Performance of the TableFilter
- Basics: the sorting performance
- Minimum settings performance
- Parsing performance
- AutoChoices performance
- AutoChoices and performance on model updates
- Adaptive choices with enabled auto choices
This page tries to provide information on the overhead that the system experiences when using a table filter header. In some cases, the programmer will need to disable some of the features in the filter, in order to achieve a better, smoother experience.
The sections below contain specific memory requirements and measures of time to show the performance of the filter header on multiple scenarios. Those numbers are, however, specific to the example and to the machine where they have been executed, and their main value is to understand the relative performance as the library features are enabled or disabled.
The example used is, in most cases, the TableFilterExample class released with this library's source code, on its basic model, with 5 columns. The computer used is, explicitly, and old Pentium D820 (dual core), with 2 GB, from late 2006.
Note: The scenarios using 1 million of rows are executed with the JVM setting -Xmx1500M. (JVM is 32 bits)
The filter header is applied on top of the sorting functionality provided in swing. The sorting is done always on the GUI event thread, and, depending on the table size, it could be considered slow. Note that, being used the GUI event thread, the whole GUI freezes until the sorting finishes; although even 1 second affects negatively the application usability, the benefits of the associated action (in this case the sorting) surely offset the drawbacks. Subjectively, a couple of seconds already exceeds the maximum time that any action should last before the system produces a proper feedback (in this case, showing the sorted / filtered table).
The following list shows the required times to sort a table of several sizes. The test is done on the TableFilterExample application delivered with the library's source code, disabling the filter header (using therefore only the sorting capabilities of swing). The sorting is applied on a column containing strings, in average 12 characters long, being unique up to 10.000 of those strings; other types, like integers, would likely be sorted in shorter times:
1000 rows: 50 ms 10000 rows: 400 ms 100000 rows: 5500 ms 1000000 rows: 77819 ms
It requires approximately 1.5 seconds when the table has 30.000 rows, time long enough to start thinking 'Have I pressed the button?'. Sorting a table with a million rows is just an exercise on patience, the programmer should find alternative ways to present a reduced set of the information.
These measures do not imply that the swing JTable component cannot be used on tables larger than 30.000 rows, only that the user would benefit of a faster computer, and will likely experience a frozen GUI on specific actions (anyway the user will probably prefer this frozen GUI than lacking the possibility to sort these tables).
This discussion could now go into different directions, like proposing some sorting mechanisms invoked on separated threads, but this would add other interface headaches on its own (and would not't be compatible with existing Swing applications, as the table models should become thread-safe, which is currently not the case).
To observe now the overhead of the TableFilterHeader, the most basic scenario corresponds to the filter header with autoChoices and adaptiveChoices disabled:
These are the settings requiring less resources from the system: requires no added memory to hold the choices on each column, or computing time to extract and process those choices.
Anyway, the simple operation of associating a filter header to a table requires some memory and computing time:
A1- 10 rows: 0.05 Mb, 184 ms A2- 100 rows: 0.06 Mb, 187 ms A3- 1000 rows: 0.07 Mb, 190 ms A4- 10000 rows: 0.27 Mb, 189 ms A5- 100000 rows: 2.58 Mb, 242 ms A6- 1000000 rows: 23.18 Mb, 634 ms
However, this memory/time is not used directly by the filter header; instead, setting the filter header implies activating the sorting capabilities of the JTable, and this activation requires resources: the reason why memory increases with the number of rows lies on the internal structures used by the TableSorter to handle the table. Said otherwise: not using the filter header and triggering a sorting on any table's column will require the given computing time and most of the memory (some memory is definitely used by the filter header's components).
Obviously, the table model affects these times; for example, using the same example, with the model containing 8 columns instead of 5, the measures for 10.000 rows read as:
A5b- 100000 rows: 2.72 Mb, 333 ms (versus 2.58 Mb / 242 ms)
Computing time is definitely required when the user enters an expression on any of the filter editors. This time has two components: the filtering itself, plus its impact on the current sorting. The worse filter is, therefore, the filter that does not't filter out any rows, as the subsequent sorting will apply on the whole set of rows. Effectively, setting the content to '*' on a string column that is not sorted implies following filtering times:
A7- 10000 rows: 240 ms A8- 100000 rows: 837 ms A9- 1000000 rows: 2193 ms
If the column has no string type, filtering implies always a type conversion; how this happens depends on the parser being used. If the default library's parser is used, there are two possibilities:
- The expression introduced by the user is parsed, and the included expression converted into the type of the column. For example, if the user enters the expression '> 123' on an integer column, the text '123' is converted into the integer 123, and the filtering involves comparing integers, which should be a fast operation.
- The expression introduced by the user is handled as a string (ultimately a regular expression), and the filtering involves regular expression matching along the whole column's contents. If the column type is not a String, there must be a conversion to String on each column value before matching the regular expression. This operation is obviously slower than using the column's type.
The first option only happens when the user enters an operator, and the column type is a primitive or the programmer has provided valid Format and Comparator instances for the type. This is why, on a column with type Integer, it is faster to introduce as expression '= 111' than '111'. Nevertheless, in most cases, the time difference can be dismissed without bothering the user with these details.
Anyway, the parser is not used in this way on every case:
- If the filter editor is not editable, the filtering corresponds to always use the operand '='.
- If the user chooses a custom filter, the filtering is done through the filter of the custom choice, so no parser is involved.
The first feature to improve the filter header usability is to enable the auto choices (which happens by default): each filter editor shows all the possible choices, as extracted from the table.
In this case, there is already a significant overhead, mostly in time: the filter handler must extract all the unique values from the table, and, in most cases, convert them into strings; then, it must sort those strings (unless the column includes a renderer on the filter editor). The overhead depends therefore on the number of columns, the number of unique elements per column, the type associated to that column and the performance of the string conversion for that type.
So, in addition to the normal overhead of attaching the table to the filter header, it must be included the time/memory required to extract the choices on each column. Following measures apply for different scenarios, with a table with 100.000 rows:
A- String column: 9400 unique values: 220 ms. B- Date column: 16750 unique values: 379 ms. C- Integer column: 26 unique values: 20 ms.
Only a small percentage subset of this time goes into extracting values from the table: what is more important here is the number of unique elements per column.
To understand the performance on the whole table, the overall overhead is calculated by adding the time required to extract the autochoices on each column (where autochoices are enabled).
If auto choices are enabled, model updates imply some additional work on the filter header.
Adding a row to the model has little impact, but removing a row implies extracting the choices for all the columns, the same as updating a row. If a cell is updated, only the choices for that column are extracted.
The main problem on this behaviour lies on the sequencing of such operations. A GUI operation can be translated into several model updates: for example, removing 3 rows and updating another. In this case, the choices are extracted from the model ONCE for each suboperation; it can be more effective in this case to disable the filter header, perform the operations and the enable again the filter header: the choices are only extracted once.
The final setting with big impact on the performance is the adaptive choices flag. If enabled, as it is by default, choices are automatically updated to only display the available ones (an unavailable choice is a choice that would filter out all the rows).
It is not needed to activate the AutoChoices feature to have adaptive choices. In this case, the adaptive behaviour happens on the custom choices provided by the programmer. This section relates, however, to the most useful scenario, when auto choices are enabled.
Enabling adaptive choices requires more memory, as the filter header keeps, to improve performance, internal structures with the filtering state. On the positive side, these structures allow faster filtering times (checks to verify if a cell is filtered in or out are done exactly once). Next table shows the initialisation times and memory required for scenarios with adaptive choices enabled vs the same scenario without adaptive choices (both with auto choices enabled):
B1- 1000 rows: 0.22 Mb, 180 ms <-- versus --> 0.14 Mb, 220 ms. B2- 10000 rows: 1.37 Mb, 385 ms <-- versus --> 0.72 Mb, 735 ms B3- 100000 rows: 7.27 Mb, 755 ms <-- versus --> 3.63 Mb, 1709 ms B4- 1000000 rows: 58.77 Mb, 3100 ms <-- versus --> 23.87 Mb, 4332 ms.
Two results are obvious: required space doubles, and initialisation time is significantly decreased when adaptive choices is enabled; note that this example does not contain more than 10.000 unique values, which is the reason why the times does not increase much more on cases B3 and B4.
The negative performance impact from the adaptive choices lies on the extraction of the available choices that happen dynamically. A change on the filter expression of one editor requires extracting the available choices of the other editors, whose times have been already measured above.