Guaranteeing the physical consistency of shared data in a main-memory DBMS

In this paper, we discuss the implementation strategy of the latch adopted in Tachyon, a main-memory DBMS (MMDBMS). The latch is a tool that guarantees physical consistency of the shared data simultaneously accessed by multiple transactions. Transactions hold the latch assigned to shared data before access, and also release the latch after access. Efficient implementation of the latch is crucial, especially in MMDBMSs since transactions have to be performed so rapidly. The latch adopted in Tachyon is characterized as follows: it maximizes system concurrency by supporting both shared and exclusive modes; provides high CPU utilization by minimizing the busy-waiting of transactions; minimizes the number of expensive context switches by reducing the number of semaphore system calls; and finally does not restrict the maximum number of latches.