Resource Containers
Gaurav Banga, Peter Drushcel, Jeffrey Mogul
OSDI 1999

Summary: Makes the case for an explicit OS resource principal in lieu
of the process or thread.

Main points:

* Using processes or threads as a resource principal is a poor match
  for many applications (multithreaded web servers, CGI scripts, etc.)
  Resource containers provide a mechanism to dynamically bind threads
  and processes into a single resource principal.

* Use LRP techniques to bind incoming packets to resource containers
  as soon as possible.

* Resource containers can be arranged in a hierarchy.  However,
  (presumably) a single scheduling policy applies to the entire system.

Notes:

* Current systems provide decent CPU accounting, and minimal memory
  accounting.  Other resources (disk bandwidth, network bandwidth,
  etc.) are completely unaccounted for.  These issues are not
  addressed by this work.

* The system claims to account for kernel state such as protocol
  control blocks and network buffers.  I'm not convinced that they
  actually do this.


Bottom line: resource containers provide a mechanism for creating
explicit resource principals.  However, existing resource management
policies are lacking.  Also, I believe that kernel state and
processing done on behalf of processes can be significant.