The systems in a shared

hmultiprocessors (CMPs) containing two to eight cores with support for up to eight hardware thread contexts per core. In the next decade we anticipate a computing landscape dominated by multi-core chips featuring hundreds of execution contexts. While many database management systems (DBMS) were designed for one- or ever-growing memory capacities, however, further increases the importance of understanding the work sharing/parallelism trade-off as more and more workloads fit in main memory. In order to achieve maximum performance in highly parallel systems the database engine must intelligently exploit work sharing, detecting and avoiding situations where work sharing will be unhelpful. 1.1 To Share, or not to Share? Intuitively, work sharing should provide performance benefits proportional to the amount of work removed from the system. Unfortunately this is not always the case, especially for in-memory workloads. Prior efforts to incorporate explicit work sharing int