Resource containers: A new facility for resource management in server systems
Gaurav Banga, Peter Druschel, Jeffrey Mogul
OSDI 1999

Why are we reading this paper?

What problem are they solving?
  Your web server is having trouble enforcing priority.
  (Or you want your X-based editor to have priority over compilation?)

What are the basic issues?
  All scheduling points must obey priorities.
  All consumption of resources must be charged to the right activity.

Let's think of points at which scheduling occurs.
  O/S process scheduler.
  Interrupt dispatcher.
  Network protocol processing for incoming packets.
  Disk scheduler.
  Physical page allocator (replacement algorithm).
  Accept queue.
  Web server's main select() loop.

How does the O/S decide how to schedule?
  Hmm, only explicit case is process scheduling.
  Two kinds of common policy:
    Based on CPU time consumed (for fairness).
    Based on absolute priority.

For fairness scheduling, we need to account for CPU time consumed.

Let's think of times when it may be confusing who to charge CPU to.
  Handling of different connections in server process.
  Fast CGI, AMPED helpers, sfsusrv.
  Interrupts.
  Protocol processing.
  Memory consumption (valuable resource, not really CPU).
  Disk I/O usage (valuable resource, not really CPU).

What do they think is wrong with how current O/S charges time?
  Charged to process, not activity.

O/S abstractions designed for "classical applications"
  Process per logical activity.
  But efficiency &c force us to handle many activities per process.

Hmm, two kinds of isolation:
  Memory (address space).
  Performance (scheduling).

Why isn't scheduling *threads* enough?

Real world example:
  I charge things to different accounts.
  I.e. resource principle different from human identity.

Accounting plan:
  Do they give us a list of exactly what is accounted for?
    Hmm, not really, 4.4 is vague.
    4.6 implies containers attached only to threads and sockets.
  Every activity associated with a resource container.
  So extra argument to each sys call?
  Every socket tagged with container ID?
  Processes tell O/S when they switch activities.

Scheduling plan:
  Where are scheduling decisions made?
    They mention thread scheduler and (4.7) protocol processing.
  AND (4.8) incoming SYNs.
    Different prios come in on different accept sockets?
  Interrupt dispatch?
  Network code (but LRP?)?
  Disk scheduler?
  Phys page allocator?

Things that are scheduled based on containers:
  (section 4.6)
  CPU time
  memory
  network QoS

They actually bind a thread to a set of containers.
  The containers the thread has recently executed for?
  This is done by O/S, not by process.

Why do they need a container hierarchy?
  Web server vs nightly backup.
  Within web server, different connections.

What is the hierarchy semantics?
  Parents charged sum of children's usage?

A container looks like a file descriptor!
  Why?

Extend to deal with file system I/O?
  Memory allocation?

Why can only fixed-share containers have children?
  Why not time-shared containers?
  Too hard to decide how to schedule parents?
    E.g. parent X has 10 children, Y has 1.
    How to time-share between X and Y?
  5.1

How do you evaluate a system like this?
  Do we care how fast it runs?
    A little, to see if it adds overhead.
  Big issue is whether it can enforce priorities.
    Overload from many low-pri...
    I.e. are there important things that weren't scheduled or accounted.
    Interrupts?
    Select -- what is the problem in Figure 11?

Can they really resist DoS attacks better?
  Example assumes they know attacker's IP address.
  And that attacker doesn't forge it.

What other aspects of a process might we want to make explicit?
  User ID.
  For access checks.
  So sfsusrv could say on whose behalf it is making a system call.