Unifying High-speed
Interconnects
Contract No: MEXT-CT-2003-509595
unisix
aims
at building general-purpose computing and storage infrastructure
for the future datacenter that support all IT activities. Today, such
datacenters include various types of servers that support compute,
storage I/O,
client network I/O, and high-speed LAN applications. The requirements
of such,
modern datacenters to support a variety of IT services (mail, dbases,
storage,
backup, compute intensive applications, online services, etc.) lead to
multi-tier architectures that require connectivity among different
levels. One
of the main problems of using multiple tiers in such architectures is
the
specialization of each tier to a specific type of functions or
applications.
For instance, a cluster used for computing cannot be used to service
storage
requests as well. This leads to an imbalance of using the available
resources.
Modern
datacenters would benefit significantly from the ability to
shift resources between tasks. However, this requires that the division
of
subsystems within a datacenter is logical and not physical. To some
degree,
today such uniform and symmetric architectures are possible since most
computing resources within the datacenter are very similar, in most
cases PCs
or workstations. However, these similar nodes are connected through
incompatible interconnects, each specialized to the functions performed
by the
specific subsystem in a datacenter. For instance, the storage area
network that
provides access to scalable storage is different from the system area
network
that connects the cluster nodes. Moreover, the use of multiple,
application
specific protocols and interconnects leads to high management
complexity and
cost, difficulty in upgrading to new generations of each technology,
and, in
some cases, to proprietary solutions due to incomplete standardization.
In our
view, these issues can compromise the future evolution of datacenters
and their
ability to effectively support their constantly increasing loads.
The goal of our proposed work is to
develop the technology required to unify the network interconnects used
today
in datacenters and provide the fundamental technology for evolving to
more
flexible and effective datacenter architectures. We will investigate
how to
replace all these dissimilar components and protocols with an
interconnect that
employs common lower-layer characteristics and provides multiple
protocol
abstractions that can handle different application needs.
