Date: Sun, 15 Mar 1998 13:44:00 -0500 From: Cory Kempf <ckempf@enigami.com> To: Raul Zighelboim <rzig@gulfsouth.verio.net>, "'scsi@freebsd.org'" <scsi@FreeBSD.ORG> Subject: Re: to raid or not to raid Message-ID: <v04003a24b131be90dab6@[208.140.182.45]> In-Reply-To: <A03CD00C69B1D01195AB00A024ECEB167110D9@kaori.communique.net>
index | next in thread | previous in thread | raw e-mail
Oh Cool, one I can answer.... At 10:53 -0500 98.03.15, Raul Zighelboim wrote: >Could anyone point me to documentation on the pros and cons of raid >5/raid 0 in a non-pro language ? >Also, any pro/cons off 'raid 0 over raid 5' ? OK, there are three major flavours of RAID, 0, 1, and 5. (assuming I haven't scrambled the numbers again...) These levels correspond to Striping, Mirroring, and sort of a combo of the two. Start with a basic disk. You do a read (writes work similarly). The system asks the disk for blocks 100-400. The disk mechanism moves the heads, and when the data is under the head reads into its internal buffer, then sends it out. This maybe happens at 8MB / second, on a fairly modern drive. In the old days, the bus was tied to that disk while it was looking for the data. Then came disconnect: the computer instead says get the data, bug me when you have it. It still takes the same amount of time, but the data can be "burst" to the host much faster, say 20 MB / sec on Ultra SCSI. And this is where RAID comes in. With Striping, we take a pair of 4 GB disks (for example), and pretend they are one 8GB disk. We ask drive A for the first 64 blocls, and while it is looking for that, we disconnect and and ask drive B for the second 64 blocks. Drive A comes back and says it is done, here is the data, and we ask it for the third 64 blocks. Drive B comes back and... The point is, the data moves at about double the speed as with a single disk. Due to overhead, it isn't twice as fast, but it is faster. The down side is that if one disk dies, you lose all the data on both drives. Thus your MTBF is cut in half. Mirroring, IMO, is a waste. The idea is that if one drive has a HW failure, well, the data is on two drives, no big deal. IME, hardware failures are much less common than soft failures, and with a soft failure, both drives are hosed with the same bad data. Typical MTBF (mean time between failures) nowdays is around 800,000 hours... or about 91 years. At least according to the manufacturor. (uh huh, right. And how did they come up with that number?) Raid 5 takes 3 disks (or more, and sort of stripes them and sort of mirrors them. With three 4 GB disks, you have 8 GB storage. Lets say you were writing the following (binary) data: 1100 1010 Drive A would get 1100, drive B would get 1010, and Drive C would get 0110, or the first XOR the second. Now, lets say drive B has a HW failure. If we XOR the data on Drive A with that on drive C, we get: 1010. In other words, we can recover from any one drive failure. The downside is that you need to do the XORs for each write (writes are ~20% of I/O ops) and have to write it out to disk, and you pay for three drives but only get two. Generally, you need hardware support to make it work though. In short, while RAID 5 is more expensive, it gives you most of the benefits of RAID 0 and 1, without the expense of RAID 1 or the risk of RAID 0. Multiple busses are used if there is not enough bandwidth on a single bus. FWIW, UW SCSI has bandwidth of around 40 MB /s. Fiber SCSI is much faster. The fastest disks out there (Seagate Cheetahs this week) only put out around 16 MB / s IME. +C -- Thinking of purchasing RAM from the Chip Merchant? Please read this first: <http://www.enigami.com/~ckempf/chipmerchant.html>; Cory Kempf Macintosh / Unix Consulting & Software Development ckempf@enigami.com <http://www.enigami.com/~ckempf/>; To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-scsi" in the body of the messagehelp
Want to link to this message? Use this
URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?v04003a24b131be90dab6>