Performance Isolation: Sharing and Isolation in Shared-Memory
Multiprocessors
Ben Verhese et al

Goal: Provide resource fairness to different tasks without sacrificing
throughput 

Main points:

* The resource principal is a software performance unit--conceptually
  similar to a resource container.

* Each SPU maintains three resource levels for each SPU: entitled
  (fair share allocation), allowed (entitled + idle resources), used
  (resources in use).

* A key factor in sharing policy is the revocation cost: how much it
  costs to transfer resources between SPUs.

* Processor allocation: high priority processes are loaned idle CPUs
  - assume that there are fewer active SPUs than CPUs (?), so
    time-sharing is not a big issue.


* Disk bandwidth is approximated by counting the sectors transfered in
  an interval, and decaying this count periodically.
  
  - Only modify C-SCAN if an SPU exceeds its fair share by some
    threshold. 

* Synchronization points can lead to priority inversion (inode lock,
  page table lock).

Comments:

* Shared resources and kernel resources are difficult to account for.
  This work creates two special SPUs: shared and kernel.
  Malicious/buggy processes could take advantage of this.

* What prevents a malicious process from increasing its priority to
  gain CPUs (this probably isn't in their threat model).

* No description of how the system finds idle pages.  Thrashing is not
  considered here...  This implies that each process's working set
  fits inside its fair share allocation.