Transparent Transaction Processing with A High-Performance Proxy for Distributed KVS
Transparent Transaction Processing with A High-Performance Proxy for Distributed KVS
复制标题
使用分布式 KVS 的高性能代理进行透明事务处理
DOI：
10.1109/candarw51189.2020.00051
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Hiroshi Matsuo
中科院分区：
文献类型：
--
作者：
Ryuya Miwa;Motoshi Miyamoto;Ryota Kawashima;Hiroshi Matsuo
分布式一致性协议在分布式KVS（D-KVS）中需要进行大量的事务处理通信。Google Spanner和Amazon DynamoDB使用从原子钟获得的高度精确的时间戳来减轻通信开销。然而，引入特殊的硬件设备会使操作管理复杂化，并会导致系统发生剧烈变化。本文提出了一种不依赖于分布式共识协议和专用硬件设备的透明事务处理方法。我们通过在集中式高性能代理上执行独占控制和事务处理，提供符合ACID的事务处理功能。代理通过管理多线程上的锁和事务，为D-KVS提供ACID兼容的事务处理功能，并保证中等规模D-KVS（数十个节点）的性能。在评估中，首先，我们评估了具有事务处理功能的代理的开销，吞吐量下降了0.2%到12%。接下来，我们确认代理处理Cassandra的轻量级事务的执行时间约为60%。最后，我们证实了代理可以处理事务，与MariaDB的事务处理性能相比，它的运行速度比MariaDB快。
Distributed consensus protocols require a large amount of communications for transaction processing in distributed KVS (D-KVS). Google Spanner and Amazon DynamoDB alleviate the communication overhead using highly accurate timestamping obtained from atomic clocks. However, introducing special hardware devices complicates operation management and drastic system changes. In this paper, we propose a transparent transaction processing method without depending on distributed consensus protocols and special hardware devices. We provide transaction processing functions that complies with ACID by performing exclusive control and transaction processing on a centralized high-performance proxy. The proxy provides ACID-compliant transaction processing functions to D-KVS by managing locks and transactions on multiple threads and guarantees the performance in medium-scale D-KVS (dozens of nodes). In the evaluation, first, we evaluate the overhead of the proxy with transaction processing functions, the throughput decreased by 0.2% to 12%. Next, we confirmed that the proxy processes the lightweight transaction of Cassandra with about 60% execution time. Finally, we confirmed that the proxy can process trans-actions, compared the transaction processing performance with MariaDB, it operated faster than MariaDB.
链式复制的理论与实践
DOI：
--
发表时间：
2010
期刊：
Erlang Workshop
影响因子：
--
作者：
S. Fritchie
通讯作者：
S. Fritchie