Replica traffic manager for data grids

Replication is used in data grids to improve the efficiency of user requests. One or more copies of files are stored on storage components, replicas, in the data grid. Currently, users send their requests directly to these replicas. There is no control over a request once it leaves the user. The focus of this paper is controlling workload traffic at data grid replicas, by managing the flow of requests to each replica. We propose the creation of a replica traffic manager that controls workload traffic sent to the individual replicas in the data grid. The traffic manager receives all user requests and manages the traffic for all replicas, by maintaining a certain number of outstanding requests at each replica. When a particular replica is heavily loaded, all incoming requests for that replica would be held in queue at the traffic manager and/or directed to another replica. Once the traffic decreases at the replicas, the queued requests would be immediately forwarded. By limiting the traffic to each replica, the traffic manager has more control over the system than otherwise possible with individual users submitting requests directly to the replicas. In our evaluations, we observe that our replica traffic manager has a significant and beneficial effect on the performance of the data grid. It provides reliable and consistent response times for user requests. In addition, the traffic manager can adjust for overloaded or failing replicas, while ensuring a starvation free environment that provides dynamic load balancing.