performance.html

<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd">
<!--
Copyright 2004-2011 H2 Group. Multiple-Licensed under the H2 License, Version 1.0,
and under the Eclipse Public License, Version 1.0
(http://h2database.com/html/license.html).
Initial Developer: H2 Group
-->
<html xmlns="http://www.w3.org/1999/xhtml" lang="en" xml:lang="en">
<head><meta http-equiv="Content-Type" content="text/html;charset=utf-8" /><title>
Performance
</title><link rel="stylesheet" type="text/css" href="stylesheet.css" />
<!-- [search] { -->
<script type="text/javascript" src="navigation.js"></script>
</head><body onload="frameMe();">
<table class="content"><tr class="content"><td class="content"><div class="contentDiv">
<!-- } -->

<h1>Performance</h1>
<a href="#performance_comparison">
    Performance Comparison</a><br />
<a href="#poleposition_benchmark">
    PolePosition Benchmark</a><br />
<a href="#database_performance_tuning">
    Database Performance Tuning</a><br />
<a href="#built_in_profiler">
    Using the Built-In Profiler</a><br />
<a href="#application_profiling">
    Application Profiling</a><br />
<a href="#database_profiling">
    Database Profiling</a><br />
<a href="#explain_plan">
    Statement Execution Plans</a><br />
<a href="#fast_import">
    Fast Database Import</a><br />

<h2 id="performance_comparison">Performance Comparison</h2>
<p>
In 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.
</p>

<h3>Embedded</h3>
<table class="main">
<tr><th>Test Case</th><th>Unit</th><th>H2</th><th>HSQLDB</th><th>Derby</th></tr>
<tr><td>Simple: Init</td><td>ms</td><td>241</td><td>431</td><td>1027</td></tr>
<tr><td>Simple: Query (random)</td><td>ms</td><td>193</td><td>267</td><td>748</td></tr>
<tr><td>Simple: Query (sequential)</td><td>ms</td><td>89</td><td>179</td><td>658</td></tr>
<tr><td>Simple: Update (random)</td><td>ms</td><td>406</td><td>772</td><td>12175</td></tr>
<tr><td>Simple: Delete (sequential)</td><td>ms</td><td>155</td><td>266</td><td>6281</td></tr>
<tr><td>Simple: Memory Usage</td><td>MB</td><td>7</td><td>13</td><td>16</td></tr>
<tr><td>BenchA: Init</td><td>ms</td><td>200</td><td>251</td><td>1075</td></tr>
<tr><td>BenchA: Transactions</td><td>ms</td><td>1071</td><td>1458</td><td>8142</td></tr>
<tr><td>BenchA: Memory Usage</td><td>MB</td><td>8</td><td>14</td><td>12</td></tr>
<tr><td>BenchB: Init</td><td>ms</td><td>787</td><td>1584</td><td>4163</td></tr>
<tr><td>BenchB: Transactions</td><td>ms</td><td>465</td><td>875</td><td>2744</td></tr>
<tr><td>BenchB: Memory Usage</td><td>MB</td><td>17</td><td>13</td><td>10</td></tr>
<tr><td>BenchC: Init</td><td>ms</td><td>348</td><td>225</td><td>922</td></tr>
<tr><td>BenchC: Transactions</td><td>ms</td><td>1382</td><td>865</td><td>3527</td></tr>
<tr><td>BenchC: Memory Usage</td><td>MB</td><td>12</td><td>20</td><td>11</td></tr>
<tr><td>Executed statements</td><td>#</td><td>322929</td><td>322929</td><td>322929</td></tr>
<tr><td>Total time</td><td>ms</td><td>5337</td><td>7173</td><td>41462</td></tr>
<tr><td>Statements per second</td><td>#</td><td>60507</td><td>45020</td><td>7788</td></tr>
</table>

<h3>Client-Server</h3>
<table class="main">
<tr><th>Test Case</th><th>Unit</th><th>H2</th><th>HSQLDB</th><th>Derby</th><th>PostgreSQL</th><th>MySQL</th></tr>
<tr><td>Simple: Init</td><td>ms</td><td>1715</td><td>2096</td><td>3008</td><td>3093</td><td>3084</td></tr>
<tr><td>Simple: Query (random)</td><td>ms</td><td>2615</td><td>2119</td><td>4450</td><td>3201</td><td>3313</td></tr>
<tr><td>Simple: Query (sequential)</td><td>ms</td><td>2531</td><td>1944</td><td>4019</td><td>3163</td><td>3295</td></tr>
<tr><td>Simple: Update (random)</td><td>ms</td><td>1862</td><td>2486</td><td>13929</td><td>4404</td><td>4391</td></tr>
<tr><td>Simple: Delete (sequential)</td><td>ms</td><td>778</td><td>1118</td><td>7032</td><td>1682</td><td>1882</td></tr>
<tr><td>Simple: Memory Usage</td><td>MB</td><td>8</td><td>14</td><td>18</td><td>1</td><td>2</td></tr>
<tr><td>BenchA: Init</td><td>ms</td><td>1264</td><td>1686</td><td>2734</td><td>2867</td><td>3225</td></tr>
<tr><td>BenchA: Transactions</td><td>ms</td><td>5998</td><td>6829</td><td>14323</td><td>11491</td><td>10571</td></tr>
<tr><td>BenchA: Memory Usage</td><td>MB</td><td>9</td><td>18</td><td>14</td><td>1</td><td>2</td></tr>
<tr><td>BenchB: Init</td><td>ms</td><td>5571</td><td>7553</td><td>11636</td><td>12226</td><td>12553</td></tr>
<tr><td>BenchB: Transactions</td><td>ms</td><td>1931</td><td>3417</td><td>3435</td><td>2407</td><td>2149</td></tr>
<tr><td>BenchB: Memory Usage</td><td>MB</td><td>18</td><td>16</td><td>13</td><td>2</td><td>2</td></tr>
<tr><td>BenchC: Init</td><td>ms</td><td>1333</td><td>968</td><td>1769</td><td>1693</td><td>2645</td></tr>
<tr><td>BenchC: Transactions</td><td>ms</td><td>6508</td><td>4330</td><td>7910</td><td>7564</td><td>6108</td></tr>
<tr><td>BenchC: Memory Usage</td><td>MB</td><td>12</td><td>21</td><td>14</td><td>2</td><td>2</td></tr>
<tr><td>Executed statements</td><td>#</td><td>322929</td><td>322929</td><td>322929</td><td>322929</td><td>322929</td></tr>
<tr><td>Total time</td><td>ms</td><td>32106</td><td>34546</td><td>74245</td><td>53791</td><td>53216</td></tr>
<tr><td>Statements per second</td><td>#</td><td>10058</td><td>9347</td><td>4349</td><td>6003</td><td>6068</td></tr>
</table>

<h3>Benchmark Results and Comments</h3>

<h4>H2</h4>
<p>
Version 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.
</p>

<h4>HSQLDB</h4>
<p>
Version 2.0.0 was used for the test.
Cached tables are used in this test (<code>hsqldb.default_table_type=cached</code>),
and the write delay is 1 second (<code>SET WRITE_DELAY 1</code>).
</p>

<h4>Derby</h4>
<p>
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 <code>FileChannel.force(false)</code>, but only twice per log file (not on each commit).
Disabling this call improves performance for Derby by about 2%.
Unlike H2, Derby does not call <code>FileDescriptor.sync()</code> on each checkpoint.
Derby supports a testing mode (system property <code>derby.system.durability=test</code>)
where durability is disabled. According to the documentation, this setting should be used for testing only,
as the database may not recover after a crash. Enabling this setting improves performance
by a factor of 2.6 (embedded mode) or 1.4 (server mode). Even if enabled, Derby is still less
than half as fast as H2 in default mode.
</p>

<h4>PostgreSQL</h4>
<p>
Version 8.4.4 was used for the test.
The following options where changed in <code>postgresql.conf:
fsync = off, commit_delay = 1000</code>.
PostgreSQL is run in server mode.
The memory usage number is incorrect, because only the memory usage of the JDBC driver is measured.
</p>

<h4>MySQL</h4>
<p>
Version 5.1.47-community was used for the test.
MySQL was run with the InnoDB backend.
The setting <code>innodb_flush_log_at_trx_commit</code>
(found in the <code>my.ini / my.cnf</code> file) was set to 0. Otherwise (and by default), MySQL is slow
(around 140 statements per second in this test) because it tries to flush the data to disk for each commit.
For small transactions (when autocommit is on) this is really slow.
But many use cases use small or relatively small transactions.
Too bad this setting is not listed in the configuration wizard,
and it always overwritten when using the wizard.
You need to change this setting manually in the file <code>my.ini / my.cnf</code>, and then restart the service.
The memory usage number is incorrect, because only the memory usage of the JDBC driver is measured.
</p>

<h4>Firebird</h4>
<p>
Firebird 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.
</p>

<h4>Why Oracle / MS SQL Server / DB2 are Not Listed</h4>
<p>
The 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.
</p>

<h3>About this Benchmark</h3>

<h4>How to Run</h4>
<p>
This test was as follows:
</p>
<pre>
build benchmark
</pre>

<h4>Separate Process per Database</h4>
<p>
For each database, a new process is started, to ensure the previous test does not impact
the current test.
</p>

<h4>Number of Connections</h4>
<p>
This is mostly a single-connection benchmark.
BenchB uses multiple connections; the other tests use one connection.
</p>

<h4>Real-World Tests</h4>
<p>
Good 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.
</p>

<h4>Comparing Embedded with Server Databases</h4>
<p>
This 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.
</p>

<h4>Test Platform</h4>
<p>
This 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.
</p>

<h4>Multiple Runs</h4>
<p>
When 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.
</p>

<h4>Memory Usage</h4>
<p>
It 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.
</p>

<h4>Delayed Operations</h4>
<p>
Some 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).
</p>

<h4>Transaction Commit / Durability</h4>
<p>
Durability means transaction committed to the database will not be lost.
Some databases (for example MySQL) try to enforce this by default by
calling <code>fsync()</code> to flush the buffers, but
most hard drives don't actually flush all data. Calling the method slows down transaction commit a lot,
but doesn't always make data durable. When comparing the results, it is important to
think about the effect. Many database suggest to 'batch' operations when possible.
This benchmark switches off autocommit when loading the data, and calls commit after each 1000
inserts. However many applications need 'short' transactions at runtime (a commit after each update).
This benchmark commits after each update / delete in the simple benchmark, and after each
business transaction in the other benchmarks. For databases that support delayed commits,
a delay of one second is used.
</p>

<h4>Using Prepared Statements</h4>
<p>
Wherever possible, the test cases use prepared statements.
</p>

<h4>Currently Not Tested: Startup Time</h4>
<p>
The 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.
</p>

<h2 id="poleposition_benchmark">PolePosition Benchmark</h2>
<p>
The PolePosition is an open source benchmark. The algorithms are all quite simple.
It was developed / sponsored by db4o.
This test was not run for a longer time, so please be aware that the results below
are for older database versions (H2 version 1.1, HSQLDB 1.8, Java 1.4).
</p>
<table class="main">
<tr><th>Test Case</th><th>Unit</th><th>H2</th><th>HSQLDB</th><th>MySQL</th></tr>
<tr><td>Melbourne write</td><td>ms</td><td>369</td><td>249</td><td>2022</td></tr>
<tr><td>Melbourne read</td><td>ms</td><td>47</td><td>49</td><td>93</td></tr>
<tr><td>Melbourne read_hot</td><td>ms</td><td>24</td><td>43</td><td>95</td></tr>
<tr><td>Melbourne delete</td><td>ms</td><td>147</td><td>133</td><td>176</td></tr>
<tr><td>Sepang write</td><td>ms</td><td>965</td><td>1201</td><td>3213</td></tr>
<tr><td>Sepang read</td><td>ms</td><td>765</td><td>948</td><td>3455</td></tr>
<tr><td>Sepang read_hot</td><td>ms</td><td>789</td><td>859</td><td>3563</td></tr>
<tr><td>Sepang delete</td><td>ms</td><td>1384</td><td>1596</td><td>6214</td></tr>
<tr><td>Bahrain write</td><td>ms</td><td>1186</td><td>1387</td><td>6904</td></tr>
<tr><td>Bahrain query_indexed_string</td><td>ms</td><td>336</td><td>170</td><td>693</td></tr>
<tr><td>Bahrain query_string</td><td>ms</td><td>18064</td><td>39703</td><td>41243</td></tr>
<tr><td>Bahrain query_indexed_int</td><td>ms</td><td>104</td><td>134</td><td>678</td></tr>
<tr><td>Bahrain update</td><td>ms</td><td>191</td><td>87</td><td>159</td></tr>
<tr><td>Bahrain delete</td><td>ms</td><td>1215</td><td>729</td><td>6812</td></tr>
<tr><td>Imola retrieve</td><td>ms</td><td>198</td><td>194</td><td>4036</td></tr>
<tr><td>Barcelona write</td><td>ms</td><td>413</td><td>832</td><td>3191</td></tr>
<tr><td>Barcelona read</td><td>ms</td><td>119</td><td>160</td><td>1177</td></tr>
<tr><td>Barcelona query</td><td>ms</td><td>20</td><td>5169</td><td>101</td></tr>
<tr><td>Barcelona delete</td><td>ms</td><td>388</td><td>319</td><td>3287</td></tr>
<tr><td>Total</td><td>ms</td><td>26724</td><td>53962</td><td>87112</td></tr>
</table>
<p>
There are a few problems with the PolePosition test:
</p>
<ul><li>
HSQLDB uses in-memory tables by default while H2 uses persistent tables. The HSQLDB version
included in PolePosition does not support changing this, so you need to replace
<code>poleposition-0.20/lib/hsqldb.jar</code> with a newer version (for example
<code>hsqldb-1.8.0.7.jar</code>),
and then use the setting
<code>hsqldb.connecturl=jdbc:hsqldb:file:data/hsqldb/dbbench2;hsqldb.default_table_type=cached;sql.enforce_size=true</code>
in the file <code>Jdbc.properties</code>.
</li><li>HSQLDB keeps the database open between tests, while H2 closes the database (losing all the cache).
To change that, use the database URL <code>jdbc:h2:file:data/h2/dbbench;DB_CLOSE_DELAY=-1</code>
</li><li>The amount of cache memory is quite important, specially for the PolePosition test.
Unfortunately, the PolePosition test does not take this into account.
</li></ul>

<h2 id="database_performance_tuning">Database Performance Tuning</h2>

<h3>Keep Connections Open or Use a Connection Pool</h3>
<p>
If your application opens and closes connections a lot (for example, for each request),
you should consider using a  <a href="tutorial.html#connection_pool">connection pool</a>.
Opening a connection using <code>DriverManager.getConnection</code> is specially slow
if the database is closed. By default the database is closed if the last connection is closed.
</p><p>
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
<a href="features.html#closing_a_database">Closing a database</a>.
</p>

<h3>Use a Modern JVM</h3>
<p>
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 <code>-server</code> command-line
option improves performance at the cost of a slight increase in start-up time.
</p>

<h3>Virus Scanners</h3>
<p>
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 <code>.db</code> are not scanned.
</p>

<h3>Using the Trace Options</h3>
<p>
If 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 <a href="features.html#trace_options">Using the Trace Options</a>.
</p>

<h3>Index Usage</h3>
<p>
This database uses indexes to improve the performance of
<code>SELECT, UPDATE, DELETE</code>.
If a column is used in the <code>WHERE</code> clause of a query, and if an index exists on this column,
then the index can be used. Multi-column indexes are used if all or the first columns of the index are used.
Both equality lookup and range scans are supported.
Indexes are used to order result sets, but only if the condition uses the same index or no index at all.
The results are sorted in memory if required.
Indexes are created automatically for primary key and unique constraints.
Indexes are also created for foreign key constraints, if required.
For other columns, indexes need to be created manually using the <code>CREATE INDEX</code> statement.
</p>

<h3>How Data is Stored Internally</h3>
<p>
For persistent databases, if a table is created with a single column primary key of type <code>BIGINT, INT, SMALLINT, TINYINT</code>,
then the data of the table is organized in this way. This is sometimes also called a "clustered index" or
"index organized table".
</p><p>
H2 internally stores table data and indexes in the form of b-trees.
Each b-tree stores entries as a list of unique keys (one or more columns) and data (zero or more columns).
The table data is always organized in the form of a "data b-tree" with a single column key of type <code>long</code>.
If a single column primary key of type <code>BIGINT, INT, SMALLINT, TINYINT</code> is specified when creating the table
(or just after creating the table, but before inserting any rows),
then this column is used as the key of the data b-tree.
If no primary key has been specified, if the primary key column is of another data type,
or if the primary key contains more than one column,
then a hidden auto-increment column of type <code>BIGINT</code> is added to the table,
which is used as the key for the data b-tree.
All other columns of the table are stored within the data area of this data b-tree
(except for large <code>BLOB, CLOB</code> columns, which are stored externally).
</p><p>
For each additional index, one new "index b-tree" is created. The key of this b-tree consists of the indexed columns,
plus the key of the data b-tree. If a primary key is created after the table has been created, or if the primary key
contains multiple column, or if the primary key is not of the data types listed above, then the primary key
is stored in a new index b-tree.
</p>

<h3>Optimizer</h3>
<p>
This 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.
</p>

<h3>Expression Optimization</h3>
<p>
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 <code>WHERE</code> clause is always false, then the table is not accessed at all.
</p>

<h3>COUNT(*) Optimization</h3>
<p>
If the query only counts all rows of a table, then the data is not accessed.
However, this is only possible if no <code>WHERE</code> clause is used, that means it only works for
queries of the form <code>SELECT COUNT(*) FROM table</code>.
</p>

<h3>Updating Optimizer Statistics / Column Selectivity</h3>
<p>
When executing a query, at most one index per join can be used.
If the same table is joined multiple times, for each join only one index is used
(the same index could be used for both joins, or each join could use a different index).
Example: for the query
<code>SELECT * FROM TEST T1, TEST T2 WHERE T1.NAME='A' AND T2.ID=T1.ID</code>,
two index can be used, in this case the index on NAME for T1 and the index on ID for T2.
</p><p>
If a table has multiple indexes, sometimes more than one index could be used.
Example: if there is a table <code>TEST(ID, NAME, FIRSTNAME)</code> and an index on each column,
then two indexes could be used for the query <code>SELECT * FROM TEST WHERE NAME='A' AND FIRSTNAME='B'</code>,
the index on NAME or the index on FIRSTNAME. It is not possible to use both indexes at the same time.
Which index is used depends on the selectivity of the column. The selectivity describes the 'uniqueness' of
values in a column. A selectivity of 100 means each value appears only once, and a selectivity of 1 means
the same value appears in many or most rows. For the query above, the index on NAME should be used
if the table contains more distinct names than first names.
</p><p>
The SQL statement <code>ANALYZE</code> can be used to automatically estimate the selectivity of the columns in the tables.
This command should be run from time to time to improve the query plans generated by the optimizer.
</p>

<h3>In-Memory (Hash) Indexes</h3>
<p>
Using in-memory indexes, specially in-memory hash indexes, can speed up
queries and data manipulation.
</p>
<p>In-memory indexes are automatically used
for in-memory databases, but can also be created for persistent databases
using <code>CREATE MEMORY TABLE</code>. In many cases,
the rows itself will also be kept in-memory. Please note this may cause memory
problems for large tables.
</p>
<p>
In-memory hash indexes are backed by a hash table and are usually faster than
regular indexes. However, hash indexes only supports direct lookup (<code>WHERE ID = ?</code>)
but not range scan (<code>WHERE ID &lt; ?</code>). To use hash indexes, use HASH as in:
<code>CREATE UNIQUE HASH INDEX</code> and
<code>CREATE TABLE ...(ID INT PRIMARY KEY HASH,...)</code>.
</p>

<h3>Use Prepared Statements</h3>
<p>
If possible, use prepared statements with parameters.
</p>

<h3>Prepared Statements and IN(...)</h3>
<p>
Avoid generating SQL statements with a variable size IN(...) list.
Instead, use a prepared statement with arrays as in the following example:
</p>
<pre>
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();
</pre>

<h3>Optimization Examples</h3>
<p>
See <code>src/test/org/h2/samples/optimizations.sql</code> for a few examples of queries
that benefit from special optimizations built into the database.
</p>

<h3>Cache Size and Type</h3>
<p>
By 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 <a href="features.html#cache_settings">Cache Settings</a>.
</p>

<h3>Data Types</h3>
<p>
Each data type has different storage and performance characteristics:
</p>
<ul><li>The <code>DECIMAL/NUMERIC</code> type is slower
    and requires more storage than the <code>REAL</code> and <code>DOUBLE</code> types.
</li><li>Text types are slower to read, write, and compare than numeric types and generally require more storage.
</li><li>See <a href="advanced.html#large_objects">Large Objects</a> for information on
    <code>BINARY</code> vs. <code>BLOB</code>
    and <code>VARCHAR</code> vs. <code>CLOB</code> performance.
</li><li>Parsing and formatting takes longer for the
    <code>TIME</code>, <code>DATE</code>, and
    <code>TIMESTAMP</code> types than the numeric types.
</li><li><code>SMALLINT/TINYINT/BOOLEAN</code> are not significantly smaller or faster
    to work with than <code>INTEGER</code> in most modes.
</li></ul>

<h3>Sorted Insert Optimization</h3>
<p>
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 <code>SORTED</code>
before the <code>SELECT</code> statement:
</p>
<pre>
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);
</pre>

<h2 id="built_in_profiler">Using the Built-In Profiler</h2>
<p>
A very simple Java profiler is built-in. To use it, use the following template:
</p>
<pre>
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));
</pre>

<h2 id="application_profiling">Application Profiling</h2>

<h3>Analyze First</h3>
<p>
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
<code>System.currentTimeMillis()</code>. But this does not work for complex applications with many modules, and for memory problems.
</p>
<p>
A simple way to profile an application is to use the built-in profiling tool of java. Example:
</p>
<pre>
java -Xrunhprof:cpu=samples,depth=16 com.acme.Test
</pre>
<p>
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 <code>jps -l</code> to get the process id, and then
run <code>jstack &lt;pid&gt;</code> or <code>kill -QUIT &lt;pid&gt;</code> (Linux) or press
Ctrl+C (Windows).
</p>
<p>
A simple profiling tool is included in H2. To use it, the application needs to be changed slightly. Example:
</p>
<pre>
import org.h2.util;
...
profiler = new Profiler();
profiler.startCollecting();
// application code
profiler.stopCollecting();
System.out.println(profiler.getTop(3));
</pre>
<p>
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.
</p>

<h2 id="database_profiling">Database Profiling</h2>
<p>
The <code>ConvertTraceFile</code> tool generates SQL statement statistics at the end of the SQL script file.
The format used is similar to the profiling data generated when using <code>java -Xrunhprof</code>.
For this to work, the trace level needs to be 2 or higher (<code>TRACE_LEVEL_FILE=2</code>).
The easiest way to set the trace level is to append the setting to the database URL, for example:
<code>jdbc:h2:~/test;TRACE_LEVEL_FILE=2</code> or <code>jdbc:h2:tcp://localhost/~/test;TRACE_LEVEL_FILE=2</code>.
As an example, execute the the following script using the H2 Console:
</p>
<pre>
SET TRACE_LEVEL_FILE 2;
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;
</pre>
<p>
After running the test case, convert the <code>.trace.db</code> file using the <code>ConvertTraceFile</code> tool.
The trace file is located in the same directory as the database file.
</p>
<pre>
java -cp h2*.jar org.h2.tools.ConvertTraceFile
    -traceFile "~/test.trace.db" -script "~/test.sql"
</pre>
<p>
The generated file <code>test.sql</code> will contain the SQL statements as well as the
following profiling data (results vary):
</p>
<pre>
-----------------------------------------
-- 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;
</pre>

<h2 id="explain_plan">Statement Execution Plans</h2>
<p>
The SQL statement <code>EXPLAIN</code> displays the indexes and optimizations the database uses for a statement.
The following statements support <code>EXPLAIN</code>: <code>SELECT, UPDATE, DELETE, MERGE, INSERT</code>.
The following query shows that the database uses the primary key index to search for rows:
</p>
<pre>
EXPLAIN SELECT * FROM TEST WHERE ID=1;
SELECT
    TEST.ID,
    TEST.NAME
FROM PUBLIC.TEST
    /* PUBLIC.PRIMARY_KEY_2: ID = 1 */
WHERE ID = 1
</pre>
<p>
For joins, the tables in the execution plan are sorted in the order they are processed.
The following query shows the database first processes the table <code>INVOICE</code> (using the primary key).
For each row, it will additionally check that the value of the column <code>AMOUNT</code> is larger than zero,
and for those rows the database will search in the table <code>CUSTOMER</code> (using the primary key).
The query plan contains some redundancy so it is a valid statement.
</p>
<pre>
CREATE TABLE CUSTOMER(ID IDENTITY, NAME VARCHAR);
CREATE TABLE INVOICE(ID IDENTITY,
    CUSTOMER_ID INT REFERENCES CUSTOMER(ID),
    AMOUNT NUMBER);

EXPLAIN SELECT I.ID, C.NAME FROM CUSTOMER C, INVOICE I
WHERE I.ID=10 AND AMOUNT>0 AND C.ID=I.CUSTOMER_ID;

SELECT
    I.ID,
    C.NAME
FROM PUBLIC.INVOICE I
    /* PUBLIC.PRIMARY_KEY_9: ID = 10 */
    /* WHERE (I.ID = 10)
        AND (AMOUNT > 0)
    */
INNER JOIN PUBLIC.CUSTOMER C
    /* PUBLIC.PRIMARY_KEY_5: ID = I.CUSTOMER_ID */
    ON 1=1
WHERE (C.ID = I.CUSTOMER_ID)
    AND ((I.ID = 10)
    AND (AMOUNT > 0))
</pre>

<h3>Displaying the Scan Count</h3>
<p>
<code>EXPLAIN ANALYZE</code> additionally shows the scanned rows per table and pages read from disk per table or index.
This will actually execute the query, unlike <code>EXPLAIN</code> which only prepares it.
The following query scanned 1000 rows, and to do that had to read 85 pages from the data area of the table.
Running the query twice will not list the pages read from disk, because they are now in the cache.
The <code>tableScan</code> means this query doesn't use an index.
</p>
<pre>
EXPLAIN ANALYZE SELECT * FROM TEST;
SELECT
    TEST.ID,
    TEST.NAME
FROM PUBLIC.TEST
    /* PUBLIC.TEST.tableScan */
    /* scanCount: 1000 */
/*
total: 85
TEST.TEST_DATA read: 85 (100%)
*/
</pre>

<h3>Special Optimizations</h3>
<p>
For certain queries, the database doesn't need to read all rows, or doesn't need to sort the result even if <code>ORDER BY</code> is used.
</p><p>
For queries of the form <code>SELECT COUNT(*), MIN(ID), MAX(ID) FROM TEST</code>, the query plan includes the line
<code>/* direct lookup */</code> if the data can be read from an index.
</p><p>
For queries of the form <code>SELECT DISTINCT CUSTOMER_ID FROM INVOICE</code>, the query plan includes the line
<code>/* distinct */</code> if there is an non-unique or multi-column index on this column, and if this column has a low selectivity.
</p><p>
For queries of the form <code>SELECT * FROM TEST ORDER BY ID</code>, the query plan includes the line
<code>/* index sorted */</code> to indicate there is no separate sorting required.
</p><p>
For queries of the form <code>SELECT * FROM TEST GROUP BY ID ORDER BY ID</code>, the query plan includes the line
<code>/* group sorted */</code> to indicate there is no separate sorting required.
</p>

<h2 id="fast_import">Fast Database Import</h2>
<p>
To speed up large imports, consider using the following options temporarily:
</p>
<ul><li><code>SET LOG 0</code> (disabling the transaction log)
</li><li><code>SET CACHE_SIZE</code> (a large cache is faster)
</li><li><code>SET LOCK_MODE 0</code> (disable locking)
</li><li><code>SET UNDO_LOG 0</code> (disable the session undo log)
</li></ul>
<p>
These options can be set in the database URL:
<code>jdbc:h2:~/test;LOG=0;CACHE_SIZE=65536;LOCK_MODE=0;UNDO_LOG=0</code>.
Most of those options are not recommended for regular use, that means you need to reset them after use.
</p>
<p>
If you have to import a lot of rows, use a PreparedStatement or use CSV import.
Please note that <code>CREATE TABLE(...) ... AS SELECT ...</code>
is faster than <code>CREATE TABLE(...); INSERT INTO ... SELECT ...</code>.
</p>

<!-- [close] { --></div></td></tr></table><!-- } --><!-- analytics --></body></html>