From: Karthik Gopalratnam (karthikg@cs.washington.edu)
Date: Wed Oct 13 2004 - 01:40:35 PDT
End-to-end Arguments in System Design
This paper attempts to identify and reinforce a principle to system
design - that of placing the onus of management of resources and
functionality on the end-points in the system.
The above principle is mainly considered in the context of
communication systems, where the authors provide a whole slew of examples
of network systems and applications where it makes sense to make the end
hosts of a network manage the requirements of specific applications as
opposed to having the lower level network components provide for the needs
of high level appliacations, such as reliability, authencity of data, etc.
The authors make the valid claim, at least for reliability, that having
the lower level network components try to provide reliability to
applications will take away significantly from the performance observed by
the applications, and still does not obviate the need for reliability
mechanisms at the higher layer. This design principle in some sense is
another facet of the flexibility / reliability / performance tradeoff
involved in any design that wants to figure out whether to push
functionality to the lower layers or keep them application specific.
However, the contributions of this paper are fairly sketchy. The
earlier TCP paper from '74 identified the reasons for managing the network
from the ends and it wold seem that this design principle was very clearly
understood in the context of network systems, and was indeed widely
applied as well. The authors' examples for why this principle should hold
seem to merely restate the original ideas involved in TCP-IP. The whole
principle of end-to-end management is really an artifact of the concept of
hierarchy and layering abstraction inherent in most system designs - the
higher layers represented by the end hosts should build on the possibly
limited functionality provided by lower layers to tailor the design to the
specific needs of the application. In that light, this paper mererly seems
to be restating what must have been obvious given two decades of system
design based on abstraction and layering.
This archive was generated by hypermail 2.1.6 : Wed Oct 13 2004 - 01:40:35 PDT