Deeply
Embedded Networks
An expanding frontier for computer
scientists lies at the intersection of the logical and physical realms. As
computing elements become embedded more pervasively in our environment, a new
generation of cyber-physical systems arises in which logical processing is
very deeply intertwined with the distributed physical environment in which it
occurs. Computing becomes less obtrusive and a more natural part of the
external world. It becomes more autonomous and less reliant on human input,
intervention, and administration. Physical objects acquire new logical
properties due to embedded computation, sensing, and actuation. Such objects
become increasingly interconnected giving rise to a new global network of
embedded devices. This network is called deeply embedded.
Deeply embedded networks enable several new
applications that enhance our interaction with the physical world. For
example, new medical applications allow reliable and secure integration of
sensors and actuators in smart assisted living facilities to improve elder
care. Deeply embedded networks also enhance social and professional
interaction by enabling new communication experiences such as tele-presence and immersive cyber-physical tele-conferencing. They improve accessibility of
information via wide-area sensor data services. Importantly, they help
advance fundamental knowledge in many environmental, biological, and physical
disciplines by providing access to the physical world in an unprecedented
spatial and temporal granularity. With the above opportunities come numerous
challenges. Deeply embedded networks differ in fundamental ways from the
current Internet and current embedded systems. Important points of departure
including the following:
· Inverted
communication/computing speed trade-off:
Current wired network bandwidth is significant compared to CPU speeds. This
leads to a network architecture that featurs a simple core and an intelligent edge. It fuels,
in part, the end-to-end argument for removing unnecessary functionality from
the common path. In contrast, in deeply embedded networks, the integration
with a natural physical environment often dictates battery-powered operation,
which makes energy the most premium resource. Since (omni-directional)
communication is much more energy-consuming than local computation,
significantly faster CPUs can be afforded than wireless links for the same
energy budget. Consequently, the speed tradeoff point between communication
and computing is inverted. This inversion has non-trivial implications on
network architecture, the end-to-end argument, and application design
decisions.
· In-network computation: Given the overwhelming amount of data that can be
collected in deeply embedded networks from the physical world, a fundamental
function of such networks is reduction of information into a more concise
actionable form rather than point-to-point communication. This implies the
need for in-network computation. In network computation takes good advantage of
the inverted communication/computing speed tradeoff leading to systems where
computation, communication and physical interaction with the external world
are much more tightly linked. This fundamental architectural shift has severe
repercussions on programming abstractions, security mechanisms, data
management tools, and network resource management algorithms.
· Scale: Compared to traditional embedded systems, deeply
embedded networks have a significantly larger scale. The complexity of
composing distributed systems of such a large scale is formidable, motivating
new tools for system composition, troubleshooting and complexity management.
· Inverting
the multiprogramming paradigm:
In-network computation in large deeply embedded networks introduces
significant resource management challenges. Current operating systems (i.e., the typical
resource managers) are designed for devices that are able to perform multiple
concurrent tasks on a single device. This gives rise to familiar
multiprogramming abstractions. In contrast, deeply embedded networks may be
composed of large numbers of individually resource-limited devices. Rather
than executing multiple tasks on each device, multiple devices are needed to
execute any particular network task. Hence, new abstractions and supporting
mechanisms are needed to capture the essence of (network) “tasks”
and “inter-task communication” in the new environment.
Research Areas:
· The inverse problem: Cooperation among multiple simple entities to
achieve a global task is an important research problem in deeply embedded
networks. Many biological systems (e.g., swarms of social insects) perform
non-trivial tasks by following very simple interaction rules. Their
collective perceived intelligence is an instance of emergent behavior. While
one cannot reprogram the behavior of biological systems or rewrite laws of
physics, it is easy to reprogram deeply embedded network nodes. This leads to
what is commonly called the inverse problem. In other words, how to design
the low-level node interaction protocols that would collectively result in a
desired global behavior? A theory is needed for synthesis of globally
efficient, intelligent behavior of choice through aggregation of simple local
actions.
· Application-specific
protocol design: Deeply embedded networks
represent an evolution from traditional communication that parallels a
similar trend in computational devices. In embedded computing, the use of
generic microprocessors has been augmented with use of ASIC components and
more recently with use of FPGAs. This trend is
increasingly towards an application-specific but retargettable
infrastructure. A research question is how to design embedded networks that
are increasingly application-specific yet retargettable
at low cost?
· Programming paradigms: New models and paradigms are needed for
distributed embedded (in-network) computation. New underlying theoretical
foundations are needed to support such paradigms. New programming languages
and distributed middleware tools must be developed around the emerging
abstractions. Network protocols must be accordingly redefined.
· Information reduction: A significant network challenge is to distill
information meaningfully for the network user. Data mining and machine
learning techniques are needed to identify data patterns, learn context, and
act autonomously with little or no human assistance. Networks must suppress
false alarms, yet provide accurate and timely information on problems
inferred from large amount of distributed measurements.
|
