Database Reference
In-Depth Information
to a planned maximum capacity of 2 TB. Initial products are expected to be
write-once read-many with future plans to handle rewritable media.
1.7.6 Direct Molecular Manipulation
Continued improvements in cost-effectiveness of devices is closely related to
storage densities of the medium. As we press toward areal densities that ap-
proach the atomic scale, there has been increased interest in technologies that
can encode data by directly manipulating atomic structures. Direct mechan-
ical manipulation attempts to make novel technology cost-competitive with
current storage devices by leapfrogging their areal densities. Some examples
include direct manipulation of atoms using a scanning tunneling microscope
(STM), nanotube devices that depend on the van der Waals interaction be-
tween crossed tubes, and IBM's Millipede device (shown in Figure 1.5) that
uses many parallel micromechanical (MEMS) styli to encode data into a poly-
mer medium. Each of these devices promises storage densities that exceed the
current magnetic limit of 1 Tbit/mm2. However, such devices are still in their
infancy.
1.8 Summary and Conclusions
Mechanical magnetic storage devices such as disk and tape have been the
dominant technology for secondary storage for the past 30 years. Although
solid state devices such as FLASH have been gaining ground over the past few
years, areal density and cost trends ensure that disk will remain competitive
for at least the next decade.
One important trend that may complicate future storage for the highest-end
scientific computing system is the growing gap between disk capacity and the
delivered bandwidth of these devices. Storage trends continue to show 40-60%
per year compounded growth rate for storage capacity, thanks to continued
improvements in areal density. However, the bandwidth delivered by these
same disk subsystems is only growing by 17-20% per year. The performance
of such systems for random access has become nearly stagnant, which favors
linear streaming read or append-only write operations. If these trends continue
unchanged, HPC systems will be forced to purchase larger numbers of disk
spindles over time that are accessed in parallel in order to maintain existing
balance ratios between the HPC system performance and storage subsystem
bandwidth. As such, the disk subsystem will likely consume a larger fraction
of the area and power budget for future systems without some technology
change.
As a result of requiring more disks to achieve performance requirements,
HPC centers are deploying file systems that are for the first time eclipsing
