15 Comments

Summary:

The latest idea from Facebook on how to revolutionize the data center is elegant in its simplicity: put a switch on hard-disk drives that slows their speed when their data is no longer hot. I wonder if Facebook will take up the cause and build it.

hard_disk_head_on_platter

The latest idea from Facebook on how to revolutionize the data center is elegant in its simplicity: Put a switch on hard-disk drives that slows their speed when their data is no longer hot. I’m not aware of any vendor actually doing this, which makes me wonder if Facebook will just take up the cause itself.

Writing on the Open Compute blog, Facebook Engineer Eran Tal makes an astute observation. Essentially, he says, HDDs must remaining spinning even when their data is no longer accessed often because should someone need something, the 30 seconds it might take for disk to resume spinning might seem like an eternity. The natural response to this point might be to suggest buying slower disks, but, as Tal points out, disks don’t start off cold:

HDDs however aren’t “born” cold; they progress into that state. In early stages, when user data is constantly being uploaded, or when data is recovered to the system due to a failure on a different machine, high disk bandwidth is a valuable asset. However, when a system’s data turns cold, there is no value to the high bandwidth.

Solid-state drives might be one solution to the problem, but they’re still relatively expensive, especially at scale. Tal has another idea, which is to put a switch on HDDs that could be flipped to slow their speeds when their data becomes cold. According to his calculations, a 3TB SATA HDD reduced from 7,2oo RPM to 3,600 RPM would cut its power consumption from 7W to 3W.

As the post’s commenters point out, the idea of dynamic-speed HDDs isn’t entirely new (see this IEEE paper from 2003, for example), and the tactic wouldn’t work in every situation. As one commenter notes, on a site such as Amazon.com, where items such as textbooks receive little attention except with the start of each new semester, efforts at changing an HDD’s speed would have to be an ongoing process.

But the thing about Facebook — and Google, Yahoo and other large web sites — is that they innovate based on their unique needs, not what’s appropriate for mainstream users. Let’s not forget, this post appeared on the Open Compute blog, part of a project based around Facebook’s custom-built servers and data centers. If Facebook actually sees real value in this idea, it wouldn’t surprise me to see the specifications for its flipswitch HDDs show up in the Open Compute library at some point.

If that happens, those mainstream users can assess the designs and decide for themselves if they’ll follow in Facebook’s footsteps like some already have with servers.

  1. Greg Knieriemen Friday, August 19, 2011

    Nexsan has been doing this in arrays for years. This is not a new technology at all.

    Share
  2. Sounds very interesting but haven’t this been around for a while?

    Share
  3. “the tactic wouldn’t work in every situation. As one commenter notes, on a site such as Amazon.com, where items such as textbooks receive little attention except with the start of each new semester” Unless…you write a program/bash script that analyzes when data is accessed. You could easy write something that analyzes time of day, day of week, and week of year to determine when data is commonly accessed. For example, data is probably accessed much less at midnight than at noon so you could slow it down for 2 hours each day. You could also then group data with common access patterns thus ensuring a hard drive could be turned down for the maximum amount of time.

    Just the thoughts of a non-hardware engineer.

    Share
    1. “For example, data is probably accessed much less at midnight than at noon so you could slow it down for 2 hours each day”

      The busiest period for one UK broadband outfit I worked with was 11pm to 2am; it’s amazing how many people come home from a night out and do some surfing, online game playing or browsing for things they have been chatting to their friends about.

      Share
    2. It’s not gonna be easy to define this period of low-activity (that’s how I read “midnight”) in a globalized world. Facebook is not limited to a certain geography so one can expect (though it is still certainly worth collecting relevant stats) that while there are big peaks, there are not really big troughs bringing the access down to a grind…I still like the idea that using access stats, one could slow down or rev up the hard discs as needed. Or may be even have farms spread across geographies .. rather than have farms all in one timezone/geographic zone.

      Share
  4. Greg Glockner Friday, August 19, 2011

    I thought most consumer SATA drives can do this already. Wondering why people haven’t done this on the servers. An interesting additional question is to group the data in terms of frequency of access.

    Share
  5. Thanks for the hat tip Greg, yes, our arrays added this feature about 4 years ago. We call it AutoMAID. Many of the drives we sell support a half-speed standby mode which we call AutoMAID Level 2.

    Share
    1. Derrick Harris Friday, August 19, 2011

      Interesting, I hadn’t heard of AutoMAID before. Can I ask whether this is a temporary standby mode, or something more permanent like FB was talking about? Or isn’t there a difference, technically?

      Share
      1. In AutoMAID terms, this is suggesting Level 1 latencies with Level 2 power savings. http://www.nexsan.com/library/automaid.aspx

        Share
  6. Vladimir Rodionov Friday, August 19, 2011

    We can cut power even more by turning these servers (with cold data) off.

    Share
  7. Western Digital’s ‘Green Power’ HDDs already do this; their IntelliPower algorithm varies the RPM

    Share
  8. Scott Elliott Friday, August 19, 2011

    My experience working with the HDD makers is that even at a million units a month they are hard pressed to do anything custom. They are in a brutal commodity business with razor thin margins. Adding even one second to their test flow can be a deal killer.

    Share
  9. Spin-down technology already exists. Not new at all.

    Share
  10. Peter Petersen Sunday, August 21, 2011

    Our Dell Equallogic SAN’s spin down inactive disks. Nothing new to see here folks!

    Share

Comments have been disabled for this post