FreeBSD Study #10

In last Frieday, FreeBSD Study #10 was held in KDDI Web Commnications' Office.

大きな地図で見る

Today's session is "Managing storage on FreeBSD (vol.2)".

• How to test the data transfer spped of devices
  1. internal kernel

     dd if=/dev/zero of=/dev/null bs=1024x1024 count=80

  2. read only from device da2

     dd if=/dev/da2 of=/dev/null bs=1024x1024 count=80

  3. write only to device da3

     dd if=/dev/zero of=/dev/da3 bs=1024x1024 count=80

  4. read and write data from device da2 to device da3

     dd if=/dev/da2 of=/dev/da3 bs=1024x1024 count=80

• (4) must be slower than (2) and (3)
• limit data transfer spped is which slower one of (2) or (3)

• iostat: statics command of device I/O
  • IOPS
  • data transfer spped
  • %busy
  • svc_t
• gstat: statics command of GEOM I/O (iostat GEOM version)

• The true right way to create gmirror
  1. set gm0 as the mirror of ada1
  2. copy whole ada0 to gm0
  3. set gm0 as boot disk, and reboot
  4. add ada0 to mirror member of gm0, and wait sync between ada0 and ada1
• The way written in FreeBSD handbook is now incorrect :-)

• New file system feature since 9.0 (1) : GPT as default
  • because MBR cannot deal partitions over 2TB
• New file system feature since 9.0 (2) : GPT HAST
  • Highly Available Storage
  • like gmirror over the TCP/IP network
  • deals /dev/hast/XXXX block
  • setting /etc/hast.conf in each machines mirroring

• ZFS
  • architecture of UFS : device -> volume manager -> file system
  • architecture of ZFS : devices -> ZPOOL -> DMU -> ZFS
  • Good purposes for using ZFS
    • many devices
    • dealing petabyte-class data
  • Not good purposes for using ZFS (just a sales talk :-P)
    • speed
    • robustness
    • scalability
  • ZPOOL : Give the name as one block from one or more devices
  • ZFS Data Set : Give namespaces onto the ZPOOL
  • DMU : internal kernel for manipulating dnode
  • dnode : ZPOOL version inode
    • ZFS Data Sets on ZPOOL are dealt like inode
    • Files ans directories on each ZFS Data Sets are dealt like inode
  • ZFS' I/O = copy-on-write
  • ZFS Pros.
    • transactionable
    • lock free
  • ZFS Cons.
    • heavy for dnode
    • updating cost for large size file
  • Tips
    • SHOULD USE amd64
    • set vfs.zfs.cache.size as 0
    • tune vfs.zfs.meta_limit for your machine