Review: Anderson et.al., Scheduler Activations

From: Steve Arnold (stevearn_at_microsoft.com)
Date: Mon Jan 26 2004 - 14:35:41 PST

  • Next message: Chuck Reeves: "Scheduler Activations: Effective Kernel Support for the User-Level management of Parallelism."

    In this paper, the authors discuss what seems to be their invention,
    "scheduler activations." Briefly this is a hybrid kernel/user-mode
    approach to thread management. Their implementation is on a
    non-mainstream machine, but the concepts seem to apply to many other
    operating systems.

     

    The authors assert that current thread management schemes do not perform
    (and are not managed well) at neither the kernel nor the user level.
    When they are management completely at the user level, they do not have
    a real-world view of the underlying system, and thus cannot make a good
    implementation. However, when managed at the kernel level, there is too
    much overhead in using them for typical parallel programs.

     

    With scheduler activations the authors aim to "demonstrate that the
    exact functionality of kernel threads can be provided at the user
    level." However, they need an enhancement to the system in order to do
    this effectively. This is where activations come in. The activations let
    the user control starting of stopped threads (among other things),
    whereas this was previously done by the kernel. "Upcalls" are made for
    relevant processor allocation events. The authors show in test results
    that they can achieve better performance with these activations.

     

    It seems to me that there might have been a debate up to this point
    about how and where threads should really be managed. The authors offer
    a solution. This must be quite successful as it has subsequently been
    implemented on other OSes. Does this method require good user code to
    use it effectively?

     


  • Next message: Chuck Reeves: "Scheduler Activations: Effective Kernel Support for the User-Level management of Parallelism."

    This archive was generated by hypermail 2.1.6 : Mon Jan 26 2004 - 14:51:04 PST