RAID0 Implementation Under Linux

Most of us who use Linux at home don't have the same requirements as businesses that consider Linux a cost-effective, open alternative to expensive and proprietary Unices. Usually RAID devices aren't a requirement of the home user, although many users running a striped swap partition report a big improvement in speed. The multiple device (md) driver, written by Marc Zyngier, brings RAID to Linux.

md is a driver (included in the standard kernel distribution since 1.3.69) that allows you to group a number of disk partitions together so that they act as a single block device. md differs from the other drivers because it doesn't truly access the physical devices that compose it. md redirects requests from the upper layer to the devices involved and is interface independent, allowing IDE, SCSI and XT disks to be grouped as a single device.

There are three modes that md can use with its devices: linear, RAID0 and RAID1. In linear mode, the physical devices are appended to each other. When the first device reaches capacity, data is sent to the next device in the group. This mode allows for the creation of a device with a greater capacity but offers no real improvement in performance. RAID0 (or striped) devices spread the data evenly across all the devices in the group. Each write is broken into “chunks”, and the chunks are placed sequentially across the physical devices. RAID0 offers performance improvements, especially with concurrent disk access. RAID1 adds mirroring to RAID0. I feel that RAID0 is the most important of these modes; therefore, it is the focus for the remainder of this article.

When planning your RAID0 implementation, there are two considerations to keep in mind: physical device layout and device size. If you use partitions on the same physical device, you will not see any real benefit. The best recommendation I can make is to use several SCSI disks with each partition having the same number of blocks. This seems to offer the best performance. md can deal with different size devices as long as there is a significant difference. Using a 1,000,000 block device and a 1,000,001 block device can lead to problems. If you were to create an md device with a 500MB, a 1000MB and a 1500MB partition, it would run fine; md would split the device into “stripe zones” of 500MB. Once 1500MB was written to the device, the first physical device would be full. The second stripe would then be used on the second and third device. After another 1000MB is written, all data would be placed on the last device. Performance decreases in this arrangement as disk usage increases.

Once you have set up the partitions to be used, the kernel will have to be recompiled with md support enabled. Run make config (menuconfig or xconfig) and select “Multiple Device Support” and either “Linear” or “RAID0” mode. Compile as usual. While rebooting with the new kernel, you should receive a message like this:

md driver 0.35 MAX_MD_DEV=4, MAX_REAL=8
raid0 personality registered

If it went by too fast and you think you may have missed it, use the following command:

dmesg | more

You now have md support in your kernel, or as a loadable module if you went that way. Next you need to obtain the tools to manage your md devices. Although md is supported in the kernel, it appears that most distributions don't include the tools. They are available from ftp://sweet-smoke.ufr-info-p7.ibp.fr/public/Linux or from the mirror in the U.S. at ftp://linux.nrao.edu/pub/linux/packages/MD-driver. Red Hat software has an RPM distribution available at ftp://ftp.redhat.com/pub/contrib/RPMS. The file md-035-3.i386.rpm contains the needed binaries. Once you have downloaded and unpacked the source, become root and run make install. The compilation is straightforward, and I've never had a problem with it. If your Linux source code tree is not located in /usr/src/linux, you will need to edit the Makefile; otherwise, it should compile out of the box.