COVER FEATURE A Survey of Storage Options Over the past 50 years, technological innovations have continually driven down computer storage costs while dramatically increasing performance and value—and the trend shows no signs of slowing. John P. f you were a personal computing enthusiast in In 1970, Fairchild Corp. invented the first 256- Scheible 1984, you could buy an IBM PC for $5,000 to Kbyte SRAM; by 1972 it was the best-selling semi- IBM $6,000—about $8,000 to $9,500 in today’s cur- conductor memory chip in the world, superseding rency. For this you would get a 0.004-GHz 8- the magnetic core memories that computers had I bit processor with 0.064 Mbytes of RAM, a 12- used since 1947. inch monochrome text-only display, a 0.16-Mbyte Both types of RAM are volatile—they lose their floppy drive, and no possibility of a hard drive. contents when the power shuts down. DRAM Compare that with a PC purchase today: a 2- requires thousands of refreshes per second. SRAM GHz 32-bit processor with 2,000 Mbytes of RAM, is faster because it does not need to be refreshed. If a 128-Mbyte video graphics card, a 700-Mbyte all things were equal, main memory would use CD-RW (and maybe a DVD-R), a 17-inch 1600 × SRAM, since it is faster than DRAM and does not 1200 display with 32 million colors—all for maybe require refresh circuitry to protect data from cor- $1,500. The old machine cost six times as much for ruption. However, SRAM is physically larger than 1/2000th of the processing power. The purchasing DRAM and it costs considerably more, making it power gain over 18 years is 1,200,000 percent. impractical for main memory, although it is still use- ful for relatively small caches. THE FORMS OF STORAGE Asynchronous and synchronous DRAM. DRAM can be Very few industries have shown such dramatic either asynchronous or synchronous. With an asyn- increases in value and function and simultaneous chronous interface, the processor must wait idly for reductions in cost as the computer storage indus- the DRAM to complete its internal operations, which try. The technology has evolved continuously over typically takes about 60 nanoseconds. With syn- the past 50 years both across different media and chronous control, the DRAM latches information within each one. from the processor under the system clock’s control. We can categorize primary storage as volatile ran- Years ago, most PCs came with asynchronous dom access memory, which loses its data upon fast-page-mode DRAM, which ran at speeds be- power loss, and nonvolatile memory, such as flash tween 80 and 100 ns. Extended-data-out DRAM and read-only memories, in which data persists. improved speed by about 20 percent. However, Data transfer forms include floppy disks, CDs, escalating CPU and motherboard bus speeds out- DVDs, and flash memory cards. Fast random access stripped the ability of both FPM and EDO to deliver forms include disk drives, flash memory, and the data in a timely manner. As system speeds increased, slower DVD-RAM. Data archive forms include especially when 66-MHz memory buses became tape, CD, and DVD. standard on PC motherboards, FPM and EDO DRAMS dragged effective speeds down by forcing Random access memory CPUs to wait to receive data from memory. The two basic types of RAM are dynamic RAM Once it became apparent that bus speeds would and static RAM. IBM’s Robert H. Dennard created have to run faster than 66 MHz, DRAM designers the one-transistor, small-capacitor DRAM in 1966. needed to overcome the significant latency issues 42 Computer 0018-9162/02/$17.00 © 2002 IEEE that still existed. They did so by implementing a The bus, or data path, consists of four 8- synchronous interface, which also offered other bit paths, allowing four separate data IBM and 3M advantages. requests at the same time. Because the mem- introduced tape The current standard, Synchronous DRAM, is ory modules are intrinsic parts of the data synchronized to the system clock. RAM accesses path, none of the three memory sockets can 50 years ago in occur in burst mode, which means that memory remain empty. Therefore, a socket without 12-inch reels accesses occur in bunches, with the first of the memory must hold a continuity module, a that stored bunch requiring more time to allow for setup. special board that carries the clock and data 1.4 Mbytes of data. SDRAM is tied to the system clock and is rated by signals. Unlike earlier memory modules, megahertz instead of nanoseconds. It must be rated which had to be the same size, Rambus to run at least as fast as the system bus—preferably allows each socket to hold a different-sized a little faster. module. Current designs. There are many different memory Other features that distinguish Rambus from designs today. Synchronous Link DRAM uses pack- SDRAM technology include a lower voltage re- ets for address, data, and control signals to operate quirement (2.5 V as opposed to 3.3 V), a much on a faster bus than standard SDRAM—up to at higher clock frequency (200 MHz to start, projected least 200 MHz. SLDRAM operates the output sig- to move rapidly to 800 MHz and 1.6 GHz), and nal at twice the clock rate. This puts the output greater difficulty in manufacturing. operation as high as 400 MHz, with some engineers The manufacturing problem revolves around the claiming it will reach 800 MHz in the near future. tight tolerances that Rambus requires. Timing and Double Data Rate SDRAM—also known as data transmission signals must travel at precise rates. SDRAMII—does precisely what the name implies. This requires all traces to be exactly the same length We can represent a clock cycle by a square wave. with the same impedance. Such precision is difficult SDRAM uses only one of the wave’s edges, to mass manufacture, and it accounts, at least partly, but DDR SDRAM references both to effectively for the long delays in Rambus product releases. double the data transmission rate. Unlike 168-pin The immediate competition in memory technol- SDRAM, DDR SDRAM uses a 184-pin plug. ogy will focus on Rambus versus SDRAM. Ram- Although DDR SDRAM does not require changes bus must overcome manufacturing and cost battles in the basic motherboard technology, it is not back- in time to compete. Otherwise, ongoing innova- ward compatible on motherboards designed for tions in SDRAM technology ensure its continuing SDRAM. market lead. Compared with DDR SDRAM, SLDRAM seems to offer a much better alternative. Its actual clock Flash memory speed is lower, reducing signal problems. Latency RAM’s volatility makes it suitable only for tem- timings are shorter. So are costs due to the royalty- porary storage. Flash memory, on the other hand, free design and operation on current bus designs. retains its contents when power is lost, making it SLDRAM’s bandwidth is also much higher than ideal for storage. Flash memory, the essential com- DDR SDRAM at 3.2 Gbps versus 1.6 Gbps. ponent in solid-state hard disks, retains data with- SLDRAM is an open standard adopted by the out a power source. Flash memory has a size IEEE. It uses a narrow 16- or 18-bit bus with packet advantage (much smaller than a floppy, CD, or disk addressing that effectively increases the bandwidth. drive), making it the preferred storage option for The technology focuses not on the memory itself, small portable devices such as cell phones and dig- which is largely unchanged from standard DRAM, ital cameras. but on the interface, or protocol, which makes mul- However, flash memory is also relatively slow to tiple simultaneous accesses possible, as long as they write, which limits its applications. If it can over- are all to different physical locations in the RAM. come the write delays, which is likely, flash mem- The current battle for memory market share cen- ory can continue to expand its market. ters on DDR SDRAM and Rambus. RDRAM versus SDRAM. Rambus is a new system Tape design that requires major changes in the bus struc- IBM and 3M introduced tape 50 years ago in 12- ture and the way the RAM carries clock signals. inch reels that stored 1.4 Mbytes of data. Today Also known as RDRAM (Rambus DRAM), high-end tape holds 1 Gbyte of data. Rambus traps the clock signal, rather than broad- Recent price decreases and capacity increases in casting it, as in previous SDRAM designs. disk drives have made them competitive with tape December 2002 43 RAID Storage IBM received the first patent for a disk array subsystem in 1978. The company worked with the University of California, Berkeley, to define levels for redundant array of independent disks and released the initial RAID definitions in 1987. Although RAID is not storage in itself, it does ciently over large distances for backup purposes change the nature of storage by adding tolerance of disk drive failure and allowing for secure and centralized backup, espe- changing the access times to read and write data. cially for disaster recovery purposes. As the “RAID The major levels are RAID-0 through RAID-6. Storage” sidebar indicates, data managers can increase reliability by using RAID—redundant array • RAID-0 (striping) accesses all drives in parallel. It provides high of independent disks. Although RAID does increase data rates, but the data rates come at the expense of reliability. reliability, it does not protect against disasters like Because RAID-0 writes the data across multiple drives, it provides the terrorist attacks of 11 September 2001, when no redundancy.