An active data-aware cache consistency protocol for highly-scalable data-shipping DBMS architectures

Abstract

In a data-shipping database system, data items are retrieved from the server machines, cached and processed at the client machines, and then shipped back to the server. Current cache consistency approaches typically rely on a centralized server or servers to enforce the necessary concurrency control actions. This centralized server imposes a limitation on the scalability and performance of these systems. This paper presents a new consistency protocol, Active Data-aware Cache Consistency (ADCC), that allows clients to be aware of the global state of their cached data via a two-tier directory. Using parallel communication with simultaneous client-server and client-client messages, ADCC reduces the network latency for detecting data conflicts by 50%, while increasing message overhead by about 8% only. In addition, ADCC improves scalability by partially offloading the concurrency control function from the server to the clients. An optimization, Lazy Update, is introduced to reduce the message overhead for maintaining client directory consistency. We implement ADCC in a page server DBMS architecture and compare it with the leading cache consistency algorithm, Callback Locking (CBL), which is the most widely implemented algorithm in commercial DBMSs. Our performance study shows that ADCC has a similar or lower abort rate, higher throughput, and better scalability for important workloads and system configurations. Both the simulation results and the analytic study indicate that the message overhead is low and that ADCC produces better behavior compared to the traditional server-based communication under high contention workloads.

A unified platform for data driven web applications with automatic client-server partitioning

“…To limit the amount of server-side data required for each client session, the developer can specify a maximum life span for the data in the server cache, as well as the heartbeat frequency of the client. Our cache consistency strategy is similar to a detection based approach for transactional client server caching [12,25,18], although the system allows for the integration of other strategies in the future.…”

“…Work in this area has focused on Transactional Client-Server Cache Consistency [12,18,25], a technique that evaluates part of a transaction at the client by shipping it the required data. This work is concerned with guaranteeing the ACID properties of a transaction, and proposes many different approaches such as the Avoidance Based Approach (Adaptive CallBack Locking) and the Detection Based Ap- Hybrid Shipping Architectures have been proposed to run queries in a distributed setting [22,23].…”

Self-tuning speculation for maintaining the consistency of client-cached data

“…(3) A demonstration that, with high contention and transaction-level temporal locality, write/write conflicts can dominate other overheads when the transaction granularity (size) is reduced. Similar to previous work [4,5,13], this study focuses on a page-server architecture with page level consistency. We compare SADCC with the two leading cache consistency algorithms, Active Data-aware Cache Consistency (ADCC) [13] and Asynchronous Avoidance-based Cache Consistence (AACC) [8].…”

“…However, putting the server on the critical path unavoidably increases the communication latency. A protocol that employs client-client and server-based communication was recently proposed to address this problem [13]. However, in current optimistic schemes (asynchronous [8,13] or deferred [1]), clients always proceed to write to locally-cached copies under the assumption that no other clients are currently using the data.…”

“…The primary contributions of this research are: (1) A new algorithm, Self-tuning Active Data-aware Cache Consistency (SADCC), that self-tunes between optimistic and pessimistic consistency control. (2) A demonstration that the integration of self-tuning speculation and parallel communication in SADCC improves throughput by 6% compared to ADCC [13] and by 14% compared to AACC [8] in a high contention environment. (3) A demonstration that, with high contention and transaction-level temporal locality, write/write conflicts can dominate other overheads when the transaction granularity (size) is reduced.…”