Options range from software RAID to IDE/ATA and SCSI RAID to chip-level implementations.
By Margit Evensta
The demand for low-cost, high-density application servers, including blade servers, is taking off rapidly. Many of these systems, typically housing two to six internal drives, already play roles at ISPs, in cluster node installations, or as network-attached workstations. Market research firm International Data Corp. predicts the high-density server market will jump from 450,000 units last year to more than 700,000 units in 2004.
For IT storage administrators, these compact servers present new opportunities and new challenges. On one hand, the compact, rack-mounted servers offer advantages in terms of cost and footprint. Yet users still require the same level of fault tolerance and reliability found in higher-end servers. While full-featured RAID subsystems could provide all the redundancy and availability needed to support these systems, they typically come at too high a cost for servers that often sell for less than $3,000. In addition, a full-featured RAID system will occupy far more system real estate than these highly compact servers.
How can IT storage managers reconcile the need for comprehensive RAID capability with the cost and space constraints of these new systems? Clearly, highly integrated, low-cost RAID offers the best fit for this environment. But a wide variety of technologies are available to serve this purpose. Here's a review of the integrated RAID options available for these platforms and the key tradeoffs.
From a pure cost perspective, hardware-based RAID configurations can't match the advantages of software RAID. Available from operating system vendors and third-party suppliers, software RAID offers many of the features and functions of hardware RAID, but at a lower cost.
Software RAID is also easy to implement. Administrators do not have to open the server chassis, install cabling, or contend with host IDs. Administering and upgrading software RAID is also simpler than hardware-based alternatives. In addition, software RAID is very flexible. Since it is tightly coupled to the operating system, it allows the use of any supported disk interface in an array, and migration to new interfaces is easy.
But the software-only approach has drawbacks. One problem is its reliance on the host CPU. Since software RAID runs on the host processor and uses system memory resources, there is a performance penalty. Every compute cycle used to run software RAID drains bandwidth from the system's host processor and limits it from performing other tasks.
Perhaps a bigger problem with software RAID, particularly operating system-based approaches, is that it is not bootable. Since the RAID functionality is contained within the operating system, it cannot mirror the operating system itself. Software RAID offers no protection from errors within the operating system or host application. Therefore, it cannot provide as high a level of reliability as RAID implemented outside the operating system.
Perhaps the fastest growing RAID solutions are those based on PCI-IDE/ATA host bus adapters (HBAs). These relatively low-cost boards are available in either intelligent or non-intelligent versions. Intelligent HBAs use on-board processors to supply fault-tolerant functionality. Non-intelligent approaches rely in part on the host processor. IT managers have to carefully consider the tradeoffs.
Non-intelligent RAID costs less but consumes CPU resources. These boards sell for a fraction of the cost of comparable SCSI-based PCI RAID systems and offer users a wide array of RAID configurations. These subsystems offer excellent flexibility for their price, but they can't perform at the same level as a comparable SCSI configuration. For example, while PCI-IDE RAID can support data rates up to 133MBps, the latest generation of SCSIUltra320reaches speeds up to 320MBps.
Newer, intelligent PCI-IDE RAID is closing the functionality gap with SCSI implementations. Some of these systems present themselves to the operating system as a SCSI RAID device and offer many of the same advanced features. However, intelligent PCI-IDE costs more than non-intelligent PCI-IDE cards, and performance is still limited by the throughput of the IDE interface.
While PCI-IDE RAID is relatively inexpensive, there are additional issues to consider, depending on the particular system environment and the availability of in-house technical expertise. These cards, and associated drives, typically sell for much less than comparable PCI-SCSI solutions. But they still require a larger investment than software RAID, zero-channel RAID, or chip-on-motherboard alternatives. Disk-drive reliability differences between IDE and SCSI are also worth reviewing. The IDE interface does not offer the established code base, queuing, and support for hot-swap capabilities found in SCSI implementations.
Ease-of-use may also be a consideration. Most RAID approaches require users to load specialized drivers and configuration utilities. But some, particularly SCSI solutions, provide better manageability through a more robust protocol.
SCSI-based PCI RAID cards usually offer a higher-performance option. These HBAs support all RAID levels and provide all the functionality required for operating system and data protection. Performance levels vary, depending on which version of SCSI the controller supports as well as the caching scheme. At the high end, Ultra320 SCSI controllers support data rates up to 320MBps. But storage administrators pay a price for that additional capability: SCSI boards generally cost more than PCI alternatives.
PCI SCSI RAID controllers offer excellent reliability due to the extensive and proven code base built up over SCSI's 20-year history. From an installation and maintenance perspective, they have about the same requirements as IDE RAID. Storage administrators must still rely on application-specific configuration utilities and specialized drivers, but SCSI-based PCI RAID is more scalable than IDE RAID and allows users to increase the number of volumes and volume size as required.
A couple of newer technologies offer storage administrators additional ways to cut costs and save space in high-density servers. For example, "RAID-on-a-chip" offers comparable capabilities and performance as PCI RAID cards, while reducing cost, power, and space. These integrated ICs typically support all RAID levels and usually combine a RAID controller, drive-attach channels, and a memory controller on a single device. RAID-on-a-chip offers a highly compact and efficient motherboard solution.
Another emerging technologyzero-channel RAID (ZCR)combines many of the cost advantages of RAID-on-a-chip with the flexibility of an add-on card. These compact cards reduce the size and cost of a traditional RAID card by taking advantage of the disk-drive controller chip embedded on the system motherboard. PCI RAID cards typically feature multiple controllers to control storage devices. By using the motherboard controller instead, a ZCR add-in card can consist of just the I/O processor and buffer memory in a much smaller footprint.
ZCR cards plug directly into a small form-factor PCI connector that is linked to the PCI bus via specially designed logic. While these compact cards offer a cost-effective alternative, users may find them performance-limited compared to other PCI RAID approaches due to the increased PCI bus transactions required to transfer data in and out of the buffer. Still, the cost benefits of eliminating the SCSI chip on the card can be significant.
Controller-based integrated RAID
More recently, manufacturers have taken advantage of advances in chip design to build more-efficient, more-integrated single-chip solutions. These new devices typically rely on a multiple-processor design to increase performance. The additional resources can be used to address the processing demands of interfaces such as Ultra320 SCSI, Fibre Channel, and emerging proposals such as Serial ATA and Serial-attached SCSI.
Controller-based integrated RAID includes an integrated I/O processor for hosting RAID functionality and drive-attach channels. It reduces cost by eliminating the need for discrete logic on the motherboard. External memory requirements are minimal, and RAID 1 functionality requires as little as 32KB of NVSRAM. Since the controller doesn't have to contend with PCI bus transactions, it can out-perform a ZCR solution for RAID 1. It can also offer better reliability than software RAID by adding write journaling to ensure integrity of the mirrored volume across re-boots, operating system crashes, and power failures. By eliminating the need for a separate processor and buffer memory, it can also cost less than a PCI RAID controller, RAID-on-a-chip, or ZCR.
As the number of potential applications for high-density servers continues to expand, the demand for comprehensive, cost-efficient RAID functionality will also grow. Clearly, a wide variety of RAID technologies are available to address these issues. The key to finding the best solution for a particular application is to assess the relative performance benefits of each approach against issues such as cost, space, and ease-of-use.
Margit Evensta is a product manager at LSI Logic (www.lsilogic.com) in Colorado Springs, CO.