Memory allocated to the application server | Since the query engine retrieves the necessary information from persistent storage, the memory allocated to the Java Virtual Machine (usually specified by the -Xmx parameter) can be kept low. We recommend setting this figure to less than 1/3rd of the total memory available on the OS. We also advise to stay below 32 GB, which should fit all reasonable use cases, and allow benefiting from the compressed Oops feature. |
Memory allocated to the operating system | On the OS running the application server, it is important to leave sufficient room to the OS cache, letting it optimize access to the persistent Lucene indexes. Indeed, once these have been loaded from the file system, the OS uses its memory cache to speed up subsequent accesses to this same data, and avoid reloading it every time from the disk. This is only possible if sufficient RAM has been left for this purpose. It is also necessary to configure the OS so that the JVM process can reserve the resources required by numerous memory-mapped files (see this article for details). On a Linux OS, this can be done by issuing the following commands: ulimit -n 512000 sysctl vm.max_map_count=262144 |
Memory monitoring | Indications of EBX® load activity are provided by monitoring the underlying database, and also by the 'monitoring' logging category. If the numbers for cleared and built objects remain high for a long time, this is an indication that EBX® is swapping on the application server. In that case, the memory allocated to the application server should be increased. |
Garbage collector | Tuning the garbage collector can also benefit overall performance. This tuning should be adapted to the use case and specific Java Runtime Environment used. |
The EBX® repository data are indexed into Lucene indexes, stored on the disk under the root directory.
Disk space: a rule of thumb for the disk size is to plan for 10 times the space occupied by the table G_BLK in the relational database.
Disk latency: in order to maintain good overall performance, it is important for the disk storing the Lucene indexes to have low latency.
The number of CPUs available for the application server must be defined considering the number of concurrent HTTP requests to be served, the complexity (CPU cost) of the implied tasks, and the background activities, including the Java garbage collection.
Large imports and more generally large transactions involving many creates, updates and deletes will be completed faster if the difference between the server load and the number of available processors allows the indexing to be efficiently run in parallel within the transaction. The persistence log category contains the following entries:
'Setting forced synchronous indexing to false (...)' to indicate that the indexing will be performed concurrently;
'Setting forced synchronous indexing to true (...)' to indicate that the indexing will not be performed concurrently.
The difference mentioned above is assessed every ten seconds and is computed using the methods getSystemLoadAverage() and getAvailableProcessors() in the Java class java.lang.management.OperatingSystemMXBean. Both numbers are written at the end of the log entry above.
The LZ4 library is used to store data to and retrieve data from the database. To speed up data access, it is required to perform a ebx-lz4.jar native installation.
See Data compression library for more information.
To speed up the web applications server startup, the JAR files scanner should be configured.
As with any database, inserting and deleting large volumes of data may lead to fragmented data, which can deteriorate performance over time. To resolve the issue, reorganizing the impacted database tables is necessary. See Monitoring and cleanup of the relational database.
A specificity of EBX® is that creating dataspaces and snapshots adds new entries to tables GRS_DTR and GRS_SHR. When poor performance is experienced, it may be necessary to schedule a reorganization of these tables, for large repositories in which many dataspaces are created and deleted.
In a data model, when an element's cardinality constraint maxOccurs is greater than 1 and no osd:table is declared on this element, it is implemented as a Java List. This type of element is called an aggregated list, as opposed to a table.
It is important to consider that there is no specific optimization when accessing aggregated lists, in terms of iterations, user interface display, etc. Besides performance concerns, aggregated lists are limited with regard to many functionalities that are supported by tables. See tables introduction for a list of these features.
For the reasons stated above, aggregated lists should be used only for small volumes of simple data (one or two dozen records), with no advanced requirements for their identification, lookups, permissions, etc. For larger volumes of data (or more advanced functionalities), it is recommended to use osd:table declarations.
The internal validation framework will optimize the work required during successive requests to update the validation report of a dataset or a table. The incremental validation process behaves as follows:
The first call to a dataset or table validation report performs a full validation of the dataset or the table.
The next call to the validation report will compute the changes performed since the last validation. The validation report will be updated according to these changes.
Validation reports are stored persistently in the TIBCO EBX® repository. This reduces the amount of memory dedicated to validation reports when datasets have a large amount of validation messages. Also, validation reports are not lost when the application server restarts.
Validation reports can be reset manually in the user interface by an administrator user (this option is available from the validation report section in EBX®). As a consequence, resetting validation reports must be used with caution since associated datasets or tables will be fully revalidated during the next call to their validation reports.
See Adaptation.resetValidationReport for more information.
Certain constraints are systematically re-validated, even if no updates have occurred since the last validation. These are the constraints with unknown dependencies. An element has unknown dependencies if:
It specifies a programmatic constraint in the default unknown dependencies mode.
It declares a computed value, or it declares a dynamic facet that depends on an element that is itself a computed value.
It is an inherited field or it declares a dynamic facet that depends on a node that is itself an inherited field.
Consequently, on large tables, it is recommended to:
Avoid constraints with unknown dependencies (or at least to minimize the number of such constraints). For programmatic constraints, the developer is able to specify two alternative modes that drastically reduce incremental validation cost: local dependency mode and explicit dependencies. For more information, see Dependencies and validation.
To use constraints on tables instead of programmatic constraints defined at field level. Indeed, if a table defines constraints at field level, then the validation process will iterate over all the records to check if the value of the associated field complies with the constraint. Using constraints on tables gives the opportunity to execute optimized queries on the whole table.
Avoid the use of the facet pattern since its check is not optimized on large tables. That is, if a field defines this facet then the validation process will iterate over all the records to check if the value of the associated field complies with the specified pattern.
Tables are commonly accessed through EBX® UI, data services and also through the Request and Query APIs. This access involves a unique set of functions, including a dynamic resolution process. This process behaves as follows:
Inheritance: Inheritance in the dataset tree takes into account records and values that are defined in the parent dataset, using a recursive process. Also, in a root dataset, a record can inherit some of its values from the data model default values, defined by the xs:default attribute.
Value computation: A node declared as an osd:function is always computed on the fly when the value is accessed. See ValueFunction.getValue.
Filtering: An XPath predicate, a programmatic filter, or a record-level permission rule requires a selection of records.
Sort: A sort of the resulting records can be performed.
In order to improve the speed of operations on tables, persistent Lucene indexes are managed by the EBX® engine.
Faster access to tables is ensured if indexes are ready and maintained in the OS memory cache. As mentioned above, it is important for the OS to have enough space allocated.
The query optimizer favors the use of indexes when computing a request result. If a query cannot take advantage of the indexes, it will be resolved in Java memory, and experience poor performance on large volumes. The following guidelines apply:
Only XPath predicates and SQL queries can benefit from index optimization.
Some fields and some datasets cannot be indexed, as described in section Limitations.
XPath predicates on a multivalued field cannot benefit from index optimization, except for the osd:search function.
XPath predicates using the osd:label function cannot benefit from index optimization
If indexes have not yet been built, additional time is required to build and persist the indexes, on the first access to the table.
Accessing the table data blocks is required when the query cannot be computed against any index (whether for resolving a rule, filter or sort), as well as for building the index. If the table blocks are not present in memory, additional time is needed to fetch them from the database.
It is possible to get information through the memory monitoring and request logging categories.
The following access lead to poor performance, and must be avoided:
Access a table after a few modifications, repeatedly. It implies the index state to be refreshed after each modification. The cost of refreshing makes this pattern ineffective. Instead, perform a single query and apply the modification when browsing the results.
If there is an ongoing access to the same table, concurrently to the previous case, it prevents outdated index files to be deleted. As a consequence, the size of the index on disk increases, and the server may run out of disk space in extreme cases. When the concurrent access is closed, the index size is back to normal. This is usually a sign that a Request or a Query is not properly closed.
The new records creations or record insertions depend on the primary key index. Thus, a creation becomes almost immediate if this index is already loaded.
The merge information in history table (the merge_info field) has a potentially high access cost. To improve performance and if the client code does not need this field, the includeMergeInfo parameter must be set to false.
See History for more information.
In order to improve performance, a fetch size should be set according to the expected size of the result of the request on a table. If no fetch size is set, the default value will be used.
On a history table, the default value is assigned by the JDBC driver: 10 for Oracle and 0 for PostgreSQL.
On PostgreSQL, the default value of 0 instructs the JDBC driver to fetch the whole result set at once, which could lead to an OutOfMemoryError when retrieving large amounts of data. On the other hand, using fetchSize on PostgreSQL will invalidate server-side cursors at the end of the transaction. If, in the same thread, you first fetch a result set with a fetchsize, then execute a procedure that commits the transaction, then, accessing the next result will raise an exception.
While TIBCO EBX® is designed to support large volumes of data, several common factors can lead to poor performance. Addressing the key points discussed in this section will solve the usual performance bottlenecks.
For reference, the table below details the programmatic extensions that can be implemented.
Use case | Programmatic extensions that can be involved |
|---|---|
Validation | |
Table access | |
EBX® content display | |
Data update |
For large volumes of data, using algorithms of high computational complexity has a serious impact on performance. For example, the complexity of a constraint's algorithm is O(n 2 ). If the data size is 100, the resulting cost is proportional to 10 000 (this generally produces an immediate result). However, if the data size is 10 000, the resulting cost will be proportional to 10 000 000.
Another reason for slow performance is calling external resources. Local caching usually solves this type of problem.
If one of the use cases above displays poor performance, it is recommended to track the problem, either by code analysis or by using a Java profiling tool.
Refreshing a Lucene index takes time. It should be avoided whenever possible.
When does a refresh happen? | In the context of a transaction, an index refresh occurs when the table has been modified and one of the conditions below occurs:
|
Coding recommendations |
|
It is generally not advised to use a single transaction when the number of atomic updates in the transaction is beyond the order of 10 5 . Large transactions require a lot of resources, in particular, memory, from EBX® and from the underlying database.
To reduce the transaction size, it is possible to:
Specify the property ebx.manager.import.commit.threshold. However, this property is only used for interactive archive imports performed from the EBX® user interface.
Explicitly specify a commit threshold inside the batch procedure.
Structurally limit the transaction scope by implementing Procedure for a part of the task and executing it as many times as necessary.
On the other hand, specifying a very small transaction size can also hinder performance, due to the persistent tasks that need to be done for each commit.
If intermediate commits are a problem because transactional atomicity is no longer guaranteed, it is recommended to execute the mass update inside a dedicated dataspace. This dataspace will be created just before the mass update. If the update does not complete successfully, the dataspace must be closed, and the update reattempted after correcting the reason for the initial failure. If it succeeds, the dataspace can be safely merged into the original dataspace.
If required, triggers can be deactivated using the method ProcedureContext.setTriggerActivation.
Authentication and permissions management involve the user and roles directory.
If a specific directory implementation is deployed and accesses an external directory, it can be useful to ensure that local caching is performed. In particular, one of the most frequently called methods is Directory.isUserInRole.