Hardware Reference
In-Depth Information
FIGURE 29.3: The energy consumed by data movement is starting to exceed
the energy consumed by computation.
the energy needed to move the data exceeds the energy used in performing
the operation on those data.
3
(See Figure 29.3.)
By 2018, further improvements to compute eciency will be hidden by
the energy required to move data to the computational cores on a chip.
The current computing infrastructure is built on the premise that comput-
ing is the most expensive component. In the current era, computing is cheap
and ubiquitous, whereas data movement dominates energy costs. This over-
turns basic assumptions about programming and portends a move from a
computation-centric paradigm to a data-centric paradigm for programming
future systems [2]. The severe constraints in data movement are also driving
the industry toward alternative technologies for data communication such as
optics and silicon nanophotonics that integrate optical technologies directly
onto silicon chips.
29.2 Implications for the Future of Storage Systems
Data-Centric Computing and Non-POSIX I/O: Whereas current
models focus on equal partitioning of computation and moving the data to
3
This observation is explained in more detail in D. A. B. Miller and H. M. Ozaktas,
\Limit to the Bit-Rate Capacity of Electrical Interconnects from the Aspect Ratio of the
System Architecture,"JournalofParallelandDistributedComputing41, 4252 (1997).
