DuckCorp » DuckCorp Infrastructure

• Lian Li EX-30
• Lian Li EX-H34B
• Mediasonic HF2-SU2S2
• Addonics MST4ESA-B

• Cooler Master STB-3T4-E3-GP (vendor info)
• Lian Li EX-332B (vendor info)
• SilverStone CFP51 (vendor info, review)

and then plug it into an old spare case with a 300W PSU (probably Enermax), with a hack to start the PSU, it could be a cheap solution.

• buy 4 * Hitachi Deskstar 7K2000
• buy necessary cables for e-SATA<->SATA conversion and being able to plug things into the external box
• buy what is neccesary for #110
• perhaps buy one rack for external cooling

• chunk size 256KB
• block size 4kB
  seems to give the best performance.

stride = chunk / block = 64kB
stripe-width = stride * ( (n disks in raid5) - 1 ) = 192kB

RAID creation: mdadm --create /dev/md0 --verbose --chunk=256 --level=5 --raid-devices=4 /dev/sda /dev/sdb /dev/sdc /dev/sdd
fs creation: mkfs.ext3 -v -m .1 -b 4096 -E stride=64,stripe-width=192 /dev/md0

Should we stay in ext3 ? ext4 seems quite stable on Orfeo. XFS seems to be ok again. dunno...

• https://raid.wiki.kernel.org/index.php/RAID_setup#RAID-4.2F5.2F6
• http://wiki.debian.org/DebianInstaller/SoftwareRaidRoot

As for the bootloader, GRUB2 should work well, but care about the bios_grub partition: http://grub.enbug.org/BIOS_Boot_Partition

• mirrored loader, so be able to boot a degraded but not broken RAID 5 (see this page for RAID1 as an example)
• mirrored bios_grub partition (no idea)

• http://grub.enbug.org/

I wonder if this new NAS should be merged with Daneel. Having more backup space is great, but it defeats the concept fo backup for the important data on the NAS, and backup runs would probably be a nuisance for the NAS performance (and not all backups occurs while we are sleeping). Perhaps optimizing the backup on Toushirou to exclude not-so-important things could help Daneel stay the same, and, later, reinstalled on 2TB disks ?

I read LVM is now able to align containers on RAID properly, but don't know how to check it.

According to my tests, ext4 is able to calculate stride and stripe-width automatically.

Unfortunately, it seems there is a bug in PV size autodetection, and only 5.46TB out of 6TB are assigned. It's quite a lot of space lost, so i'm investigating. mdadm --detail /dev/md1 gives:

     Array Size : 5860071168 (5588.60 GiB 6000.71 GB)

  PV         VG          Fmt  Attr PSize PFree  
  /dev/md1   Daneel_main lvm2 a-   5.46t 449.10g

Stats (bonnie++):

Version  1.96       ------Sequential Output------ --Sequential Input- --Random-
Concurrency   1     -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
Machine        Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec %CP
Daneel           4G   370  98 150080  28 96812  20  1655  98 358202  35 628.0  11
Latency             22100us    1002ms    1124ms   12510us   37295us   76125us
Version  1.96       ------Sequential Create------ --------Random Create--------
Daneel              -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete--
files:max            /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP
       128:131072:0   877  13   112   1 11067  47   565   8   106   1  5034  23
Latency              5566ms    2576ms     824ms   67313ms    3054ms    1264ms
1.96,1.96,Daneel,1,1281615311,4G,,370,98,150080,28,96812,20,1655,98,358202,35,628.0,11,128,131072,,,,877,13,112,1,11067,47,565,8,106,1,5034,23,22100us,1002ms,1124ms,12510us,37295us,76125us,5566ms,2576ms,824ms,67313ms,3054ms,1264ms

Using the mkfs largefile option may be interesting for the partition dedicated to stuffZ.

Seems the default for mdadm chunk is now 512KB, and LVM VG extend size matches, so i guess it is not needed to use special options anymore. Nevertheless, not sure about alignments exactly.