l4-hurd
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: making paging decisions


From: Ludovic Courtès
Subject: Re: making paging decisions
Date: Tue, 26 Oct 2004 21:54:01 +0200
User-agent: Mutt/1.5.4i [Guile enabled]

Hi,

Today, 4 hours, 20 minutes, 37 seconds ago, Neal H. Walfield wrote:
> Indeed.  This is a problem.  What I have in mind is having a per-user
> "policy" combined with a economic model of short and long term loans.
> The former being more tested; the latter more academic.

The GNUnet folks have an enlightening paper that describes a simple, yet
powerful, economic model [0].  Basically, the idea is that "most of the
time", there are more resources available than needed (this may not be
true if we consider physical memory in a running system, unless there is
1 or more GB of RAM or so).  So when enough resources are available, the
server would pretty much allocate as many resources as requested by the
clients.

However, in times of resource shortage, the server makes decisions about
which request to honor based on prior behavior of clients.  The server
can keep track of its clients' behaviors by maintaining a "trust level"
(an integer) for each of them: each client specifies a priority (an
integer) for the requests it makes which represents how much trust it
can "afford" for this request to be served, and the server may decrease
its level of trust in this client by this number.

In the case of physical memory, there still needs to be a number of
guaranteed frames, because otherwise things would be too complicated
(new tasks would have a hard time trying to get physical memory since
physmem would have to reclaim physical memory to other tasks before it
can hand it to new tasks).  However, extra frames could be handled this
way: any task can request physical memory to physmem (it must specify
the priority of its request); under memory pressure, physmem reclaims
physical memory to the task with the lowest "trust level".  This creates
a real incentive to not hold more physical memory than needed.

Of course, there may be important details that escaped me.  :-)

Thanks,
Ludovic.

[0] "An Excess-Based Economic Model for Resource Allocation in
     Peer-to-Peer Networks",
    http://www.ovmj.org/GNUnet/papers.php3?xlang=English




reply via email to

[Prev in Thread] Current Thread [Next in Thread]