Reliable, Inexpensive RAID Backup
As a topic, backups is one of those subject likely to elicit as many answers as people you ask about it. It is as personal a choice as your desktop configuration or your operating system. So in this article I am not even going to attempt to cover all the options. Instead I describe the methods I use for building a reliable, useful backup system. This solution is not the right answer for everyone, but it works well for my situation.
Everyone knows they should be doing backups. But do you? How many times have you started a backup schedule only to let it slide after a few weeks? Sounds a bit like an exercise or diet regime, doesn't it?
I had several goals when designing a new backup system for my home and colocated web server: reliability of stored data, automation of the backup process and relative low cost. Human error is the weakest element of any backup system, so a 100% hands-off system was my goal.
In "Scary Backup Stories", Paul Barry discusses failed backups. The common thread of his stories was somewhere in the chain of events a person had forgotten a very important step. The first story he tells highlights how one team forgot to format the tapes. They had religiously followed their backup plan, backing up onto the unformatted tapes, only to discover the tapes were useless.
I did some reading and settled on a RAID-5 array of hard drives as the most reliable way to store data. It can survive a single drive failure and recover from it when you replace the failed drive. Unlike tape, CDR or DVD backups, it doesn't need someone to swap media or format and rotate tapes. None of the RAID methods can survive a two-drive failure, so RAID-5 is as good as it gets.
RAID-5 achieves its reliability by writing the data across a number of disks, along with error detection information. The information is spread in such a way that no single-disk failure can destroy the archive. And when you replace the failed drive it automatically rebuilds the data that was on that section of the RAID.
The base system would be my recently retired colocated web server box. It has a nice rackmount case, a 400MHz AMD processor and 768MB of RAM. I added a beefier power supply (Antec 350W from Best Buy) to replace the 250W unit that came with the case. The system already had a SCSI controller and a 5GB SCSI drive that I'd be using for the root filesystem. Yes, 5GB is small by today's standards, but this system was built and installed in 1999. It ran without failure until it was removed in December 2002, because the ISP went out of business. The minimal install of Red Hat 8 takes about 400MB, so this drive works just fine for its new purpose.
SCSI usually is the first choice for reliable RAID hardware, but it is expensive--not only the drives but the controllers, too. Also important reason is speed: SCSI handles multiple accesses to the drive more efficiently than IDE drives. But for my application speed wasn't a deciding factor.
IDE RAID controllers are becoming more affordable but are still in the $200+ price range as of this writing. A less expensive alternative is to add several IDE controller cards to the system and put one drive per channel (2 drives per card) on them. These PCI IDE cards are less than $25 each, and they support the newer 133MHz IDE bus speeds.
I chose to install two PCI cards for use as RAID controllers. This left the IDE controllers on the motherboard free for adding other drives at a later time. They also could be used to quickly back up a drive that I didn't want to copy over the network.
There are two good reasons for limiting backup to a single drive per channel. First, if one drive fails it can disrupt the other drive on the channel, causing a catastrophic two-drive failure. The other reason is speed. With two drives on an IDE chain, the throughput is halved, as I understand it, so it makes sense to use only a single drive. An argument also can be made for using only one drive per controller card. At that point, though, you might as well invest in a dedicated RAID card.
My drive choice had already been made. For some time, I'd been using a second Maxtor drive in each of my systems as a backup drive, mirroring the live filesystem to it with rsync. And I have been using Maxtor drives for years without a single failure, unlike Fujitsu drives, which seem to drop dead within a year (I have three of them in the junk box). I suppose this means that as soon as this article is published, all of my reliable drives will fail at the same time.
You need to have three drives for a minimum RAID-5 system. The drives all should be the same size, because the total size is calculated using the smallest drive size, multiplied by 1-number of drives. So, three 30GB drives yield a RAID-5 of about 60GB of storage. At the time, I had two 40GB and one 30GB drives on hand. So I wasted about 20GB of space in building this system in the interest of getting it up and running as quickly as possible.
It may be possible to resize the array by adding more drives at a later time, but unless you have a second backup of the data, you probably don't want to try this. Instead I'd recommend buying a larger drive, copying the RAID to it and rebuilding the RAID filesystem from scratch.
I'm not going to get into a discussion of the different RAID levels. Suffice it to say that for my purposes RAID-5 fit the bill. It provides larger storage space than the single drive and the capability of surviving and recovering from a single-drive failure.
When dealing with requirements like mine, I really don't see any need to have hardware RAID. I don't need speed, the backups run when the LAN is usually idle and the only other load the machine has is running the SETI@home client in the background.
The goal here was to install as plain a Linux system as possible, so in the event of a failed RAID root filesystem, it could be reinstalled with a minimum of hassle. I have several systems running on Red Hat 8.0, so I chose it as my distribution. The instructions, though, should apply to any modern Linux distribution that has RAID support enabled in the kernel by default.
I did a minimal install of Red Hat 8.0, selecting individual packages and turning off everything that didn't look important. RH may call it a minimal install, but it still includes a number of things you probably don't need. Check the box that says select all packages, then go through the list and turn them off. If you turn off too much, the configuration program will resolve the dependencies before the final install and prompt you with a list of packages that need to be added.
Use Disk Druid to partition your drives. For the drives that will be used in the RAID, format them as Software Raid and select a partition size that covers the full drive. Remember to configure another drive/partition as the root partition with swap and /boot. RAID systems can be booted from a root partition that lives on the RAID, but it is a bit tricky to set up, and I wanted to keep this as straightforward as possible.
To create the RAID system select the RAID button from the choices in Disk Druid. The partitions you selected as Software Raid will be selected by default. Enter a mountpoint (I used /backup) and the RAID level (5 in my case, really the only option that makes sense to me). Format it with your favorite journaling filesystem. I used ext3 for my system, but ReiserFS should work equally as well. I tend to prefer ext3 to ReiserFS mostly because it is backward-compatible with ext2. This way, if anything happens to the journal I can still access the data as an ext2 filesystem.
Continue with a normal install. You can put as much or as little on the system as you wants. I selected the minimal install and had to install the samba-common, samba-clients and cups-libs packages before smbmount could be used to backup Windows machines.
Reboot your system and confirm RAID is running by entering df to see what filesystems are mounted and what their capacities are. Here's my current output:
/dev/md0 is the RAID device, and as you can see I've done a good job of filling it with backups. Which brings me to the next step--actually backing up your systems. I use rsync and SSH along with smbmount for my backups. Set up your systems so the root user on the backup system can access root on all the systems that need to be backed up. Set it up so the backup system's root user can log in without being asked for a password.
Do this set up by generating a key pair on the backup machine with ssh-keygen -t DSA, and then copy the .ssh/id_dsa.pub file into the .ssh/authorized_keys2 file on all of the systems to be backed up. This authorizes the backup system to access all of the target system's files. If you only need to back up a subset of the files,you could use a user other than root on the target system.
Because this system has access to all of your other systems, it needs to be as secure as possible. Don't run any other services on it, and make sure you always use SSH to log into the machine, so its root password isn't exposed to the rest of the network.
I use rsync to handle the copying of only the files that have changed since the last backup. This program efficiently calculates the differences and transfers the changes, saving time and bandwidth. With rsync I am able to do nightly backups of my colocated web server--after an initial eight-hour backup of the base system over my 256KB cable modem connection.
I modified an rsync backup script by email@example.com to fit my needs. It creates a lockfile to prevent two instances from running at the same time, which is a possibility if something hangs during a backup. It dumps a copy of all the RPMs installed on the target system into a file in the target's /etc/ directory, using this command:
ssh firstname.lastname@example.org "rpm -qa > /etc/rpm_qa.txt"
This way you know what RPMs were installed on the system.
The script uses the backup-dir feature of rsync to create daily directories that contain the files that have changed. This way you end up with a current, full and complete backup and seven directories, named after the days of the week, with the files that changed on that day. This is much easier to restore than a old-fashioned, full backup and incremental changes.
The script could be modified to fit a different backup schedule by changing the way the directory used by the backup-dir argument is named. See the associated listing, linux_inc, for the script to handle backing up Linux machines.
For Windows systems (I have only one, my wife's computer) I mount the Windows shares to the backup system using smbmount, and then use rsync on the local filesystem to make the backup. See the associated listing windows_inc for the backup script to handle Windows machines.
All of this is automated with a crontab:
MAILTOemail@example.com # Backup the windows machine at 7pm 0 19 * * * /backup/scripts/windows_inc # Backup Linux machine at 2am 0 2 * * * /backup/scripts/linux_inc
In the scripts provided, do a search for "target" and replace it with your machine's name or IP address to customize the script for your setup. Make a separate copy for each machine to backup, and add it to root's crontab using crontab -e.
The last feature of the system is automated shutdown when the power fails. The system uses an Asus P5A motherboard with an ATX power supply, so it is capable of shutting itself off. I have it connected to an APC 500 power backup with a USB connection.
I installed the latest version of apcupsd to handle shutting down the system when the power has been out for two minutes. The ext3 filesystem and the RAID should be able to prevent any data corruption without a UPS attached, but why take the chance?
My system has been running backups for about a month. Nightly reports are e-mailed to me (from root's cron job) that detail the files backed up. The only hitch I ran into was when the Windows machine was off it would delete the archive--not a good thing! So I added error checks to the smbmount, and not it does not try to do a backup if mounting the Windows shares fail.
Hopefully this article has convinced you that automated backups can be done with a minimum of hassle. It is possible to remove much of the human element from the backup process, but not completely. You still need to monitor your system to make sure things are running smoothly.
Brian Lane is a software developer from Seabeck, Washington, where he lives with his wife and son. When he isn't writing software for www.shinemicro.com, he is working on various Linux projects which can be found at www.brianlane.com.