Disco: running commodity operating systems on scalable multiprocessors

Reviewer: Armin B. Cremers

Virtual machine monitors, a popular operating systems approach in the 1970s, are investigated as a means of extending modern system software so it can run efficiently on large-scale shared-memory multiprocessors without a massive implementation effort. Virtual machines add an intermediate level between multiple copies of commodity operating systems and the scalable multiprocessor hardware in order to hide certain novel attributes of the machines, such as its size and aspects of the nonuniform memory architecture (NUMA). As part of the Stanford FLASH shared-memory multiprocessor project, a “NUMA-aware” prototype implementation called Disco has been developed to combine commodity operating systems into a new performant system software base. In a simulation-based experiment, this prototype was used to run multiple copies of SGI's IRIX operating system. The results show that many traditional problems of the virtual machine approach do not occur in this approach. Disco succeeds in supporting a global buffer cache functionality that is transparently shared across all virtual machines. This is achieved by combining a suitable emulation of the DMA engine with standard distributed filesystem protocols. There are indications that many of the techniques also apply to more loosely coupled environments, such as networks of workstations.