Cache Topologies

Caching topology defines the strategy for data storage in a clustered cache. It refers to the behavior of cache cluster under different scenarios. These behavior specifications help in choosing the cache topology which suites best to an application’s environment.

NCache provides the following set of topologies for cluster caches.

• Mirrored Cache (Two-Server Active/Passive)
• Replicated Cache
• Partitioned Cache
• Partitioned-Replica Cache (Asynchronous and Synchronous)

Reference Data Versus Transactional Data: Cached data can be differentiated into two broad types i.e. Reference and Transactional. Reference refers to the type of data that has a high ratio of reads than writes e.g. product catalog prices may change after one or few days. On the other hand, transactional refers to the data that has almost the same ratio of reads and writes e.g. each ASP.NET Session is updated with all the activities of each application user.

Each of these caching topologies is explained in more detail below.

Mirror Cache

Mirror cache consists of two server active/passive nodes. One server node is active node which is basically the coordinator node. The coordinator node is the oldest server node in any cache cluster having membership management authorities. The other node is passive/mirror node.

Clients Connection: All cache clients can only connect with the active node. Passive node is added for creating the exact backup of active node.

Backup Support: Mirror cache provides data reliability by creating a backup through asynchronous replication of all updates on mirror/passive node. Due to asynchronous replication between active and passive node, cache clients get better performance. Active and passive servers can be interchanged at any time.

Failure Recovery: If active node goes down due to any reason, then passive node automatically becomes active without any intervention. All the cache clients connect to the new active node and can perform operations without any performance degradation and downtime.

Runtime Scalability: As there is a single active node in this topology, its scalability is the same as that of a local cache. But due to mirroring of data, it also provides fault tolerance which lacks in local cache.

State Transfer: Mirror node always replicates the state from active node when it joins the cache cluster.

Replicated Cache

Multiple Server Nodes: In replicated cache, there can be two or more server nodes. Each node is replica of the cache, i.e., all server nodes contain the same set of cached data.

More than One Backup: It has more than one backup of the cached data. With an increase of server nodes in cluster, the number of backups also increase.

Replication Mode: Replications are performed synchronously on all other nodes of the cache. In synchronous parallel replication, there is a chance that two parallel requests for updating the same data may result in different state on different replica nodes. But using a global token performs updates on all nodes in a sequenced fashion to avoid data integrity issues.

Replication Mode: Replications are performed synchronously on all other nodes of the cache. In synchronous parallel replication, there is a chance that two parallel requests for updating the same data may result in different state on different replica nodes. But using a global token performs updates on all nodes in a sequenced fashion to avoid data integrity issues.

Clients Load Balancing: Cache clients are load balanced automatically by the cluster for an even distribution of load on the cache nodes. This ensures that any one node is not overwhelmed by the cache client requests. Client load balancing is configurable and can be turned Off.

High Performance in Reference Data: Replicated cache is suitable when data read are more than data writes. For any cache client, reads are local to the server which means that cache client has direct access to all data.

Runtime Scalability: Server nodes can be added or removed in replicated cache like other topologies at run time.

Low Performance in Transactional Data: As more server nodes are added to scale the replicated cache, write overhead increases. So this topology is not suitable for applications having transactional type of data.

Storage Limitation to Each Server Node: Adding more nodes does not increase the storage capacity because every node has same data set.

High Availability: Replicated topology is best for applications where high availability and reliability with zero down time is the priority. This is because this topology contains multiple backups of cached data set.

State Transfer: Every server node replicates the state from the coordinator node (the senior most node in terms of cluster joining) when it joins the cache cluster.

Partitioned Cache

Multiple Server Nodes: This topology has two or more servers. When any cache item is added in the cache it is saved only to the relevant server node which means that every server node has a unique set of data.

Data Distribution: Data is distributed/partitioned among all server nodes on the basis of the hash code of cache key. A hash based distribution map is generated which is then distributed to all server nodes.

Clients Connections: Cache clients establish connections with every server node of the cluster. Data distribution maps are also shared with cache clients and cache clients are aware of the actual data distribution. When a cache client wants to send a request to cache servers, it first determines to which server node that key belongs to and sends the request accordingly on the basis of the distribution map.

One Hop Reads and Writes: Based on the distribution map, cache client adds/fetches cached item to/from the exact server node. In this manner cache reads and writes both become one hop operation and there are no delays.

No Backups/Fault Tolerance: When any server node leaves the cluster, then data residing on that node is lost because there is no replica for any node. That's why there is no fault tolerance in terms of cached data.

Runtime Scalability: Server nodes can be added or removed at runtime when ever needed.

High Performance: Scaling the cache will not affect the performance of cluster cache for both reference and transactional data because read/write is one hop operation in this topology. It will result in high performance in reads/writes without affecting the performance of cluster cache.

Storage Scalability: When more nodes are added in partitioned cache, storage capacity is increased because each partition in the cluster has a unique set of data.

Reliability: This topology is not fault tolerant because there is no backup of data and if any server node goes down, then data belonging to that partition is permanently lost.

State Transfer: When a node joins the cache cluster in Partitioned topology, existing nodes sacrifice a portion of their data, according to hash based distribution, for the newly joined node. Newly joined node transfers its data share from existing nodes which results in an even distribution of the cached data among all cluster nodes. State transfer occurs in background. Therefore during state transfer, user cache operation does not suffer.

Partitioned-Replica Cache

Multiple Server Nodes: This topology has two or more servers. When any cache item is added in the cache it is saved only to the relevant server node which means that every server node has a unique set of data.

Data Distribution: Data is distributed/partitioned among all server nodes on the basis of the hash code of cache key. A hash base distribution map is generated which is then distributed to every server node.

Fault Tolerance: It provides fault tolerance up to one node replica. Each server node in the cluster has one active partition and its passive/mirror replica on the other node. Thus when one node is down data can be restored from its replica.

Double Memory Consumption: Each server node consumes double memory than the configured size because that node contains one active partition and one replica. Both active and replica has the same size as configured for cache.

Clients Connections: Cache clients are connected to all server nodes in this topology. Data distribution maps are also shared with cache clients and cache clients are aware of the actual data distribution. Operations are directly sent to the node containing data with the help of distribution map on the basis of cache key.

Runtime Scalability: Server nodes can be added or removed at runtime whenever needed.

Storage Scalability: When more nodes are added in partitioned-replica cache, storage capacity is increased because each partition in the cluster has a unique set of data.

High Performance: Scaling the cache will not affect the performance of cluster cache for both reference and transactional data because read/write is one hop operation in this topology. It will give high performance in reads/writes without affecting the performance of cluster cache. Addition of new nodes to the cache cluster results in load distribution of existing members resulting in more throughput.

State Transfer: When a node joins the cache cluster in partitioned-replica topology, existing nodes sacrifice a portion of their data, according to hash based distribution, for the newly joined node. Newly joined node gets its data share transfered from existing nodes which results in an even distribution of the cached data among all cluster nodes. Similarly, when a node leaves the cache cluster, data from its backup node is distributed among existing partitions. State transfer also occur between active and passive nodes. State transfer occurs in background. Therefore, during state transfer, user cache operation does not suffer.

Replication strategy/modes: Following are the data replication modes provided in this topology.

Synchronous Replication

Cache clients have to wait for replication to the backup partition as part of the cache operation. Due to synchronous replication it can be assured that if cache client request is executed successfully, then data is already stored to the backup node.

Asynchronous Replication

Cache clients request returns immediately after operation is performed on active partition node. Operations are replicated asynchronously to the replica node. If any node goes down disgracefully, then operations that remain in the replicator queue could be lost.