Comments on: Architecting The Internet Data Center: Pt II http://storagemojo.com/2006/07/26/architecting-the-internet-data-center-pt-ii/ Data storage info & analysis Tue, 07 Feb 2012 16:02:02 +0000 hourly 1 http://wordpress.org/?v=3.3.1 By: Robin Harris http://storagemojo.com/2006/07/26/architecting-the-internet-data-center-pt-ii/comment-page-1/#comment-4377 Robin Harris Thu, 27 Jul 2006 19:57:15 +0000 http://storagemojo.com/?p=202#comment-4377 David, good to hear it isn't hardwired into ZFS. I guess I have to amend my comment to: In 10 years the ZFS team will be real glad they made block size a variable. David, good to hear it isn’t hardwired into ZFS. I guess I have to amend my comment to:

In 10 years the ZFS team will be real glad they made block size a variable.

]]> By: David Magda http://storagemojo.com/2006/07/26/architecting-the-internet-data-center-pt-ii/comment-page-1/#comment-4376 David Magda Thu, 27 Jul 2006 18:47:43 +0000 http://storagemojo.com/?p=202#comment-4376 I believe the 128 KB block size is the current default. According to the tutorial presentation at: http://www.sun.com/software/media/real/zfs_learningcenter/high_bandwidth.html they can go up to 32MB (don't remember off-hand which segment they mention it in). Given the design though put into ZFS, I'd be very surprised if they had such arbitrary limits. Doing some digging, the current minimum write size is 4KB, and maximum is 128KB, but both values are defined by variables that could be easily changed (ZIL_MIN_BLKSZ and ZIL_MAX_BLKSZ, respectively) : http://blogs.sun.com/roller/page/realneel?entry=the_zfs_intent_log I believe the 128 KB block size is the current default. According to the tutorial presentation at:

http://www.sun.com/software/media/real/zfs_learningcenter/high_bandwidth.html

they can go up to 32MB (don’t remember off-hand which segment they mention it in). Given the design though put into ZFS, I’d be very surprised if they had such arbitrary limits.

Doing some digging, the current minimum write size is 4KB, and maximum is 128KB, but both values are defined by variables that could be easily changed (ZIL_MIN_BLKSZ and ZIL_MAX_BLKSZ, respectively) :

http://blogs.sun.com/roller/page/realneel?entry=the_zfs_intent_log

]]> By: Robin Harris http://storagemojo.com/2006/07/26/architecting-the-internet-data-center-pt-ii/comment-page-1/#comment-4370 Robin Harris Thu, 27 Jul 2006 05:00:35 +0000 http://storagemojo.com/?p=202#comment-4370 Good question David. I thought about interweaving some of the ZFS info, and thought better of it since it isn't, AFAIK, in use at any IDCs. Yet you are correct, ZFS uses a number of techniques to conserve I/Os, including: <ul> <li>All writes are full stripe, thanks to RAID-Z's variable width stripes</li> <li>Writes are ganged together for maximum performance</li> <li>As you noted, no overwrites, which also contributes to resiliancy, along with a maximum 128k block size</li> </ul> In 10 years I think that 128k block limit will be one of the few things the team will wish they'd made much larger. Good question David. I thought about interweaving some of the ZFS info, and thought better of it since it isn’t, AFAIK, in use at any IDCs. Yet you are correct, ZFS uses a number of techniques to conserve I/Os, including:

  • All writes are full stripe, thanks to RAID-Z’s variable width stripes
  • Writes are ganged together for maximum performance
  • As you noted, no overwrites, which also contributes to resiliancy, along with a maximum 128k block size

In 10 years I think that 128k block limit will be one of the few things the team will wish they’d made much larger.

]]> By: David Magda http://storagemojo.com/2006/07/26/architecting-the-internet-data-center-pt-ii/comment-page-1/#comment-4363 David Magda Wed, 26 Jul 2006 23:24:14 +0000 http://storagemojo.com/?p=202#comment-4363 Regarding I/O costs and the characteristics of Google's system, does Sun's ZFS have the same (or similar) advantages? Most writes are in large chunks (usually 128kB), and occur on any free block--they don't overwrite the data that they're replacing--so you possibly save on seek time. Regarding I/O costs and the characteristics of Google’s system, does Sun’s ZFS have the same (or similar) advantages? Most writes are in large chunks (usually 128kB), and occur on any free block–they don’t overwrite the data that they’re replacing–so you possibly save on seek time.

]]>