Using used for fun and savings

By Paul Murphy, author of The Unix Guide to Defenestration

Sometimes the cost of high end technology comes down a lot faster than its utility, thereby creating opportunities to combine savings with significant technical advantages.

One of the areas where that's happening right now (late 2003) is with respect to RAID array storage. Sun's T3 array, for example, used to be extremely expensive; now refurbished units are hitting the market at nickel and dime rates. You can, for example, expect to get a pair of fully configured T3s, each with 256MB of cache, a PCI controller and cable, and nine 18.2GB FC-AL disks for around $3,100 each.

Later on, you can expand storage to 1000 times the 327MB my Sun 386i came with in 1989 by upgrading to the 36.4GB disk and eventually double that again by switching to the 73GB units - all without changing any software or the physical setup.

Note that people may tell you that need Stortools to manage a T3, but this isn't true. Get the manual from Sun and you can use telnet (ugh) if you have to.

The best way to set them up depends on your application but you get the biggest benefits on RDBMS applications. In that environment it makes sense to use software mirroring to take advantage of having two controllers and then set each array up to give you two blocks of four disks in RAID0 (striping only) mode each. That gives you 144GB of fully mirrored and spared disk space for all of about $6200.

Note that treating the array as a two blocks of four logical devices isn't necessary since you can treat the whole array as one device but speeds recovery on failure or upgrade.

$6,200 may not sound like much of a bargain given that you can now buy a 180GB ATA drive for about $800 and put two of them in a mirrored array for under $2,400 but, of course, there are critical differences in access speed and reliability.

All of the important parts on a T3 can be replaced on the fly -just don't take things out until you have the replacement part ready; there's only a 30 minute run-time window with parts removed. Stacking two of these beside your machine and mirroring them while holding out a spare disk in each set pretty much gives five nines for reliability.

Reliability is, of course, great but it's performance that makes the killer difference. Sun to the contrary, the T3 maxes out at 93MB/Sec in transfer rate, or about one third that of a US3 array. The difference, however. comes from the use of a smart 256MB on board cache. Get a US3 disk and you get about 20MB of on disk cache, but it's purely for flow through buffering and not that effective for mixed I/O. For the T3 you set your database for raw I/O (You can use direct I/O too, but I don't like to) and you can expect the traditional I/O bottleneck to disappear for almost everything.

There are exceptions; US3 will be faster for large sequential reads of the kind you get in a poorly designed data warehouse. For most things, however, it's disk delay, not controller delay, that governs performance and that 93MB/Sec limit rarely affects you. The bottom line is simple: this is advanced technology whose come down in price means that ordinary systems can now get near perfect reliability and much higher throughput for a few thousand bucks and some set-up work.