From: Jenny Liu (jenny.liu@cal.berkeley.edu)
Date: Wed Oct 27 2004 - 03:06:38 PDT
"Explicit Allocation of Best-Effort Packet Delivery Service" describes a
framework for providing different classes of service for different users
based on pre-allocated capacities (stated in terms of throughput,
geographic scope, and probability of assurance). The framework involves
profile service meters placed throughout the network that read/tag
packets as being within specified profile or not, and differential
packet dropping at routers. The paper then presents simulation results
and discusses dealing with non-responsive connections.
The proposed framework is flexible. Profile meters can be placed in
many different places within the network and will still work. Profile
meters along the edges of the network can be modified to affect
administrative change in the classes of service. The framework works
with TCP (rather than strictly on top of TCP) to avoid slow-start and
increase utilized bandwidth.
However, in most cases it is probably unreasonable for a single entity
to make absolute guarantees about availability of resources on the
global internet because the global internet is not controlled by any
single entity and things can go wrong out of the control of any given
entity that can still affect the service provided by that entity. The
framework also makes use of RIO built on top of RED, which suffers the
same downfalls as RED. In particular, a router running RIO drops
packets with some probabilty indepedent of how many connections are
using that router, when the router actually needs to drop a number of
packets equal to the number of connections in order to make an
appropriate aggregate back off and prevent congestion.
The provided framework is a good incremental improvement of what's
currently out there, but could be improved upon by starting over and not
building on top of components (TCP, RED) that are fundamentally flawed
to begin with.
This archive was generated by hypermail 2.1.6 : Wed Oct 27 2004 - 03:06:38 PDT