Talk by Uwe Röhm, U Sydney: Serialisable Snapshot Isolation on Multicore Servers, Jan 8, 14:15
Written: 07.01.2019 11:15 Written By: Ankur Sharma
Serialisable Snapshot Isolation on Multicore Servers
Database systems strive to provide efficient read/write access to large, shared data sets. At the core of these systems sits the concurrency control component whose role is to isolate concurrent transaction executions. Modern database workloads are dominated by queries though, which makes snapshot databases based on a multi-version storage layer an attractive system design. It is well known that snapshot-based algorithms scale better for read-only transactions than locking-based systems. However, a major pitfall is that the standard snapshot isolation (SI) algorithm allows non-serialisable executions...
This talk revisits the development of snapshot based concurrency control algorithms and discusses an efficient approach to provide serialisable snapshot isolation inside a database system - with almost the same performance than standard SI. We further take a look at scalability of SI-based database engines on multicore servers. We will show that current implementations of SI do not scale well as number of CPU cores increases, and then discuss approaches to avoid this scalability bottleneck with modern database systems on modern multicore servers.
About the Speaker:
Uwe Roehm is Associate Professor in database systems at the University of Sydney. He is a computer science graduate from the University of Passau, Germany, and completed his PhD at ETH Zurich in the area of scheduling combined OLTP/OLAP workloads in a cluster of databases. Much of his research has dealt with transaction management and replication, especially how to ensure sufficient freshness in values read. His current research interests are cloud data management, database engines on modern hardware, and in-database support of complex data-intensive computations, eg. for Data Science. He has been a visiting researcher at Microsoft, and held visiting positions at the Karlsruhe Institute of Technology (KIT), and at TU Munich. He currently is guest professor at the data management group of TU Darmstadt, Germany.
Tuesday, Jan 8, 14:15
E1 4, 024