Load Balancing of Distributed Datastore in OpenDaylight Controller Cluster

Abstract

Software defined networking controller platforms support clustering architectures to provide scalability and availability to large-scale carrier-grade networks. OpenDaylight, a well-known open source project, provides a clustering and distributed datastore architecture. The datastore is distributed into shards such that a subset of the shards can be located in any cluster member. To guarantee strong consistency in the datastore so all shard replicas have the same value, only the shard leader is responsible for accepting data updates. Even though OpenDaylight supports a distributed shard leader election algorithm, elected leaders are not distributed over the cluster members due to a lack of centralized control. Since update requests for the datastore are inherently concentrated in the leader, this results in unbalanced CPU usage of the controller cluster, as well as degradation of throughput in the distributed datastore. In this paper, we propose a shard leader distribution algorithm that maximizes the throughput of the distributed datastore by distributing the shard leaders to cluster members as evenly as possible. The shard leader distribution algorithm restricts the maximum number of leaders that a cluster member may have by monitoring the status of the shards within the cluster members. In the performance evaluation, we prove that shard leader distribution significantly improves the throughput of the datastore by load balancing data update requests to the cluster.