Oracle Coherence provides explicit support for efficient caching of huge amounts of automatically-expiring data using potentially high-latency storage mechanisms such as disk files. The benefits include supporting much larger data sets than can be managed in memory, while retaining an efficient expiry mechanism for timing out the management (and automatically freeing the resources related to the management) of that data. Optimal usage scenarios include the ability to store many large objects, XML documents or content that will be rarely accessed, or whose accesses will tolerate a higher latency if the cached data has been paged to disk.
This feature is known as a Serialization Paged Cache:
The result is a feature that organizes data in the cache based on the time that the data was placed in the cache, and then is capable of efficiently expiring that data from the cache, an entire page at a time, and typically without having to reload any data from disk.
The primary configuration for the Serialization Paged Cache is composed of two parameters: The number of pages that the cache will manage, and the length of time represented by each page. For example, to cache data for one day, the cache can be configured as 24 pages of one hour each, or 96 pages of 15 minutes each, etc.
Each page of data in the cache is managed by a separate Binary Store. The cache requires a Binary Store Manager, which provides the means to create and destroy these Binary Stores. Coherence 2.4 provides Binary Store Managers for all of the built-in Binary Store implementations, including disk (referred to as "LH") and the various NIO implementations.
A new optimization is provided in Coherence 2.4 for the partitioned cache service, since - when it is used to back a partitioned cache - the data being stored in any of the Serialization Maps and Caches is entirely binary in form. This is called the Binary Map optimization, and when it is enabled, it gives the Serialization Map, the Serialization Cache and the Serialization Paged Cache permission to assume that all data being stored in the cache is binary. The result of this optimization is a lower CPU and memory utilization, and also slightly higher performance.
Explicit support is also provided in the Serialization Paged Cache for the high-availability features of the partitioned cache service, by providing a configuration that can be used for the primary storage of the data as well as a configuration that is optimized for the backup storage of the data. The configuration for the backup storage is known as a passive model, because it does not actively expire data from its storage, but rather reflects the expiration that is occurring on the primary cache storage. When using the high-availability data feature (a backup count of one or greater; the default is one) for a partitioned cache service, and using the Serialization Paged Cache as the backing storage for the service, we strongly suggest that you also use the Serialization Paged Cache as the backup store, and configure the backup with the passive option.
When used with the distributed cache service, special considerations should be made for load balancing and failover purposes. The partition-count parameter of the distributed cache service should be set higher than normal if the amount of cache data is very large or huge; that will break up the overall cache into smaller chunks for load-balancing and recovery processing due to failover. For example, if the cache is expected to be one terabyte in size, twenty thousand partitions will break the cache up into units averaging about 50MB in size. If a unit (the size of a partition) is too large, it will cause an out-of-memory condition when load-balancing the cache. (Remember to make sure that the partition count is a prime number; see http://primes.utm.edu/lists/small/ for lists of prime numbers that you can use.)
In order to support huge caches (e.g. terabytes) of expiring data, the expiration processing is performed concurrently on a daemon thread with no interruption to the cache processing. The result is that many thousands or millions of objects can exist in a single cache page, and they can be expired asynchronously, thus avoiding any interruption of service. The daemon thread is an option that is enabled by default, but it can be disabled.
When the cache is used for large amounts of data, the pages will typically be disk-backed. Since the cache eventually expires each page, thus releasing the disk resources, the cache uses a virtual erase optimization by default. This means that data that is explicitly removed or expired from the cache is not actually removed from the underlying Binary Store, but when a page (a Binary Store) is completely emptied, it will be erased in its entirety. This reduces I/O by a considerable margin, particularly during expiry processing as well as during operations such as load-balancing that have to redistribute large amounts of data within the cluster. The cost of this optimization is that the disk files (if a disk-based Binary Store option is used) will tend to be larger than the data that they are managing would otherwise imply; since disk space is considered to be inexpensive compared to other factors such as response times, the virtual erase optimization is enabled by default, but it can be disabled. Note that the disk space is typically allocated locally to each server, and thus a terabyte cache partitioned over one hundred servers would only use about 20GB of disk space per server (10GB for the primary store and 10GB for the backup store, assuming one level of backup.)