Grid Vendors Shut Down Whole Nodes To Be Really Green

As I talk to vendors about storage solutions for non-OLTP applications, from backup and archiving to supporting massive, object-based Web applications like photo sharing, I've been seeing more solutions based on the RAIN (Redundant Array of Independent Nodes) architecture.Rather than having dedicated controllers and disk shelves, they use an array of "industry standard" 1 or 2U servers for both intelligence and storage. They typically have a small number of access nodes that other systems talk to via standard protocols like CIFS and NFS or the vendor's API. The access nodes slice, dice, and julienne the data before pushing it out to a larger set of storage nodes. So I have a question to these vendors "Why not shut down storage nodes when they're not in use?"

The RAIN architecture has a lot going for it. Unlimited, fine-grained scalability being the biggest, but using a pair of Xeons for every 4 1 TB drives as a typical RAIN of 1U servers does means you're going to use a lot of power.

As I talked to vendors, some hinted that they may be implementing MAID (Massive Arrays of Inactive Disks), which spins down drives that aren't in use, but none had plans for node shutdowns. Western Digital's Raid Edition drives draw 11.5W when active and 8.5W when idle, so spinning them down to their 1W sleep state would reduce a systems power draw.

But in a RAIN system, disk drives are a small part of the power story. The Xeon 5000 processors in most 1U servers draw 40W to 130W each, depending on speed and number of cores, so in even the most anemic single-processor server the processor draws as much juice as all four drives together. Add in memory, Ethernet interfaces, and the rest of the motherboard, and we're talking real juice here.

Since 1 and 2U servers almost always include IPMI and/or Wake on LAN, it would be simple for an access node to wake up a storage node when it needs access to the node's data. Throw in a wake from hibernate mode that could bring the system up in a minute or two and I ask again: "Why not shut down storage nodes to save energy?"