COD: Database / operating system co-design

Trends in multicore processors and hardware virtualization pose severe structural challenges to system software in general, and databases in particular. On the one hand, machines are becoming increasingly heterogeneous in terms of, i.e., cache hierarchies, processor interconnect, instruction sets, etc., thereby making it increasingly difficult to develop optimal software for all possible platforms. On the other hand, virtualization forces databases to share resources with other applications without the databases having any knowledge about the run-time conditions. As part of a long-term effort in our group to revisit the entire software stack of the data center, in this paper we explore how to enhance the interaction between databases and OS to allow the database better cope with varying hardware characteristics and runtime conditions. The goals is to integrate the database’s extensive internal knowledge of its own resource requirements –including cost models– into the system-wide and run-time view of the hardware configuration and application mix available inside the OS.