Active Storage Networks
-----------------------
David Nagle and Greg Ganger
Carnegie Mellon University


Delivery of persistent data is the dominant use of network bandwidth,
in both LAN and WAN environments. Despite this fact, network
implementations and storage services continue to ignore the particular
needs, characteristics and capabilities of their partners in network
storage delivery.  For example, current network infrastructures do not
support Network-attached Storage (NAS) in providing redundancy, secure
management, and striping, often forcing NAS back to central server
bottlenecks. Worse, network-attached storage components and future
networking technologies (e.g., Active Networks) are evolving
independently, perhaps in counterproductive ways and certainly not in
collaborative ways.

We believe, however, that these two core infrastructural technologies
(i.e., NAS and Active Nets) can assist each other to enhance existing
services and provide new ones.  Fundamentally, both attempt to improve
system-level performance, reliability, scalability, and extensibility
by leveraging the almost-free and rapidly growing wealth of
device-based processing cycles and memory.  This work proposes to
evolve both technologies in a complementary way, creating a consistent
programming interface that simplifies development, enables migration
of functionality, and allows each technology to help overcome the
limitations in the other.

For example, client-based RAID over NAS suffers significant
performance and/or reliability problems in heavily-loaded or
long-latency networks.  Migrating core RAID functionality into the
network (e.g., switches, routers) positions key functions in more
central locations much closer to storage. This can increase
reliability and performance without having to resort to a server-based
RAID controller.  NAS can also help Active Networks.  For example, if
the network does not trust functions provided by a client, network
components that integrate NAS's security model can turn directly to
NAS for trusted code, providing exactly the same degree high-level of
security as NAS provides.

As a another example, consider data types that place real-time
constraints on data movement through the network and within storage.
An Active Network that understands the scheduling requirements should
be able to ensure timely data delivery. However, if the data source
(i.e. storage) is not ready to transmit or receive the data, then
schedules will fail. Likewise, NAS attempts to solve this problem by
integrating real-time scheduling into storage, but will fail if the
network is unable to deliver data. If Active Nets and Network-attached
Storage are integrated, they could cooperate to provide end-to-end
delivery guarantees. Further, the two could negotiate (and
renegotiate) buffer sizes and transmission periods based on observed
and expected traffic patterns, allowing each more flexibility for
local optimizations.