PANOPTICON: An Omniscient Lock Broker for Efficient Distributed Transactions in the Datacenter

For datacenter applications that require tight synchronization, transactions are commonly employed for achieving concurrency while preserving correctness. Unfortunately, distributed transactions are hard to scale due to the decentralized lock acquisition and coordination protocols they employ. In this paper, we show that it is possible to achieve scalability for distributed transactions by using a lock broker architecture, and present the design and development of such a framework, called PANOPTICON. Panopticon achieves scalability by divorcing locks from the data items and striving to improve lock access locality. A lock can be hosted at the lock broker or at a different server than the server that hosts the corresponding data item. The lock broker mediates the access to data shared across servers by migrating the associated locks like tokens, and in the process gets to learn about the access patterns of transactions. We show that the broker can drastically improve the lock access locality and, hence, the performance of distributed transactions by employing simple rules. Our experiments show that Panopticon performed significantly better than distributed transactions as the the number of data items and number of servers involved in transactions increase. Moreover, as the history locality (the probability of using the same objects in consecutive transactions) increase, Panopticon’s lock migration strategies improved lock-access locality and resulted in significantly better performance.