Tuning Linux MDRAID 5 for acceptable performance

tune_cassandra_raid.sh

#!/bin/bash
# Linux RAID5 tuning
# Al Tobey 2011-09-19 <al@ooyala.com>

for sysent in /sys/block/sd*
do
    drive=$(basename $sysent)
    sector_size=$(blockdev --getss /dev/$drive)

    # it should already be CFQ, make sure, I use ionice extensively
    echo cfq > /sys/block/$drive/queue/scheduler

    # allow 256 in-flight BIO's per drive for better IO merging
    echo 256 > /sys/block/$drive/queue/nr_requests

    # set each drive's readahead to 256k
    disk_ra=$((2**18 / $sector_size))
    blockdev --setra $disk_ra /dev/$drive
done

# significantly increase the number of entries in the stripe cache (default 128)
echo 16384 > /sys/block/md7/md/stripe_cache_size

# set RAID5 /data device to 8MB RA
sector_size=$(blockdev --getss /dev/md7)
blockdev --setra $((2**23)) /dev/md7

# these next few are guesses and not verified, but not really dangerous either

# set the RAID1 /commit device RA to 0, since it's practically write-only
blockdev --setra 0 /dev/md6

# disable RA on boot, root, and var, underlying device RA should kick in (need to verify)
blockdev --setra 0 /dev/md1 # /boot
blockdev --setra 0 /dev/md3 # /
blockdev --setra 0 /dev/md5 # /var