Carnegie Mellon and Intel Labs researchers have built an experimental computing cluster they say has the potential to radically cut power usage in datacentres such as those used by Amazon and Facebook, while meeting the same capacity, availability, throughput and latency requirements.
The cluster, called Fast Array of Wimpy Nodes (Fawn), is made up of large numbers of nodes of embedded processors, such as those used in netbooks, combined with between 2GB and 16GB of flash memory.
The researchers claim the resulting system is cost-competitive with other high-throughput approaches, such as the use of large numbers of hard-disk drives (HDDs), or systems based on DRAM memory, while consuming less power. DRAM memory is faster, but more costly, than flash.
Flash memory-based storage is becoming increasingly popular in enterprises as a way of serving frequently accessed data faster and more efficiently than systems based on HDDs, according to industry analysts. Enterprise storage systems using this technology introduced this month include a flash storage array from Sun and a storage virtualisation product from IBM that integrates SSDs.
The Fawn project is led by David Andersen, Carnegie Mellon University assistant professor of computer science, along with Michael Kaminsky, senior research scientist at Intel Labs Pittsburgh. They detailed Fawn in a paper presented on 12 October [PDF] at the 22nd Association for Computing Machinery (ACM) Symposium on Operating Systems Principles in Montana.
Fawn is designed for massively parallel, input/output-intensive, high-load systems that need to retrieve small items, such as those used by Amazon, LinkedIn and Facebook for objects such as thumbnail images, wall posts or Twitter messages, Andersen and Kaminsky said.
"Unfortunately, small-object random-access workloads are particularly ill served by conventional disk-based or memory-based clusters," they wrote in the paper.
"The poor seek-performance of disks makes disk-based systems inefficient in terms of both system performance and performance per watt. High-performance DRAM-based clusters, storing terabytes or petabytes of data, are both expensive and consume a surprising amount of power — two 2GB DIMMs consume as much energy as a 1TB disk."
Power usage is becoming an increasing part of the cost of datacentre operations, the researchers noted. "Future datacentres may require as much as 200 MW, and datacentres are being constructed today with dedicated electrical substations to feed them," they wrote.
The researchers found that a prototype 21-node Fawn system using 500 MHz AMD Geode LX embedded processors, network and support hardware, could serve up to 1,300 256-byte queries per second per node while consuming less than five watts.
The prototype achieved 364 queries per joule, more than an order of magnitude better than traditional disk-based clusters, Andersen and Kaminsky said.
Experiments with newer processors found that Fawn-type systems could achieve even better efficiency, the researchers said.
"Our preliminary experience using Intel Atom-based systems paired with Sata-based flash drives shows that they can provide over 1,000 queries per Joule, demonstrating that the Fawn architecture has significant potential for many I/O-intensive workloads," they wrote.
The project is backed by money from the National Science Foundation, Google, Intel and Network Appliance.