The Java Virtual Machine (JVM) is configured in <KIM_HOME>/bin/config/config(.bat)
Unlike the configuration in /config, the one in /bin/config is applied before the JVM has started. Thus, /bin/config/config(.bat) contains all options that cannot be modified after start-up. They can either be changed in the config script itself, or defined as environment variables.
- JAVA_HOME - The full path to the virtual machine.
In a production environment, we recommend installing the JVM in a path with no spaces or special characters in it. For example, on Windows, we recommend c:\java instead of c:\Program Files\Java. - KIM_MAX_JAVA_HEAP - The maximum memory that the KIM Server is allowed to allocate.
Set to 2 GB for basic load and to 8 GB for maximum load (for 64 bit machine only). See the memory section below for details on memory consumption of resources.
Memory
There are multiple factors that affect how much memory KIM requires. The total is the amount of memory required for each component.
Size of GATE documents during annotation
If you enable semantic annotation in the KIM Server, you should reserve some memory for the in-memory representation of the documents during the annotation. The amount of memory per document depends on the document size. Please estimate the maximum size of a single document, and then refer to memory consumption of GATE documents to determine the required memory.
Remember to multiply that number by the number of parallel annotation executions, configured here.
To reduce memory consumption, set small number of parallel annotation executions or disable semantic annotation altogether.
Size of in-memory resources for semantic annotation
The pipeline for semantic annotation keeps some resources in-memory for improved performance. To enable parallel annotation executions, multiple copies of the pipeline are kept in-memory - one for each potential execution. Some of the in-memory resources are shared between the pipeline copies.
Thus, the amount of memory consumed for the in-memory resources for semantic annotation is calculated using the following formula:
Mem = sum for each enabled pipeline{P1 + (N-1) *P}
where:
- Mem is the total memory consumed
- P1 is the memory required for the first instance of a pipeline
- N is the number of parallel annotation executions
- P is the amount of memory required for a pipeline copy
To reduce the required memory for this component, you can:
- reduce P1, by having a simpler pipeline
- reduce N as described above
- reduce P by making sure the pipeline copies share as much in-memory resources as possible
- disable semantic annotation altogether. That is useful in read-only public servers.
BigOWLIM In-memory index
The BigOWLIM in-memory index takes a significant amount of memory determined by the BigOWLIM configuration.
Query result cache
The query result cache is implemented via SoftReferences, so its size is automatically reduced.
Leaving additional memory enables KIM to cache more query results, which improves performance.
Other components
The other standard KIM components use a fairly constant amount of memory. In KIM Server 3.6, it is 550MB. This amount will be larger if there are custom components, initialized as KIM Extensions or services.
| About memory use Avoid swapping. KIM only keeps in-memory data that is critical for performance. If the operating system puts some of KIM memory in swap, the speed advantage of keeping data in-memory will not only be negated, but reversed. If there is not enough physical memory, please reduce the memory requirements of KIM, using the advice above. |