Automatic data replica manager in distributed caching and data processing systems

1. A method of data storage, the method comprising:

determining a latency distance from a primary node to two or more replica nodes in a tier;

choosing a preferred replica node of the two or more replica nodes based on the determined latency distances; and

migrating replica data from a second replica node of the two or more replica nodes to the preferred replica node using a fusion cache method comprising:

write-caching new data from the primary node to the preferred replica node;

copying the replica data from the second replica node to the preferred replica node; and

maintaining the replica data existing data on the second replica node, wherein the migration of the replica data from the second replica node to the preferred replica node is performed during operation of a system comprising the preferred replica node.

2. The method of claim 1 , wherein the two or more replica nodes have a smaller latency distance to the primary node than all available nodes of a datacenter cluster comprising the primary node and the two or more replica nodes.

3. The method of claim 1 , wherein one or more nodes of a datacenter cluster as a primary node has a ranking of preferred replica nodes that are ranked based on network delays associated with paths between the primary node and respective ones of the replica nodes, and

wherein the one or more nodes of the datacenter cluster comprise:

a solid-state drive tier as a cache tier comprising a cache partition for storing data of local virtual machines and a replica partition for storing replica data from other nodes; and

a hard disk drive tier as a back end storage tier.

4. The method of claim 3 , wherein the one or more nodes further comprise a physical host for running virtual machines, the virtual machines sharing the solid-state drive tier and the hard disk drive tier in a respective node of the datacenter cluster, the method further comprising periodically updating the ranking of the preferred replica nodes to account for changes in network delays.

5. The method of claim 3 , wherein the ranking of the preferred replica nodes are further based on workload characteristics comprising access frequency of data associated with the preferred replica nodes and service level agreements (SLAs) associated with the preferred replica nodes.

6. A method of using replicated data to enable parallel prefetching from nodes of a datacenter cluster, the method comprising:

splitting a dataset into two parts;

loading the two parts of the dataset from corresponding nodes comprising a primary node of the datacenter cluster, and a replica node of the datacenter cluster;

determining a difference between an access speed corresponding to the replica node and an access speed corresponding to the primary node exceeds a threshold, or determining a utilization ratio of throughput corresponding to the replica node is less than an operating threshold; and

triggering parallel prefetching of the dataset from the primary node and the replica node, such that a makespan of the parallel prefetchinq is less than reading the two parts of the dataset from the primary node.

7. A method of using replicated data to enable parallel prefetching from multiple nodes of a datacenter cluster, the method comprising:

splitting a dataset into two parts;

loading the two parts of the dataset from corresponding nodes comprising a primary node of the datacenter cluster, and a replica node of the datacenter cluster;

determining a difference between an access speed corresponding to the replica node and an access speed corresponding to the primary node exceeds a threshold, or determining a utilization ratio of throughput corresponding to the replica node is less than an operating threshold;

determining a ratio of a total size of the dataset to an access speed of a solid-state drive of the primary node is greater than or equal to a maximum of a ratio of a size of a first of the two parts from the solid-state drive of the primary node to an access speed of the solid-state drive of the primary node and a ratio of a size of a second of the two parts from a solid-state drive of the replica node to an access speed of the solid-state drive of the replica node; and

triggering parallel prefetching.

8. The method of claim 7 , wherein λ 1 and λ 2 account for corresponding network delays of the datacenter cluster.

9. The method of claim 7 , further comprising:

determining a ratio of the access speed of the solid-state drive of the primary node to a sum of the access speed of the solid-state drive of the primary node and the access speed of the solid-state drive of the replica node; and

achieving a network makespan equal to a ratio of the total size of the dataset to a sum of the access speed of the solid-state drive of the primary node and the access speed of the solid-state drive of the replica node.