raid_ephemeral.sh

#!/bin/bash -eu
#
# this script will attempt to detect any ephemeral drives on an EC2 node and create a RAID-0 stripe
# mounted at /mnt. It should be run early on the first boot of the system.
#
# Beware, This script is NOT fully idempotent.

METADATA_URL_BASE="http://169.254.169.254/latest"

DEBIAN_FRONTEND=noninteractive apt-get -y install mdadm curl

# Configure Raid - take into account xvdb or sdb
root_drive=`df -h | grep -v grep | awk 'NR==2{print $1}'`

if [ "$root_drive" == "/dev/xvda1" ]; then
  echo "Detected 'xvd' drive naming scheme (root: $root_drive)"
  DRIVE_SCHEME='xvd'
else
  echo "Detected 'sd' drive naming scheme (root: $root_drive)"
  DRIVE_SCHEME='sd'
fi

# figure out how many ephemerals we have by querying the metadata API, and then:
#  - convert the drive name returned from the API to the hosts DRIVE_SCHEME, if necessary
#  - verify a matching device is available in /dev/
drives=""
ephemeral_count=0
ephemerals=$(curl --silent $METADATA_URL_BASE/meta-data/block-device-mapping/ | grep ephemeral)
for e in $ephemerals; do
  echo "Probing $e .."
  device_name=$(curl --silent $METADATA_URL_BASE/meta-data/block-device-mapping/$e)
  # might have to convert 'sdb' -> 'xvdb'
  device_name=$(echo $device_name | sed "s/sd/$DRIVE_SCHEME/")
  device_path="/dev/$device_name"

  # test that the device actually exists since you can request more ephemeral drives than are available
  # for an instance type and the meta-data API will happily tell you it exists when it really does not.
  if [ -b $device_path ]; then
    echo "Detected ephemeral disk: $device_path"
    drives="$drives $device_path"
    ephemeral_count=$((ephemeral_count + 1 ))
  else
    echo "Ephemeral disk $e, $device_path is not present. skipping"
  fi
done

if [ "$ephemeral_count" = 0 ]; then
  echo "No ephemeral disk detected. exiting"
  exit 0
fi

# ephemeral0 is typically mounted for us already. umount it here
umount /mnt

# overwrite first few blocks in case there is a filesystem, otherwise mdadm will prompt for input
for drive in $drives; do
  dd if=/dev/zero of=$drive bs=4096 count=1024
done

# Create /dev/md0
partprobe
mdadm --create --verbose /dev/md0 --level=0 -c256 --raid-devices=$ephemeral_count $drives
echo DEVICE $drives | tee /etc/mdadm.conf
mdadm --detail --scan | tee -a /etc/mdadm.conf
blockdev --setra 65536 /dev/md0

# Tuning for ec2 instance-store vols: http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/InstanceStorage.html#disk-performance
#
# "When configuring software-based RAID, make sure to change the minimum reconstruction speed"
echo $((30*1024)) > /proc/sys/dev/raid/speed_limit_min
#
# "Because of the way that Amazon EC2 virtualizes disks, the first write to any location on a standard instance store
# volume performs more slowly than subsequent writes. For most applications, amortizing this cost over the lifetime of
# the instance is acceptable. However, if you require high disk performance, we recommend that you pre-warm your drives
# by writing once to every drive location before production use."
# 
# "Note: The I2 high I/O instance type uses direct-attached solid state drives that provide maximum performance at launch
# time, without pre-warming."
#
# Pre-warm instance-store vols
# - This is VERY SLOW
# - Repeat until per-block write performance stabilizes across runs (?)
# - Try multiple instances to detect and throw out bad apples...
time pv -terb /dev/zero -s "`blockdev --getsize64 /dev/md0`" | dd of=/dev/md0 bs=1M || true

# Create + mount fs
mkfs -t ext3 /dev/md0
mount -t ext3 -o noatime /dev/md0 /mnt

# Remove xvdb/sdb from fstab
chmod 777 /etc/fstab
sed -i "/${DRIVE_SCHEME}b/d" /etc/fstab

# Make raid appear on reboot
echo "/dev/md0 /mnt ext3 noatime 0 0" | tee -a /etc/fstab