Design And Implementation Of High-contention Processing Architecture For Shared-Nothing Databases

Internet applications’ increasing number of users and the demand to be online all the time paved the way for the fast development and widespread application of distributed databases.Among distributed databases,Shared-Nothing distributed databases,famous for their outstanding scalability and high availability,are favored by customers and database vendors.They can leverage low-cost midrange servers to deploy largescale distributed systems so that they can meet users’ demands for performance at a lower cost.To get better scalability,Shared-Nothing distributed databases normally adopt a stateless calculation layer,thus they will have transaction processing and transaction states placed in different nodes,which may cause a long conflict detection chain.Under a high-contention workload,this configuration will cause a hefty dropdown in performance.To tame this problem,we designed a scheme to resolve the issue of SharedNothing architecture databases under high contention.Our contributions are enlisted below:1.We designed and implemented a high-contention detection module and a highcontention processing module.It can detect the high-contention status of the system in time and activate the corresponding high-contention processing strategy agilely.For dynamic high-contention workload,we utilized a partition-based method to manage high-contention data and placed high-contention transactions in the same high-contention processing node to be processed by the high-contention processing strategy,thus enabling the system to process high-contention transactions more effectively.2.For optimistic concurrency control under distributed environment,we designed and implemented a pre-locking method to reduce the conflict detection chain of high-contention transactions in Shared-Nothing architecture,thus reducing the system consumption of resources and increasing overall throughput.3.We designed a high-contention processing strategy based on a local cache to reduce the time for reading high-contention data,thus reducing the conflict detection chain of high-contention transactions and enhancing system performance.4.To resolve the single-node bottleneck caused by a single high-contention processing node,we introduced the load-balance strategy to process high-contention workload while avoiding the single-node bottleneck.5.We designed and implemented a prototype system based on Foundation DB to examine our ideas.In conclusion,we researched the performance issue of Shared-Nothing architecture distributed databases under high-contention workload.We designed and implemented a high-contention transaction processing solution.We conducted a set of experiments.Experiments illustrated that our design and optimization of high-contention transaction processing can improve shared-nothing distributed databases’ performance under the high-contention workload.