Review of Disco

From: Muench, Joanna (jmuench_at_fhcrc.org)
Date: Mon Mar 01 2004 - 13:42:44 PST

  • Next message: Brian Milnes: "DISCO Review; warning long"

    Bugnion, et al. (1997) revive the concept of virtual machine monitors by
    presenting them as a solution to the lag of operating system capabilities
    behind hardware performance. While implementing a better OS is the ideal
    solution, Disco provides a less expensive alternative with most of the
    performance benefit. The target system is the FLASH shared memory
    multiprocessor, which has the challenge of being a ccNUMA. The paper
    specifically addresses how Disco provides a superior interface to this type
    of memory system.

    Disco is a virtual machine monitor that provides an abstraction layer
    between commodity operating systems and the underlying system. The
    individual components that make Disco possible aren't particularly new, but
    the application is. Probably most important is the use of physical memory,
    which Disco handles with use of dynamic page migration and replication. This
    provides the ability for a non-NUMA aware OS to run efficiently on the
    underlying NUMA architecture. Another important component is that the
    virtual CPU executes on the real CPU, minimizing the performance hit of a
    virtual machine.

    The disappointment of the paper was discovering it was only run on a
    simulation OS, rather than the real thing. The simulation did, however,
    provide some insights into potential performance issues. One of the
    significant overheads in running Disco was the high overhead of servicing
    kernel TLB-faults. The authors thought that larger pages could reduce this
    overhead. The good news was that with only 8 processors the pure IRIX system
    bogged down due to a bottleneck at the spinlock protecting IRIX memory
    management data structures. Partitioning the system into different virtual
    machines reduced the execution time by 60%, a significant improvement.

    The bit I liked most about the paper was the (brief) section on SPLASHOS, as
    an example of being able to run a very lightweight OS efficiently. In
    general this paper, important as it is, proved a bit of disappointment in
    terms of clarity of writing.


  • Next message: Brian Milnes: "DISCO Review; warning long"

    This archive was generated by hypermail 2.1.6 : Mon Mar 01 2004 - 13:43:00 PST