SANs CAN BRING TRUE DATA SHARING WITHIN REACH.

Magnetic disk storage represents the primary storage technology for most all of the world's mission-critical, performance sensitive, and online data. Since its inception in 1956 with the RAMAC 350, the evolution of magnetic disks has continued to produce rich features and increased functionality with each new offering. The RAMAC had a fourteen-inch diameter platter and contained 5MB. The new IBM Micro drive has a one-inch diameter and contains 340MB. The disk industry's future roadmap promises even more progress and advanced functionality.

From a pure technology perspective, the areal density race is accelerating and magnetic disk densities have now reached 26.5 gigabits per square inch. This compares to just over 13 gigabits per square inch as recently as eighteen months ago. At today's areal densities, an area on a disk surface the size of a child's thumbnail can hold the contents of five standard dictionaries. The super-paramagnetic limits that were expected to appear around the 20 gigabit per square inch point didn't occur. The super-paramagnetic limit theoretically occurs when the magnetic particles are so close together that they become unstable. We can now envision a 3.5-inch diameter magnetic disk with capacities up to 500GB in five more years with seek times in the 3-5 millisecond range. This obviously poses a significant access density problem, as the potential amount of traffic in and out of a single actuator becomes a bottleneck. [Note: access density is the ratio of I/Os per second per gigabyte of capacity].

Disk data rate improvements have consistently lagged the capacity growth trend line. The RAMAC transferred data at just 10KB/sec. The Library of Congress, now estimated to be 18TB in size, would take 2083 days to transfer at the original RAMAC data rate. Raw data rates directly from the drive could reach 60MB/sec within five years aided by disk spin rates reaching 15,000rpm. At some point, areal density and pure feeds and speeds progress will slow. Our long time fascination with feeds and speeds will give way to more meaningful factors that will become the competitive edge for future disk drives. Let's examine some of the most likely events that lie ahead for the disk storage industry.

Overall functionality of the disk subsystem has quickly become the primary differentiating factor in disk subsystem selection. As is already the case with most all magnetic tape drives, widespread use of loss-less data compression and compaction will further emerge to improve volumetric efficiencies and reduce storage costs while improving data transfer speed for magnetic disks. To date, the use of disk data compression techniques has lagged that of tape drives. Given the speed and much lower costs of microprocessors today, the overhead of inline data compression and decompression techniques is easily offset by the higher effective channel speeds. Architectural enhancements such as migration of data directly between SAN intelligent disk, tape, and possibly optical disks (also known as intelligent RAID or server-less backup/recovery) are being developed and will offload significant I/O traffic from the servers creating a seamless single-level hierarchical storage system. Server-less backup is now viewed as a "killer app" for the SAN marketplace and should become a de-facto standard function for the enterprise-level SAN within the next three years.

Adding storage intelligence along with the notion of a storage operating system may potentially become part of future disk controllers as the path toward SAN pushes knowledge of the data closer to the data itself and away from the heterogeneous and often non-compatible servers. It is not yet clear, however, where, within the SAN, specific elements of storage intelligence will reside. As various RAM chip technologies fall in price, become non-volatile, and increase in capacity, expect to see more RAM storage imbedded in conventional drives for indexes, directories, and specific-addressable content that will eliminate mechanical movement and further improve search and access times.

Dual actuator per disk assembly designs are now being explored as a potential solution to the current imbalance between the slower progress of disk performance and the 60% or better annual increases in capacity. At current areal density progress, we could expect a 500GB disk drive to appear by 2005. That is obviously a lot of data for one actuator to service and provide an acceptable level of performance. Dual actuators add costs and some complexity to today's basic drive. Adding cost has historically been hard to justify to anyone. The issue of whether each actuator should address a common surface area or have the capability of addressing separate areas on the disk remains open.

Virtual disks have been successfully implemented for several years beginning with the IBM 3850 Mass Storage Subsystem in 1974, which gave helical scan tape cartridges the image of multiple 3330 disks to the operating system. Virtual disks offer more functionality by dynamically allowing a physical drive to present multiple device images or personalities to the connected servers that can enhance or favor the performance of certain applications while enabling logical sharing. Virtual disks will play an even bigger role with SANs as a variety of non-similar physical disks can be virtualized into a storage pool presenting similar formats. Combined with virtual tape subsystems, virtual disks set the stage for breaking the bond of the server operating system understanding the physical geometry of its attached storage. The SAN breaks the long-standing relationship of directly connecting storage to a server; virtualization breaks the long-standing need for the server to know the physical device image.

As we can now see, the role of future disk subsystems in conjunction with SANs and the path toward the information utility are closely related to each other. The SAN breaks the dedicated physical connection of storage to the server, virtual storage breaks the need to support a specific storage format from the server, and data sharing, at last, gives us the ability to store and retrieve data from the storage subsystems with data integrity. It is now widely recognized that departmental servers and enterprise servers alike must be able to communicate with each other and to share data, as well. Having application and server platforms that utilize the same consistent version of data and information have created the perception that a single copy of that data is the most desirable arrangement. Data sharing is the action of accessing information from multiple applications or users.

There are three generally accepted levels of data sharing today with various sub-levels. Storage Sharing is the physical sharing of or partitioning of disk or tape subsystem hardware that provides simultaneous connection to multiple servers. They share channels, cache, bandwidth, disks, and MIPS, but not the same physical instance of data. Nearline tape libraries were the first systems to popularize this type of data sharing in the late 1980s. The second type of data sharing is called Data Exchange or Data Copy Sharing. The controller uses data-movers to copy data via a bus or channel to another control unit and creates a second copy of the data converted to the proper data format (ECKD, FBA, EBCDIC, ASCII, etc.) of the new server. The result creates two physical copies of the data, increasing the storage and management requirement. Updates need to be synchronized. The second copy is created without the expense of host MIPS or channel bandwidth.

The third type of data sharing is True Data Sharing. This method allows multiple servers to access the same data simultaneously and eliminates the need for data replication and the maintenance overhead of managing multiple copies of data. True Data Sharing allows any server to have the capabilities of direct read, write, and update access to source files or databases. True data sharing is evolving, but a careful understanding of the tradeoffs with each level of data sharing is crucial to any successful implementation. The path of disk subsystem progress over the last 45 years has been most impressive; the next 10 years should make all of this progress seem small by comparison.