Hardware Reference
In-Depth Information
the processing elements, data-centric models provide facilities to describe how
the data is laid out on the system and apply the computation to the data
where it resides (in situ). There are numerous software-based approaches be-
ing pursued that enable data to be operated upon where it was produced,
or while in transit. Even more aggressive technological approaches try to em-
bed computing capability into the storage devices. This includes everything
from embedding intelligence directly into disk systems such as the Netezza, to
the re-emergence of processor-in-memory and processor-near-memory tech-
nologies. Chapter 30 explores the fact that all of these innovative solutions
are fundamentally orthogonal to the standard POSIX view of I/O consis-
tency, and open the door to new object-storage approaches that obey a more
transactional or \functional language" semantics to preserve consistency and
correctness.
Data Movement: The data movement challenges may also accelerate de-
velopment of nanoscale optical/photonic solutions. Although much of our data
transmission within a datacenter still uses wires, the intrinsic resistance of the
material used by \wires" (electrical connections of some form) limits any so-
lution involving electrons. The principal problem is that the energy consumed
to transmit a bit is proportional to the distance it must travel (due to the
resistance of the metal used to conduct the electricity). Copper is as good a
conductor as can be expected for a common material at room temperature.
For communications (wire replacement), photonics has the benefit of hav-
ing energy costs that are nearly independent of the distance that data travels,
whereas the standard electrical wires have a strong distance-dependent energy
cost. Therefore, photonic technology has enormous potential to overcome the
fundamental limitation of wire resistance, if they can overcome the challenge
of improving the eciency of laser light sources.
Photonics offers huge bandwidth density improvements. Whereas an elec-
trical communication channel can only carry one signal, multiple wavelengths
of light can share the same optical conduit without interference using Wave
Division Multiplexing (WDM). Photonics will play an essential role in over-
coming limits of wires and break through the boundary of on-chip and off-chip
communications costs, and cost/distance relationships [3]. The implications
for dense wave division multiplexing (DWDM) using photonic device technol-
ogy for interconnects is explored in more detail in Chapter 32.
Resilience: Lastly, nanoscale components with aggressive scaling in volt-
ages, currents, photons, etc. can raise energy eciency, but the probability of
failures, and more critically soft errors, will also rise. Creating systems that
are inherently resilient to both hard and soft failures is perhaps the most
daunting challenge for exascale computing. Worse yet, we depend on our stor-
age systems as the \safe location" to recover application state for the most
commonly used application resilience methods, so the requirements for the
resilience of storage systems are far higher than that of the compute devices
to which they are connected.
