Upgrading from 6x 500GB RAID5 to 4x 2TB RAID6

Table of Contents

Shutdown server and move to test bench

Prepare new disks

Disconnect old disks

Connect 4 new disks

Enable SMART

  • Samsung disks had SMART disabled by default

    for i sd{b..e} ; do smartctl -s on /dev/$i ; done

Random write

  • Suse FDE
  • Default cipher and mode is "aes-cbc-essiv:sha256"

  • Default keysize is 128-bits

    for i in sd{b..e} ; do dd_rescue -m 10M /dev/urandom /dev/$i ; done

    for i in sd{b..e} ; do echo -n $TEST_PW | cryptsetup luksFormat --key-file=- /dev/$i ; done

    for i in sd{b..e} ; do echo -n $TEST_PW | cryptsetup luksOpen --key-file=- /dev/$i crypt-$i ; done

    for i in sd{b..e} ; do dd_rescue /dev/zero /dev/mapper/crypt-$i & done

    for i in sd{b..e} ; do cryptsetup luksClose /dev/$i crypt-$i ; done

NOTE: dd_rescue doesn't stop when it reaches the end of the output device!

Benchmark

New Kernel

  • Compile new kernel for missing crypto modules

Build and install XTS

$ echo "CONFIG_CRYPTO_XTS=m" >> .config
$ make modules
# cp crypto/xts.ko /lib/modules/2.6.31-pmp/kernel/crypto/
# depmod -A
# modprobe xts
# mount -o remount,rw /boot
# cp .config /boot/config-2.6.31-pmp 
# mount -o remount,ro /boot

Partition types

  • Update parted to 1.8.8

  • DOS partition # cat > sfdisk.tempraid0.format ,1024,fd ,2048,fd

    ^D
# for dev in {b..e} ; do cat sfdisk.tempraid0.format | sfdisk -uM /dev/sd$dev ; done

  • GPT partition (annoying buggy parted means manual label creation) # for dev in {b..e} ; do parted /dev/sd$dev mklabel ; done # for dev in {b..e} ; do parted /dev/sd$dev mkpart primary 0 1024 mkpart primary 1024 3072 ; done

  • Arrays # mdadm --create --verbose --metadata=1.2 --level=raid0 --raid-devices=4 /dev/md66 /dev/sd{b,c,d,e}1 # mdadm --create --verbose --assume-clean --metadata=1.2 --level=raid6 --raid-devices=4 /dev/md67 /dev/sd{b,c,d,e}2 # for dev in md66 md67 ; do mkfs.ext4 -m 0 /dev/$dev && mkdir -p /mnt/$dev && mount /dev/$dev /mnt/$dev ; done

  • Results (some 3908MB, some 3750MB) # bonnie++ -f -d /mnt/md66 -s 3750 -n 0 -u root Version 1.03d ------Sequential Output------ --Sequential Input- --Random- -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 msdos,3908M,,,374639,44,110712,13,,,317465,19,372.9,0,,,,,,,,,,,,, msdos,3908M,,,400310,49,107278,14,,,343700,19,366.2,0,,,,,,,,,,,,, gpt,3750M,,,373408,44,106484,13,,,335955,22,383.3,0,,,,,,,,,,,,,

    gpt-raid0,3750M,,,379201,45,105965,13,,,334808,19,379.7,0,,,,,,,,,,,,,
gpt-raid6,3750M,,,129930,19,47485,6,,,120628,8,360.5,0,,,,,,,,,,,,,

# bonnie++ -f -d /mnt/md67 -s 3750 -n 0 -u root

  • No obvious difference for RAID0

  • Stop # for dev in md66 md67 ; do umount /mnt/$dev ; mdadm --stop /dev/$dev ; done

Cipher and RAID Benchmarks

CHUNK, CRYPT_KEY_SIZE, CIPHER, NCQ
64, 256, lrw-benbi, 1
hastur,3750M,,,77502,10,29570,4,,,76041,5,328.2,1,,,,,,,,,,,,,
hastur,3750M,,,77913,11,29512,4,,,75647,6,322.5,0,,,,,,,,,,,,,
hastur,3750M,,,77937,10,29149,4,,,76300,5,322.9,0,,,,,,,,,,,,,

64, 256, cbc-essiv, 1
hastur,3750M,,,77501,10,29415,4,,,75281,5,330.2,1,,,,,,,,,,,,,
hastur,3750M,,,77880,10,28867,4,,,76466,6,326.1,1,,,,,,,,,,,,,
hastur,3750M,,,78451,10,29881,4,,,73462,5,330.0,1,,,,,,,,,,,,,

64, 256, xts-plain, 1
hastur,3750M,,,76612,10,28949,4,,,76289,5,309.5,0,,,,,,,,,,,,,
hastur,3750M,,,78131,10,29525,4,,,76528,5,317.0,1,,,,,,,,,,,,,
hastur,3750M,,,77624,10,29122,4,,,76401,5,322.7,1,,,,,,,,,,,,,

64, 512, lrw-benbi, 1
hastur,3750M,,,76133,11,28562,4,,,71292,5,289.9,0,,,,,,,,,,,,,
hastur,3750M,,,76470,10,27983,4,,,70690,5,312.1,0,,,,,,,,,,,,,
hastur,3750M,,,74969,10,28450,4,,,69769,5,299.2,1,,,,,,,,,,,,,

64, 512, cbc-essiv, 1
hastur,3750M,,,76839,10,27509,4,,,71383,5,292.8,0,,,,,,,,,,,,,
hastur,3750M,,,74941,10,28484,4,,,71224,5,283.4,1,,,,,,,,,,,,,
hastur,3750M,,,75474,10,29006,4,,,74678,5,307.6,1,,,,,,,,,,,,,

64, 512, xts-plain, 1
hastur,3750M,,,76980,10,28483,4,,,77147,6,321.1,0,,,,,,,,,,,,,
hastur,3750M,,,78038,10,28926,4,,,75617,5,331.4,0,,,,,,,,,,,,,
hastur,3750M,,,77566,11,29640,4,,,73846,5,326.2,0,,,,,,,,,,,,,

128, 256, lrw-benbi, 1
hastur,3750M,,,76583,10,27767,4,,,78447,6,328.8,0,,,,,,,,,,,,,
hastur,3750M,,,76805,10,28012,4,,,78318,5,325.3,1,,,,,,,,,,,,,
hastur,3750M,,,75651,10,28269,4,,,80328,5,319.7,0,,,,,,,,,,,,,

128, 256, cbc-essiv, 1
hastur,3750M,,,73652,10,27561,4,,,76043,5,312.7,0,,,,,,,,,,,,,
hastur,3750M,,,74091,10,27321,4,,,74952,5,302.3,0,,,,,,,,,,,,,
hastur,3750M,,,74929,10,26958,4,,,75964,5,308.0,1,,,,,,,,,,,,,

128, 256, xts-plain, 1
hastur,3750M,,,73128,10,27043,3,,,74821,5,316.5,0,,,,,,,,,,,,,
hastur,3750M,,,73468,10,26747,3,,,73302,5,292.7,0,,,,,,,,,,,,,
hastur,3750M,,,73360,10,26452,3,,,72755,5,297.9,0,,,,,,,,,,,,,

128, 512, lrw-benbi, 1
hastur,3750M,,,74233,10,27291,4,,,74660,5,310.1,1,,,,,,,,,,,,,
hastur,3750M,,,74540,10,26414,4,,,73114,5,288.7,1,,,,,,,,,,,,,
hastur,3750M,,,73015,9,27081,3,,,73988,5,303.1,0,,,,,,,,,,,,,

128, 512, cbc-essiv, 1
hastur,3750M,,,74490,10,27901,4,,,74019,5,292.8,0,,,,,,,,,,,,,
hastur,3750M,,,75435,10,27576,4,,,73163,5,308.6,0,,,,,,,,,,,,,
hastur,3750M,,,74645,10,27287,4,,,73003,5,316.2,1,,,,,,,,,,,,,

128, 512, xts-plain, 1
hastur,3750M,,,75475,10,28163,4,,,78059,6,340.1,1,,,,,,,,,,,,,
hastur,3750M,,,75484,10,27796,4,,,76429,5,321.1,0,,,,,,,,,,,,,
hastur,3750M,,,75956,10,27038,4,,,75931,5,311.3,1,,,,,,,,,,,,,
  • NCQ=1 is significantly faster
  • sdc and sde doing significantly more work

Fix IO load problem

  • Even after rearranging disks in the array, the same disks (Samsungs) have roughly double the TPS on writes. chunk=64 :: key_size=512 :: cipher=aes-xts-plain :: RA /dev/mapper/crypt-md67=256 /dev/md67=512 raw=128 hastur,3750M,,,70425,9,26040,3,,,63046,4,295.2,0,,,,,,,,,,,,,

  • RAID5, for comparison hastur,3750M,,,80460,10,29684,4,,,76040,5,270.2,0,,,,,,,,,,,,,

  • RAID0, for comparison hastur,3750M,,,99602,12,40017,6,,,94923,7,291.6,0,,,,,,,,,,,,,

  • Taking dmcrypt out of the equation evens out IO load. It's misalignment of dmcrypt/ext4, not the Samsungs. # cat bench.raid0.nocrypt.chunk32.out hastur,3750M,,,429599,49,98253,12,,,310307,22,320.6,0,,,,,,,,,,,,, hastur,3750M,,,418548,49,100704,13,,,345801,24,309.8,0,,,,,,,,,,,,,

    # cat bench.raid0.nocrypt.chunk64.out
hastur,3750M,,,426720,51,103288,13,,,304494,19,337.2,0,,,,,,,,,,,,,
hastur,3750M,,,420937,50,105754,13,,,341867,19,333.3,0,,,,,,,,,,,,,


# cat bench.raid0.crypt_align256.chunk64.out
hastur,3750M,,,99446,11,40287,6,,,98194,7,308.0,0,,,,,,,,,,,,,
hastur,3750M,,,98978,12,40615,6,,,100049,7,307.4,0,,,,,,,,,,,,,

  • Mitigated by setting readahead and stripe cache, it seems

Prepare new array

Alignments

  • Partitions to HDD sectors: 512B or 4kB)
  • RAID chunks to HDD sectors: 4kB
  • dmcrypt sectors to RAID stripes: max_stripe_width * chunk_size
  • LVM sectors to ?
  • ext4 sectors to ?

Partition

Alignment

Layout

  • 300GB RAID0 (75GB per disk)

  • 3850GB RAID6 (1925GB per disk)

    for dev in {b..e} ; do parted /dev/sd$dev mklabel ; done

    Warning: The existing disk label on /dev/sdb will be destroyed and all data on this disk will be lost. Do you want to continue? Yes/No? y
    New disk label type? [gpt]?
    ...

    for dev in {b..e} ; do parted /dev/sd$dev -- mkpart primary 0 76800 mkpart primary 76800 -0 ; done

  • Buggy parted cli means creating labels manually.

  • -- argument lets you specify -0 on command line
  • Check with "unit s print" for dev in {b..e} ; do parted /dev/sd$dev unit s print ; done

Decided against partitioning.

  • RAID0 isn't expandable
  • LVM will replace the separation from partitioning

RAID

Alignment

  • With bitmap, default offset is 136 sectors (68Kb)
  • Check with mdadm -E
  • Data offsets: 4GB raid0 16 sectors (8kB) , 4GB raid6 24 sectors (12kB)
  • No alignment necessary since offset is multiple of physical sector size (512b or 4kB)

Chunk Size

Encryption

Alignment

Cipher

  • aes-cbc-essiv vs aes-xts-plain
  • aes-xts needs double keysize to feed equal parts to aes and xts
  • aes-xts has no ESSIV so no :hash is specified
  • Default luksFormat password hash is sha1. It's not vulnerable in the same way as signed certs are.
  • Specifying --hash for luksFormat is supported but seems largely unnecessary

  • 2048 is safe offset

    cryptsetup -c aes-xts-plain -s 512 --align-payload=2048 luksFormat /dev/md6

    cryptsetup luksOpen /dev/md6 crypt-md6

Logical Volumes

Alignment

Layout

  • Home, media - separate to allow home quotas to be configured

PV

pvcreate --metadatasize 193k /dev/mapper/crypt-md6

VG

vgcreate vg-md6 /dev/mapper/crypt-md6

LV

lvcreate -n media vg-md6 -L2500G
lvcreate -n home vg-md6 -L500G

Filesystems

  • EXT4

Resize reservation

  • -E resize=
  • tune2fs -l lists max fs blocks. (Default allows for 16TB)

Bytes per Inode

  • -i 65536 (reduces inode overhead, default is 16384)

Stride and Stripe

  • http://www.ep.ph.bham.ac.uk/general/support/raid/raidperf11.html
  • http://busybox.net/~aldot/mkfs_stride.html
  • Stride controls the space between metadata blocks. If stride == RAID chunk size there would be a bitmap in every chunk. Worst case would be a bitmap every stripe_width which would put all the bitmaps on a single disk.

  • Stripe-width lets the FS calculate the number of disks allowing parallel IOs stride(64k) = raid_chunk_size / ext4_block_size = 64k / 4k = 16 stripe_width(4) = raid_data_disks * stride = (4-2) * 16 = 32

  • Resizing an array

    tune2fs -E stripe-width=$NEW_STRIPE_WIDTH resize2fs

Create Home

  • No reserved space. Stride and stripe-width for 2 data disks and 64k chunk.

    mkfs.ext4 -m 0 -E stride=16,stripe-width=32 /dev/vg-md6/home

Create Media

  • No reserved space. 64k per inode. Stride and stripe-width for 2 data disks and 64k chunk.

    mkfs.ext4 -m 0 -i 65536 -E stride=16,stripe-width=32 /dev/vg-md6/media

fstab

/dev/vg-md6/home        /mnt/md6-home   ext4    defaults,noatime,nosuid,noauto,acl  0       3
/dev/vg-md6/media       /mnt/md6-media  ext4    defaults,noatime,nosuid,noauto,acl  0       3
/dev/vg-md6/home        /home   ext4    defaults,noatime,nosuid,noauto  0       3

crypttab

/dev/mapper/crypt-md6   /dev/md6        none            luks

Copy Data

Start new array in 3-of-4 disk degraded state

# mdadm --assemble --run /dev/md6 /dev/sd{h,i,j}
# cryptsetup luksOpen /dev/md6 crypt-md6
# vgchange -a y vg-md6
# mount /mnt/md6-media
# mount /mnt/md6-home

Start old array

# mdadm --assemble /dev/md2
# mdadm --assemble /dev/md3
# cryptsetup luksOpen /dev/md2 crypt-md2
# cryptsetup luksOpen /dev/md3 crypt-md3
# mount /dev/mapper/crypt-md2
# mount /dev/mapper/crypt-md3

Copy data

Reconfigure

rc.local

  • Clear out old script

Samba

# sed -i -e 's%md3/media%md6-media%g' /etc/samba/smb.conf
  • Replace valid users lists with groups: [mediasys] ... valid users = @group-name

SNMP

  • Add disk space graphs in cacti
    • Had to restart snmpd to update GetMountedPartitions query

Mediatomb

# vim /etc/mediatomb/config.xml
 <home>/mnt/md6-media/metadata/mediatomb</home>

mtdaapd

# sed -i -e 's%md3/media%md6-media%g' /etc/mtdaapd.conf

Switch Services

mediasys

# for mntpt in /export/mediasys/media/{movies,videos,music,tvshows,photos}/{library,meta} ; do umount $mntpt ; done
# for mntpt in $(grep '^[^#]*md6-media.*bind' /etc/fstab | cut -d' ' -f1) ; do mount $mntpt ; done

Sync RAID

Stop old array

Add 4th new disk

# mdadm --manage /dev/md6 --add /dev/sdb

Sync new array

Benchmark

  • Post sync. Read performance doesn't look right # bonnie++ -q -f -x 3 -s 3750 -n 0 -u root -d /mnt/md6-media name,file_size,putc,putc_cpu,put_block,put_block_cpu,rewrite,rewrite_cpu,getc,getc_cpu,get_block,get_block_cpu,seeks,seeks_cpu,num_files,seq_create,seq_create_cpu,seq_stat,seq_stat_cpu,seq_del,seq_del_cpu,ran_create,ran_create_cpu,ran_stat,ran_stat_cpu,ran_del,ran_del_cpu hastur,3750M,,,80818,11,24548,4,,,56356,5,305.6,0,,,,,,,,,,,,, hastur,3750M,,,86100,12,24631,4,,,59027,5,305.8,0,,,,,,,,,,,,, hastur,3750M,,,87435,11,24239,4,,,59217,5,312.1,0,,,,,,,,,,,,,

SideBar