Will I or Will I not RAID? IWill or IWon’t?

Many thanks to Iwill for letting us review their RAID card and to Todd for his terrific assistance.

Bottom line summary: If you use Windows 98 and manipulate big graphics or audio or video files, this controller is for you.
If not, it probably isn’t. Don’t expect better OS or app/game loading. NT/Windows 2000 users can do much the same thing in software with their OSs.

Is IDE RAID for you?

There’s been quite a bit of attention on RAID cards as of late. Should you get one? Will it help you, and where?

Reviews of these cards have tended to post a lot of numbers. Too often, we attach more importance to numbers than they deserve. We think a number is somehow more real, objective and authoritative. Maybe we like numbers so much because we think we don’t have to think when we have them.

Sometimes, though, you can’t impose simplicity on a situation. A single number is just not capable of truly reflecting reality.

The purpose of a benchmark is to measure performance, not become it. If you don’t get better performance, it doesn’t matter what the benchmarks say.

I’ve had the IWill RAID controller card working for a couple days. I’ve been generating a lot of benchmarking numbers too, but the more numbers I have, the less confidence I have in them.

All you need to know about IDE RAID

RAID is a mechanism by which you get multiple drives to do things together they can’t do by themselves. There are basically two things you can do with IDE RAID cards:

1) Make two drives do what one normally does so the job gets done faster and/or

2) Have another drive backup work you do on the first while you’re doing it, so if something bad happens to the first, you can use thesecond.

Since most of you are more interested in speed rather than backup, I just looked at that. Others have reported the card backups reliably, so I wouldn’t be too concerned about it functioning well as a backup.

How RAID makes it go faster

RAID can make hard drives serve the CPU information faster the same way a head waiter can serve a table faster with two waiters rather than one. Rather than have one waiter bring out three entrees at a time to a table of six, the head waiter can instruct two waiters to bring out three entrees each so everyone can be served in one trip rather than two.

However, computer programs don’t usually ask for information like that. They tend to ask for relatively small items one at a time.

It would be like one customer at a table ordering six peas, and another three carrots, followed by another two similiar orders, then another two, and another, and another . . . . Using two waiters that way may be a little more efficient than one, but nowhere near as much when you have big orders. Since it takes a little time for the head waiter to give instructions, it can sometimes happen that you would have been better off with one waiter.

Like most people, hard drives tend to get more work done if you leave them alone to do one task rather than constantly interrupting them and moving them around.

If the reading head of a hard drive has to keep jumping around to different places to load data, it has to spend time getting from one place to another, and while it’s going someplace else, it can’t do its job.

So if you have a big long contiguous file like a video; it should normally sit in one place longer and thus actually work more than if it has to jump around a lot grabbing relatively small chunks like HTML pages.

A RAID system will do much better when it gets to deal with big orders. What’s a big order in the computer world? A big file, like a very large graphic or sound or video file, something that takes a lot of time to load or save in the real world.

I said a big file, not a big application. A big application usually consists of a lot of little files being grabbed in sequence, like a lot of orders of sequential orders of six peas and three carrots.

So, does it speed up things? Well, that depends on what program you ask and when you ask it 🙂

What I found:

I believe benchmarks should be application-related, so that’s what I used. Synthetic benchmarks are good for telling you why certain things are happening, but if you have the greatest theoretical performance in the world and if it never shows up in the real world, what good is it?

Different applications access data at much different speeds, as you’ll see below. Some of this is due to the nature of the activity and the size of the files being used. Some is just due to the programming done to get data.

Whatever the reason, the one thing you can’t assume is general consistency, that if one specific application does better, all apps will do better.

Winbench:

I ran tests multiple times under both Windows 98SE and Windows 2000, using both Business Disk Winmark 99 and High-End Disk Winmark 99.

(All tests run using Windows 98 and 2000 with a Coppermine 550E thanks to Proton Computers, (very nicely put together 550E package with slotket and Alpha cooler) running at 733Mhz, using two 13.5Gb IBM 22GXP drives, 256MB Micron PC133 RAM, Matrox G400 card. RAID setup consists solely of two drives working in tandem.)

We have this nice table of numbers. Almost all of them are
significantly higher than what you’d get from a single drive. It would seem that the IWill card is doing something right.

But is it? How solid are these numbers? How much faith can you put in them?

Unfortunately, the guy who made them has to say, “Not necessarily too much.”

There’s no question that the RAID enabled disks generally did better than a single drive in these tests. How much better?

That’s the problem. There’s an awful lot of variation in some of these scores, especially when the IWill card in place.

Some of the synthetic benchmark data posted elsewhere seems to indicate that disk transfer rates jump around quite a bit while the card is in use. I occasionally noticed the same thing when installing OSs with the card in place; it would tend to race and slow down, race and slow down.

A few times, you had scores run under identical conditions differing more than 20% from low to high using the IWill card. Nor is it comforting to see little relationship between the numbers I came up with and numbers other reviewers generated using more or less the same equipment.

While I feel comfortable saying the IWill RAID will usually do better, I can’t say sometimes how much improvement you can consistently expect. Some scores are pretty consistent, others are not.

Some of that appears to be the card, some of it the benchmark itself. Even worse, the improvements you might expect from better disk access don’t necessarily show up in benchmarks that test the whole system.

Here’s the range of improvement:

Windows 98:

Business Disk: -10%–13%

AVS Express 3.4: 13%-20%

Microstation SE: 12%-16%

Photoshop 4.0: 33%-34%

Premiere 4.2: 19%-24%

Sound Forge: 19%-31%

Visual C++: -1%-23%

Windows 2000:

Business Disk: 29%-34%

AVS Express 3.4: 11%-32%

Microstation SE: 4%-20%

Photoshop 4.0: 37%-42%

Premiere 4.2: 35%-65%

Sound Forge: 28%-42%

Visual C++: 23%-28%

The Business Disk Winmark from Windows 98 doesn’t follow the general pattern. Since it is likely a big part of your nongaming computer use involves business/Web applications under Windows 98, this is not good.

The Front Page scores are simply not believable, period. I have scores for Front Page that significantly exceed the theoretical maximum two ATA66 drives could possibly put out.

Additionally, Front Page is the type of application that doesn’t use big chunks of files, so the scores should be at the bottom end of the speed spectrum, not four or five times better than anything else. Since those scores are used in the HE average, it pretty much invalidates that average too for comparative purposes.

Even worse, though, you cannot use the High End benchmark as a general predictor of activity, even if you could take out the Front Page score. Some apps do much better running Win98, some do much better under Win2K.

For example, using base single-drive scores, this is how Windows 2000 compares to Windows 98:

AVS Express 3.4: 102% better

Microstation SE: 103% better

Photoshop 4.0: 8% worse

Premiere 4.2: 21% worse

Sound Forge: 57% worse

Visual C++: 25% worse

When you have such major variances, an average is worthless. If you just looked at the overall number, Win 2000 looks pretty good. If you mainly use Sound Forge, though, you are in a heap of performance trouble.

Business Winstone and Content Creation:

I got between a 3-6% improvement in Win 2000 scores, about a 3-4% decrease in Windows 98 scores.

An Alternate Approach

After seeing all these inconsistencies, I yearned for something better. So I came up with two simple tests.

First, everyone has to load their operating system, and that takes more time than loading anything else. People usually want to boot faster, and would expect products like this one to do just that. So what I did was time the boot process in both Windows 98 and 2000 with RAID and without to see how much they took.

Boot-up:

The IWill RAID card adds time throughout the bootup in Windows 98: at initialization, in recognition of the device, and in loading the OS. Without the card, bootup takes about 45 seconds, with it, the boot takes about 1:10.

On the other hand, it does seem to take less time to boot into Windows 2000.

Application loading:

Photoshop takes a relatively long time to load, so I measured how well it loaded with RAID and without, under Win98 and 2000. After the first time it was used, loading took 4-5 seconds across the board, so no great improvement there.

The Big Order:

RAID is supposed to handle big files better than a lot of little ones. So what I did was use a Photoshop benchmark (PS5bench) to generate a random file, resized the picture several times. By the time I was done, I had 10Mb, 40Mb, 60MB and finally 160MB files. I saved those files, noting the time it took to save, then frequently opened and saved the files and measured how long it took. I did so both immediately and after reboots, made no difference in the results.

Again, there were inconsistencies, especially with the smaller files, but the RAID setup definitely saved a lot of time opening and closing bigger files; especially in Win2K, where operations often took only about half the time with RAID that they did without. (In Win98, time savings were more like 20-25%).

So if you handle big files a lot, this could be a real boon to you.

But . . . .

If you are regularly using these kinds of files, there’s a good chance you are using NT or Windows 2000. If you are, both operating systems let you implement this kind of RAID in software, and the synthetic benchmarking done elsewhere seems to indicate that it does much the same thing as a RAID card, at the expense of somewhat more CPU time.

And for you overclockers . . .

This may be a quirk of my system, but I could not get the RAID system to run at more than 133Mhz, though it runs fine at over 140Mhz without it.

What should I do?

Probably the easiest way to handle this is to say what it will and will not do for you.

Will my machine generally seem faster? No.

Will it load my operating system faster? Probably not.

Will it load my games and apps faster? Probably not.

Will it load big files for me to edit faster? Yes.

Will it back up files for me reliably? Yes.

Final thoughts:

I’ve owned an IWill 2935UW SCSI controller for a while, and it has always worked flawlessly. I had high hopes for this card which were not met. However, this doesn’t mean the IWill card is bad or the Promise card (hacked or not) is necessarily better; it just means there are inherent limitations to RAID and I set my hopes too high. The card certainly is advantageous for some situations, just not all.

The far greater disappointment is the unreliability of the disk benchmarks being used. Not only in my own testing, but in comparing my results to others. There are just these occasional huge discrepancies, not just with a RAID setup, but just running single disks. It makes you wonder how valid some of these benchmarks are.

That doesn’t mean all benchmarks are bad; the general ones like Winstone and Content Creator were far more consistent. Nor is Winbench necessarily a particularly bad one. For instance, I could change the SiSoft Sandra Disk Benchmark by 20% just by changing the partition being measured.

It’s likely that hard drives are much harder to measure consistently than other pieces of equipment due to their nature (constantly changing storage), and RAID just exacerbates the problem.

Email Ed