Database Reference
In-Depth Information
the size of archival storage systems. Considering the sustainability paradox,
presented in Section 3.2.1, it is likely that users will need to more carefully
consider how best to use archival storage systems to manage their most im-
portant data.
In addition, as the number of spindles continues to increase, the likelihood
of device or even RAID system failures occurring increases, as does the rebuild
time for RAID systems that are able to recover. These are challenges that have
recently been addressed with innovations in offering new levels of RAID. This
problem is discussed in some detail in Chapter 2. However, at some point
storage devices will not be able to handle error detection or recovery on their
own and will require innovation in other technologies to solve (e.g., file system
checksums).
Several technologies emerged to keep power consumption under control and
to fill the widening gap between primary storage (DRAM) performance and
secondary storage (disk) performance, such as MAID and FLASH.
Solid state non-volatile random access memory (NVRAM) technologies such
as FLASH are becoming cost competitive with the high end of disk technol-
ogy and may soon reach parity with consumer disk storage due to dramatic
rise of mass-market applications. NVRAM technologies address issues of poor
response to random accesses due to lack of mechanical “seek” cost, power
dissipation, and bandwidth scalability of conventional disk devices. However,
cell wear-out and poor write performance of FLASH relative to the mechan-
ical devices keep impact on high-end scientific computing storage marginal.
While load-leveling technology offers some protection against wear-out, the
algorithms are still subject to edge cases where intensive recopying of data
is required. In the interim, FLASH may offer advantages for read-intensive
storage applications such as data-mining applications, but for write-intensive
applications (such as data output from time-evolution simulation codes or
checkpoint/restart) we may need to wait for commercialization of alternative
NVRAM technology that doesn't suffer from cell wear-out such as MRAM,
or phase-change devices.
Tape subsystems are also seeing new demands that run counter to their orig-
inal performance characteristics. Formerly, there was emphasis on streaming
performance of tape systems for moving small numbers of very large files.
However, over time archival storage, even at scientific computing facilities,
has been increasingly dominated by large numbers of small files. For a sequen-
tial access medium, managing many small files using current tape technology
poses daunting technical challenges, especially as older, smaller capacity de-
vices are replaced by media that hold more and more data. Tape speed is
stable, which is desirable as increased speed means increased wear-and-tear
on the tape. User wait time to first byte of data will continue to increase
linearly as tape media capacity increases. Current high-performance archival
storage management software, such as the high-performance storage system
(HPSS), are historically geared toward handling large files where data can be
streamed to tape. HPSS is planning to deliver a feature to enable aggregation
