The Cost of Serializability on Platforms That Use Snapshot Isolation

Serializable isolation for snapshot databases

“…We describe the implementation of a prototype based on our algorithm in the InnoDB storage manager for MySQL. The detailed evaluation we give here includes a new benchmark based on a small modification to TPC-C (Transaction Processing Performance Council, 2005), which is more complex and representative of typical OLTP workloads than the SmallBank benchmark presented in (Alomari et al, 2008a) and used in (Cahill et al, 2008). Our results indicate that Serializable SI provides serializable isolation with performance close to snapshot isolation and generally much better than strict two-phase locking.…”

“…All transactions start by looking up the customer's name in the Account table to retrieve the CustomerID. Here, we translate the SQL descriptions of the transaction programs given in (Alomari et al, 2008a) into read and write operations that can be processed by Berkeley DB. This assumes that there are Berkeley DB handles called account, savings and checking referring to the three tables in the schema.…”

“…The original SmallBank paper, (Alomari et al, 2008a), implemented the benchmark in a relational DBMS. Here we describe how we implemented SmallBank for Berkeley DB, which provides a general purpose storage manager, but no query processing or SQL layer.…”

“…However, we cannot use TPC-C directly to compare different ways of making applications serializable, since TPC-C itself generates only serializable executions on SI-based platforms, as described above. Thus in this thesis we have used a benchmark called SmallBank, published in (Alomari et al, 2008a), which was designed to offer a diverse choice among modifications that will ensure serializable execution on SI. The SmallBank transaction mix is based on the example of an SI anomaly from (Fekete et al, 2004) and provide some functionality reflecting a simple banking system.…”

“…In this thesis we have used three different benchmarks to evaluate our prototypes. Firstly, the SmallBank benchmark from (Alomari et al, 2008a) was adapted to run on Berkeley DB. Secondly, a new microbenchmark that we call sibench was designed and implemented for InnoDB to highlight the performance of its concurrency control algorithms including the SI and Serializable SI levels that we added.…”