Program guide > Performance considerations
Byte array maps
You can store the key-value pairs in the maps in a byte array instead of POJO form, which reduces the memory footprint that a large graph of objects can consume.
The amount of memory that is consumed increases with the number of objects in a graph of objects. By reducing a complicated graph of objects to a byte array, only one object is maintained in the heap instead of several objects. With this reduction of the number of objects in the heap, the Java™ run time has fewer objects to search for during garbage collection.
The default copy mechanism used by WebSphere eXtreme Scale is serialization, which is expensive. For instance, if using the default copy mode of COPY_ON_READ_AND_COMMIT, a copy is made both at read time and at get time. Instead of making a copy at read time, with byte arrays, the value is inflated from bytes, and instead of making a copy at commit time, the value is serialized to bytes. Using byte arrays results in equivalent data consistency to the default setting with a reduction of memory used.
When using byte arrays, note that having an optimized serialization mechanism is critical to seeing a reduction of memory consumption. For more information, see Serialization performance.
Configure byte array maps
You can enable byte array maps with the ObjectGrid XML file by modifying the CopyMode attribute that is used by a map to the setting COPY_TO_BYTES, shown in the following example:
<backingMap name="byteMap" copyMode="COPY_TO_BYTES" />
See ObjectGrid descriptor XML file for more information.
You must consider whether or not to use byte array maps in a given scenario. Although you can reduce the memory use, processor use can increase when you use byte arrays.
The following list outlines several factors that should be considered before choosing to use the byte array map function.
Comparatively, memory reduction may not be possible when using byte array maps for some object types. Consequently, several types of objects exist for which you should not use byte array maps. If you are using any of the Java primitive wrappers as values, or a POJO that does not contain references to other objects (only storing primitive fields), the number of Java Objects is already as low as possible–there is only one. Since the amount of memory used by the object is already optimized, using a byte array map for these types of objects is not recommended. Byte array maps are more suitable to object types that contain other objects or collections of objects where the total number of POJO objects is greater than one.
For example, if you have a Customer object that had a business Address and a home Address, as well as a collection of Orders, the number of objects in the heap and the number of bytes used by those objects can be reduced by using byte array maps.
When using other copy modes, applications can be optimized when copies are made if objects are Cloneable with the default ObjectTransformer or when a custom ObjectTransformer is provided with an optimized copyValue method. Compared to the other copy modes, copying on reads, writes, or commit operations will have additional cost when accessing objects locally. For example, if you have a near cache in a distributed topology or are directly accessing a local or server ObjectGrid instance, the access and commit time will increase when using byte array maps due to the cost of serialization. You will see a similar cost in a distributed topology if you use data grid agents or you access the server primary when using the ObjectGridEventGroup.ShardEvents plug-in.
With byte array maps, objects are not inflated when communicating from a client to a server unless the server needs the POJO form. Plug-ins that interact with the map value will experience a reduction in performance due to the requirement to inflate the value.
Any plug-in that uses LogElement.getCacheEntry or LogElement.getCurrentValue will see this additional cost. To get the key, you can use LogElement.getKey, which avoids the additional overhead associated with the LogElement.getCacheEntry().getKey method. The following sections discuss plug-ins in light of the usage of byte arrays.
Indexes and queries
When objects are stored in POJO format, the cost of doing indexing and querying is minimal because the object does not need to be inflated. When using a byte array map we will have the additional cost of inflating the object. In general if your application uses indexes or queries, it is not recommended to use byte array maps unless you only run queries on key attributes.
When using the optimistic locking strategy, you will have the additional cost during updates and invalidate operations. This comes from having to inflate the value on the server to get the version value to do optimistic collision checking. If you are just using optimistic locking to guarantee fetch operations and do not need optimistic collision checking, you can use the com.ibm.websphere.objectgrid.plugins.builtins.NoVersioningOptimisticCallback to disable version checking.
With a Loader, we will also have the cost in the eXtreme Scale run time from inflating and reserializing the value when it is used by the Loader. You can still use byte array maps with Loaders, but consider the cost of making changes to the value in such a scenario. For example, you can use the byte array feature in the context of a read mostly cache. In this case, the benefit of having less objects in the heap and less memory used will outweigh the cost incurred from using byte arrays on insert and update operations.
When using the transactionEnd method in the ObjectGridEventListener plug-in, we will have an additional cost on the server side for remote requests when accessing a LogElement's CacheEntry or current value. If the implementation of the method does not access these fields, then we will not have the additional cost.
Parent topic:Access data in WebSphere eXtreme Scale
Parent topic:Performance considerations for application developers
CopyMode best practices
Plug-in evictor performance best practices
Locking performance best practices
ObjectTransformer interface best practices
Query performance tuning