The Test System:
Our report is aiming toward the usual desktop user; we therefore compare our 3 test disks first of all as single disks. As more and more users are RAIDing their systems with the handily available on-board RAID controllers, we will also test them under RAID 0. We do not include server performance. As baseline we also include test results with our ATA 100 reference disk (Western Digital Caviar 400BB). Also in reference to what most desktop users are still using, we ran our tests with the FAT 32 file system. Readers, who compare their benchmarking results with our tests, should be aware that tests running under FAT 32 are usually returning up to 5% better scores than when running under NTFS. We disabled all hard disk functions like noise dumping or SMART monitoring, which may negatively impact performance.

The Benchmarking Tests:
We used eTestingLab's WinBench 99 version 2.0 as main benchmarking tool. WinBench 99 has become a kind of industry standard for system benchmarking and can be downloaded for free here. WinBench 99 is a subsystem-level benchmark that measures the performance of a PC's graphics, disk, and video subsystems in a Windows environment. We used 2 parts of the WinBench 99 test suite:

• The "Disk Inspection Test" measures disk transfer rates and access time.
• The "High-end Disk Mark" measures disk performance with applications.

The High-end Disk Mark is based around the following 7 read and write intensive (I/O bound) applications:

• Adobe Photoshop 4.0
• Adobe Premiere 4.2
• Sound Forge 4.0
• Visual C++ 5.0
• Micro Station SE
• FrontPage 98
• AVS/Express 3.4

The 99 "High-end Disk" test suit is returning individual scores for the above listed high-end applications and an overall "High-end Disk" score. These scores are representing "real-world performance" better than only disk transfer rates. Another advantage is that the impact of RAID 0 can be clearly seen with the subtests of the High-End test suit. Because many readers are using SANDRA for benchmarking their system, we took the SANDRA 2002 "File System Benchmark" as well. This however with certain reservations as this SANDRA benchmark is known to show wide fluctuations when benching RAID 0 arrays.A final test was done with Ziff Davis Content Creation 2001. This is a system-level, application-based benchmark that measures a PC's overall performance when running Windows-based, 32-bit, content creation applications. It uses the following applications:

• Adobe Photoshop 6.0.1
• Adobe Premiere 6.0
• Macromedia Director 8.5
• Macromedia Dreamweaver UltraDev 4
• Microsoft Windows Media Encoder 7.01.00
• Netscape Navigator 6/6.01
• Sonic Foundry Sound Forge 5.0c (build 184)

Content Creation 2001 is by far not as strong as the WinBench 99 disk tests discriminating between hard disk performances, because it measures overall system performance and hard disk performance is naturally only a small part of the score. But on the other hand: CC 2001 shows more "realistically" the impact of differences in hard disk performance on overall system performance than a "hard-disk-only" test like the "High-end Disk" suit in WinBench 99. CC 2001 has to be ordered on CD from eTestingLabs.

We are NOT using the well-known HDTach 2.61 hard disk benchmark tool. Although HDTach 2.61 is fine for taking transfer rates for single disks, we found 2 major issues with this utility: the disk access time is obviously calculated in a rather strange way which makes it incomparable what the manufacturer's rating and other benchmarks. That is what INTEL has to say about this issue here. The other issue is with RAID 0. We found that HDTach 2.61 is showing far too low transfer rates for RAID 0 arrays.

RAID 0 Stripe, File Cluster Size, and PCI Latency:
A word regarding the ongoing discussion about the optimal cluster and stripe size for RAID 0 operations: As shown in this BBS thread here it seems there is no single optimal solution for every RAID controller and every hard disk. The current status of the discussion looks like this:

RAID 0 block stripe size:
Users report generally best results with a 16K or 34K stripe size. Anything above 34K seems to reduce performance and is not recommended unless the array is used for working with exceptionally large files such as in video editing.

File cluster size:
It seems that for usual desktop tasks the Maxtor D740X performs slightly better when the file system is formatted in 34K clusters. The Western Digital JB series seem to prefer 16K or even 8 K. We have tested the 1000JB and the 120GXP in 8K/16/ and 32K and found only random differences. It seems that a RAID 0 stripe size of 16K and a file cluster size of 16K is the best overall compromise for desktop jobs. We used therefore a 16K stripe size / 16K cluster size combo for our tests. User who have only 2 disks available and wish to format their RAID 0 array which contains the OS with an other than the default size of 4K in NTFS will find here a how-to-do.

PCI latency:
Already a while ago it was discovered that VIA chipsets have obviously a problem handling the increased burst speed of high-end ATA disks. They start to interrupt the burst shortly after the cycle has started. These interruptions taking place every 1.2 µs reduce transfer rates of high-end ATA 100 and ATA 133 disks. An in-depth explanation can be found here at techchannel.de which discovered the issue shortly after boards with the KT266A chipset started retailing. In particular for optimal RAID performance with VIA chipset it is necessary to lower PCI latency. This can be achieved by patches lowering the PCI latency for the devices attached to the PCI bus by making adjustments to the chip that manages the PCI bus. VIA is offering a "Raid Performance Patch" here. There is also a non-VIA PCI latency patch, named after the author, "Breeze patch" available here. Both patches increase the disk transfer rate by approximately 10 MB/sec. According to many users and our own results the Breeze patch is performing slightly better than the VIA "RAID Performance Patch". We used for our tests version 1.9d of the Breeze patch.