XPS: EXPL: SDA: Scalable Concurrency Control Techniques for Distributed Systems

点击翻译按钮获取中文摘要

Virtually all distributed computing applications, from transactions on databases to updates on social media platforms, involve concurrent operations on data objects. For these applications, concurrency control mechanisms represent significant performance overheads. These applications typically exhibit strong and persistent patterns in data access. Motivated by the importance of the problem, this project investigates the use of dynamic data- and lock-access patterns in distributed computations to significantly improve the performance of concurrency control mechanisms for scalable systems, specifically, in conventional cloud environments and key-value stores such as BigTable, HBase, and Cassandra. In contrast to conventional techniques that collocate locks with corresponding data items, this project relies on a modular lock service that decouples lock locations from corresponding data objects, and maintains lock state of all data items in a small set of storage nodes. This design choice motivates a number of questions for this research: (i) where and when should lock states be migrated into the lock service? (ii) when should lock state be repatriated to the data store? (iii) how should the lock service be scaled out? (iv) what are fault-tolerant, low-overhead, deadlock- and livelock-free protocols for these operations? and (v) how can long-lived data access patterns be leveraged in such systems? Building on preliminary results that demonstrate the feasibility and considerable promise of the approach, the project develops algorithms, protocols, analyses, and open-source software, along with comprehensive validation in the context of a diverse set of applications.The project will result in a novel framework for concurrency control in scalable distributed systems. The concurrency control service has a number of desirable features: (i) modularity -- the service can be instantiated at runtime, with minimal change to underlying data storage organization and access mechanisms; (ii) extensibility ? the service adapts dynamically to load and service requirements; and (iii) high performance through the use of efficient algorithms exploiting data and lock access patterns. These features are achieved through a novel mix of algorithms for lock migration and collocation, statistical models for dynamic lock and data access, protocols for lock state management, associated proofs of correctness and fairness, fault tolerance, performance, and scalability. The concurrency control service is fully validated on private as well as public clouds on a mix of applications drawn from Online Transaction Processing and Machine Learning.The project directly impacts an important class of cloud-based applications by providing a modular and extensible lock service. The service relieves burden on the application programmer while providing high performance and elastic throughput. Beyond this, the project includes a number of educational initiatives aimed at undergraduate and graduate education, along with outreach efforts aimed at enhancing representation of minority groups. These include development of instructional material, curricula, organization of and presentations at workshops and summer schools, and recruitment initiatives aimed at students from under-represented groups.