes more than one index could be used.
Example: if there is a table
The SQL statement In-Memory (Hash) IndexesUsing in-memory indexes, specially in-memory hash indexes, can speed up queries and data manipulation. In-memory indexes are automatically used
for in-memory databases, but can also be created for persistent databases
using
In-memory hash indexes are backed by a hash table and are usually faster than
regular indexes. However, hash indexes only supports direct lookup ( Use Prepared StatementsIf possible, use prepared statements with parameters. Prepared Statements and IN(...)Avoid generating SQL statements with a variable size IN(...) list. Instead, use a prepared statement with arrays as in the following example:
PreparedStatement prep = conn.prepareStatement(
"SELECT * FROM TABLE(X INT=?) T INNER JOIN TEST ON T.X=TEST.ID");
prep.setObject(1, new Object[] { "1", "2" });
ResultSet rs = prep.executeQuery();
Optimization Examples
See Cache Size and TypeBy default the cache size of H2 is quite small. Consider using a larger cache size, or enable the second level soft reference cache. See also Cache Settings. Data TypesEach data type has different storage and performance characteristics:
Sorted Insert Optimization
To reduce disk space usage and speed up table creation, an
optimization for sorted inserts is available. When used, b-tree pages
are split at the insertion point. To use this optimization, add
CREATE TABLE TEST(ID INT PRIMARY KEY, NAME VARCHAR) AS
SORTED SELECT X, SPACE(100) FROM SYSTEM_RANGE(1, 100);
INSERT INTO TEST
SORTED SELECT X, SPACE(100) FROM SYSTEM_RANGE(101, 200);
Using the Built-In ProfilerA very simple Java profiler is built-in. To use it, use the following template: import org.h2.util.Profiler; Profiler prof = new Profiler(); prof.startCollecting(); // .... some long running process, at least a few seconds prof.stopCollecting(); System.out.println(prof.getTop(3)); Fast Database ImportTo speed up large imports, consider using the following options temporarily:
These options can be set in the database URL:
|