The Internet Backplane is a synthesis of trends in distributed
operating systems [cite], from the World Wide Web, and from the
development of the Network Computer and other appliances which rely
on network resources for basic services.  It builds on the fact that
the new high-performance networks that underlie computational grids
provide far more reliability than the commercial Internet can.  That
reliability gives us a level of control that makes it possible to
manage the resources shared by closely coupled distributed operations
in ways that were previously difficult, if not impossible.  It
enables us, in other words, to consider the global network as an
extension of the processor backplane, if we only had a low overhead
mechanism for fine-grained naming and access to data, analogous to
physical addresses and bus transfers.  It is this analogy which gives
rise to the notion of an Internet Backplane: a common namespace for
fine-grained management of distributed resources.  Under the UT
NetSolve Research Grid we will explore the use of the Internet
Backplane to implement distributed storage services within NetSolve
(e.g.advanced scheduling and fault-tolerance capabilities) and
explore its use in advanced grid applications generally.
-
The first requirement allows IBP names to be passed across the
network without modification, and so they can be freely embedded in
application level protocols and data structures.  The World Wide Web
addresses this requirement through the use of URLs.  Uniform
interpretation of URLs is achieved by making all names relative to an
IP address or Internet domain, which are interpreted with a high
level of consistency throughout the Internet.  IBP names have the
same structure as URLs, consisting of an IP address and a
locally-interpreted name.  However, IBP names have more complicated
semantics associated with them than URLs and are not designed to be
used manually in documents.
-
The second requirement means that the API must not distinguish
between different kinds of IBP names.  A complete set of storage
operations must be enabled on the network, and in particular it must
be possible to modify data objects stored anywhere in the network.
Dereferencing must not require additional information such as a
storage resources must be very flexible and must allow storage to be
-
used quite freely.  User accounts and passwords are not ruled out in
the allocation of storage, but their use should not be so universal
or so restrictive as to unduly limit the usability of the system.
One of the assumptions of the IBP design is that storage is cheap
enough and plentiful enough to allow it to be shared more freely than
it is currently in order to facilitate communication.  The IBP
allocation scheme must be flexible enough to allow storage resources
to be shared as freely as desired with the level of security required
by the user.
-
NetSolve will make use of IBP to gain control the location of data
objects in order to better schedule computation tasks.  By allocating
storage and managing copies of data objects (e.g. numerical matrices,
simulation input images, checkpoint files, etc.) the NetSolve agent
can implement application-specific policies regarding data locality,
such as caching or movement of data directly between servers without
being returned to the NetSolve client between calls.