An Adaptive Scheduling Framework for Distributed Key-Value Stores Using RDMA

Many applications need to cope with Key-Value data, which imposes great pressure on Key-Value storage especially on large-scale workloads. To improve the throughput of Key-Value storage in modern distributed environments, we present an efficient scheduling framework to distribute the workloads onto different nodes. We focus on how to adaptively forward batching requests to a certain number of nodes each disposing a portion of Key-Value items corresponding to a surrogate key. To reduce the overhead of both round-trip notification and the contention derived from skewed workloads, an asynchronous communication method is presented to boost the compaction and coordination in each scheduler. It can be shown that the scheduling framework can fully exploit the high throughput of one-sided writes of modern RDMA networks, such that both the workloads and contention imposed on Key-Value servers can be significantly reduced. We conduct intensive experiments using YCSB benchmark on top of 100-gbps RMDA network. The results show that our proposed method can improve the Key-Value throughput by a factor of two when serving an in-memory Key-Value store with up to 256 work threads.