Hands On - Hardware - Raiding your hard drive

In 1988, David Patterson, Garth Gibson and Randy Katz published a paper entitled A case for Redundant Arrays of Inexpensive hard Disks, or Raid for short. It explained how several cheap hard disks could work together to deliver greater capacity, performance and/or fault tolerance.

The authors came up with five configurations and named them Raid Levels 1 to 5. In 1993, four more configurations were devised, with the Levels of 0, 6, 10 and 53.

In practice, Raid takes multiple hard disks and makes them act as one better performing drive. Consequently, even though there may be two or more physical disks in the PC, the operating system (OS) sees just one single drive, with one drive letter.

The most common implementations of Raid are Levels 0 and 1, known respectively as striping and mirroring. Striping takes multiple drives and distributes the data evenly between them. Consequently, when reading or writing, the system is in fact accessing not one, but two or more drives simultaneously, which theoretically improves data throughput.

There's no loss in storage either, with the total capacity equalling the sum of the separate disks - so a pair of striped 40Gb drives will result in a single 80Gb volume, which could perform twice as fast as a single disk.

The downside to striping is that Raid and the OS treat the multiple drives as a single volume, so if anything goes wrong with just one of the physical disks, you effectively disrupt the entire volume. Statistically speaking, a pair of striped hard disks relying on each other is therefore more likely to fail than a single disk.

The answer to this is Raid Level 1, or mirroring, where a second hard disk literally makes a copy of everything on the first. If one disk fails, then the other one takes over seamlessly.

You can then install a replacement disk, and have the Raid system use this as the new backup volume. The upside is higher reliability than a single disk, but with little or no benefit in performance.

The real downside, though, is that a pair of mirrored 40Gb disks will only be treated by the OS as a single 40Gb volume, so you'll have paid twice as much just for the additional reliability.

The beauty of Raid is that you can combine several levels at a time. If you have four hard disks, you can stripe two of them, then use the second pair to mirror them.

If you started with four 40Gb disks, you'd now have a single 80Gb volume that performs very well, and is reliable too. This is known as Raid Level 0+1, although performance nuts may find the temptation not to stripe all four disks too much to bear.

Raid requirements

Of course, you're going to need something that treats these disks as a single volume, and that's the Raid controller. While it's possible to create software Raid volumes using some server OSs, it's not advisable in terms of the stress on your processor and OS itself.

It's much better to have a dedicated hardware Raid controller on an expansion card, or integrated onto the motherboard.

If you have two, four or even more hard disks all firing at the same time, the bandwidth of the controller bus becomes very important. This is why, in the past, Raid has been limited to the world of SCSI, where bus bandwidths are traditionally higher.

Today, however, the common UltraDMA100 EIDE standard has no less than 100Mbps of bandwidth at its disposal and, better still, access to hard disks that are much cheaper than their SCSI counterparts.

As a result, companies like HighPoint, Promise and Adaptec have produced hardware Raid controllers for EIDE hard disks, and the latest models operate with the full UltraDMA100 bandwidth.

Raid chipsets from HighPoint and Promise in particular are being integrated into more sophisticated motherboards as an additional selling point. For example, Abit's VP6 (dual FC-PGA) and KT7A-Raid (single Socket A) motherboards both use HighPoint's HPT370 Raid controller, which offers Raid 0, 1 or 0+1.

Enter the VP6

While Raid can theoretically support any number of hard disks, the electronic limitations of an EIDE controller restrict support to only four. Since Raid performs at the lowest common denominator, it's also advisable to use identical disks where possible. I took four Seagate Barracuda ATAIII 40Gb UltraDMA100 drives, at £120 ex VAT each, and set to work.

My test system used the Abit VP6 motherboard, a pair of 933Mhz Pentium III processors, 256Mb of CAS2 PC133 SDRam and an ATi Radeon 64Mb graphics card. With dual chips, a hungry graphics card and four high performance hard disks, I used the muscle of an Enermax 431w ATX power supply, bought from www.overclockers.co.uk, for just under £100 including VAT and delivery.

Setting up Raid

Throwing caution to the wind, I ignored the reliability of Raid Level 1 and opted instead to see what benefits there were to striping two or four hard disks, compared to using just a single drive.

Each configuration used Windows 2000 Professional and the NTFS file system. The latest updates were installed, with the HighPoint controller BIOS at 1.0.3b1, and the VIA 4-in-1 driver at 4.29V.

Setting up Raid on the HighPoint controller was extremely simple. Just connect your drives to the appropriate EIDE channels on the motherboard, enter the main CMOS and ensure that the Raid controller is enabled and that ATA100Raid is set as one of your boot devices, then press Ctrl & H to enter the specific HighPoint BIOS when instructed.

Within the HighPoint BIOS, simply choose the Create Raid option, then follow these four steps: first, choose between Raid Level 0, 1 or 0+1 (I chose 0 for striping). Second, select the connected disks you'd like to use in this array. Third, choose the block size (see later), and fourth, start the creation process.

Striping is instant, although it will delete any information on your disks, and create a single space that subsequently needs to be partitioned and formatted. Mirroring requires existing data to be duplicated, so it could take some time.

Results

First things first: as I explained earlier, striping two or more disks produces a single large drive. With four 40Gb hard disks striped together, Windows saw them as a single 149Gb formatted drive, which is undeniably huge! Obviously, anyone who requires single huge disks for editing massive video files will see the benefit of Raid striping.

But what about those block sizes? The HighPoint controller offers block sizes of 4Kb, 8Kb, 16Kb, 32Kb and 64Kb, which refer to the smallest amount of data that can be addressed on the array.

As far as I understand it, if you choose a block size of 64Kb, any file smaller than this will still occupy 64Kb of your disk - larger block sizes are therefore less efficient, especially if your typical files are small, but they do offer greater performance. It's up to you although, if in doubt, why not got for 16Kb in the middle?

Under SYSmark 2000, a single Seagate disk set up as a normal drive scored 203. Two disks, connected to separate channels and striped with a 4Kb block size, scored 190, although testing again with a 64Kb block size increased the score significantly to 208. Four drives, striped with a 4Kb block size scored 201, which rose to an impressive 215 with a 64Kb block size.

From these results, I'd say that striping with EIDE Raid certainly offers quite a performance increase, especially with larger block sizes, or with four drives if you can afford it.

Striping also gives you a large single volume, which is great for certain applications, although don't forget about the lower reliability which statistically reduces with a greater number of drives striped. Obviously, disk-intensive applications like database or audio/visual editing work will benefit most from Raid striping.

With a striped system I'd certainly back up my data regularly, although the performance geek in me would probably still use all my available drives exclusively in a striped configuration.

It's your call, although if you're dealing with a server, mirroring is obviously the most sensible use of multiple drives. Remember that you can always go for four disks in a Raid 0+1 (striping and mirroring) configuration, giving you the best of both worlds!

At the end of the day, Raid striping offers an excellent performance option for anyone with a suitable controller, especially when you consider that additional EIDE disks cost around £100 each.

Promise also sells its EIDE Raid controller on a PCI card for £86 ex VAT. If you've had any experiences with any kind of Raid, I'd like to hear them!

Fast memory

The speed of your PC's memory has a massive impact on overall system performance, but it's not just tied to the clock of your system bus. As any regular visitor to their CMOS knows, there are tons of memory options that can greatly improve performance and, just like comedy, it's all in the timing.

Without getting too bogged down in detail here, the main thing to worry about is the latency setting, which defines the time delay between SDRam receiving data and actually reading it.

Essentially, there are two common latency settings for SDRam, known as CAS2/CL2 and CAS3/CL3, and, since these refer to actual timings, it won't surprise you to learn that the shorter CAS2/CL2 is preferable.

Sadly, you can't just go into your CMOS and change the latency setting from CAS3 to CAS2 and reap the benefits. The actual memory itself must be capable of being accessed that bit more quickly and yours may not be up to the job - it doesn't matter whether it's PC66, PC100 or even PC133 SDRam, as all are available in either CAS2 or CAS3 versions.

If you've got CAS3 memory, I'm afraid you'll be asking for trouble if you run it at CAS2 latency - even if your system starts up, it will become very unreliable, forcing you to pop back into the CMOS and change it to CAS3 again.

If your memory is rated for CAS2, it will almost certainly have a sticker on it saying as much. The bad news is that unless you specifically asked for CAS2 memory, you'll almost certainly have been sold the slower and cheaper CAS3 flavour.

In the past, CAS2 SDRam always carried quite a price premium over conventional CAS3 SDRam. Today, however, the situation has changed, and at the time of writing, you're looking at spending less than a fiver extra for the CAS2 version of a 256MB SDRam DIMM.

But the big question is how much faster is CAS2 over CAS3 anyway? I took the plunge and ordered a stick of 256Mb CAS2 PC133 SDRam to find out - it cost me £65.69 ex VAT from www.dabs.com in March.

My test system consisted of a 933Mhz Pentium III running on an Abit VP6 (VIA 694X) motherboard, with a GeForce2 GTS graphics card and an 18Gb Seagate Cheetah XL Ultra160 SCSI hard disk formatted with FAT32. I installed Windows 98 SE with the latest drivers, including the VIA 4-in-1 4.29V.

I ran SYSmark 2000 and Quake III Arena (XGA 16bit) first using 256Mb of CAS3 PC133 SDRam, then again with 256Mb of CAS2 PC133 SDRam; both featured their full 256Mb on a single DIMM.

When using CAS3, the advanced CMOS settings were left on their SPD defaults, while the CAS2 configuration was used with a Turbo setting and, of course, a latency of two.

Under SYSmark 2000, the CAS3 system scored 178, while the CAS2 scored 182. Under Quake III Arena, the CAS3 system scored 94.1fps, compared to 102.3fps with CAS2.

The SYSmark score is certainly only fractionally better, but then I've become used to seeing small performance increases under standard Office application benchmarks when using new types of memory.

However, the Quake score is much more impressive, producing almost a 10 per cent increase just by changing the memory. I would also bet that intensive graphics applications will see the benefit.

So should you buy CAS2? Well, if you were considering swapping your existing CAS3 SDRam for it, then I'd say the benefits will not outweigh the cost of investment.

If you were buying or building a new system, though, I'd definitely go for CAS2, as it's measurably faster and only fractionally more expensive - just ensure that your CMOS is driving it with the latency of two that it deserves.

Contacts:

Gordon Laing welcomes your comments on the Hardware column. Contact him via the Personal Computer World editorial office or email hardware@pcw.co.uk. Please do not send unsolicited attachments.