Mobile computing processors and flash memory are the basis
of a new, energy-efficient server architecture created by researchers at Carnegie
Mellon University
and Intel Labs Pittsburgh. Cutting data center energy use is imperative for
companies whose electricity bills are skyrocketing thanks to ever-higher chip
densities and ever-hotter operating temperatures.
Fast Array of Wimpy Nodes (FAWN) consists of 21 nodes, each
with a low-cost, low-power, off-the-shelf processor and a 4GB compact flash
card. At peak utilization, the Linux-based cluster operates on less energy than
a 100-watt light bulb, yet can perform up to 100 times as many queries using
the same amount of energy as a conventional disk-based cluster, according to
the researchers. Crafting software to leverage the capabilities of these small
processors is the FAWN project’s top mission.
“Dave [Andersen] and
I are focusing on the software—how to build system-level software and
applications to exploit these properties,” said Michael Kaminsky, senior
research scientist at Intel Labs Pittsburgh. Kaminsky and Andersen, a Carnegie-Mellon
assistant professor of computer science, are leading the FAWN project.
FAWN is ideally suited for some kinds of processing, yet
ill-suited to other kinds.
The architecture’s strong suit is handling key-value storage
systems, in which small pieces of information must be accessed quickly.
Key-value storage is used by social networking sites such as Facebook.com.
Because FAWN operates according to the same principles as parallel processing,
it is less well-suited for computing tasks that are best done by a single
processor.
Some types of cloud-based computing could benefit from the
FAWN architecture, said Andersen. “There is an aspect of cloud computing that
is very suitable for this. Amazon, for example offers a cloud storage service,
cloud database and cloud compute service. Of those three, the storage and
database services will work well on FAWN, but not necessarily the compute
services,” said Andersen.
Data-intensive analytic applications could also benefit from
the architecture because FAWN helps relieve bandwidth limitations in getting
data in or out of the CPU, said Andersen. Software built for distributed
processing, such as Hadoop, could benefit as well, he added.
The current FAWN cluster uses 500MHz AMD
processors. The team is currently at work on a 100-node FAWN cluster built with
Intel dual-core 1.6GHz Intel Atom processors, which are commonly used in handheld
devices and netbook computers.
The work was supported in part by gifts from Network
Appliance, Google and Intel, and by a grant from the National Science Foundation.
