Hardware Reference
In-Depth Information
Chapter 29
Parallel Computing Trends for the
Coming Decade
John Shalf
Lawrence Berkeley National Laboratory
29.1
Technology Scaling ::::::::::::::::::::::::::::::::::::::::::::::: 326
29.1.1
Classical Scaling Period (1965{2004) :::::::::::::::::::: 326
29.1.2
End of Classical Scaling (2004) :::::::::::::::::::::::::: 326
29.1.3
Toward Data-Centric Computing (2014{2022) :::::::::: 328
29.2
Implications for the Future of Storage Systems ::::::::::::::::: 329
29.3
Conclusion :::::::::::::::::::::::::::::::::::::::::::::::::::::::: 331
Bibliography :::::::::::::::::::::::::::::::::::::::::::::::::::::: 331
The broader technology industry has come to depend on the rapid, predictable,
and cheap scaling of computing performance and storage density. For decades,
exponentially increasing capability could be procured at roughly constant an-
nual cost, and that expectation has permeated computing resource planning
and decision making. For the past twenty years, we have become accustomed to
a very steady technological progression where improvements in silicon lithog-
raphy (Moore's Law) have translated directly into improved computing and
storage speed, density, and energy eciency. All of our assumptions about
how to program these systems implicitly assume the progression will continue
unabated. In 2004, however, a confluence of events changed forever the ar-
chitectural landscape that underpinned our current assumptions about what
to optimize for when we design new algorithms and applications. Storage
technologies must be refactored to handle massive parallelism, the severe con-
straints on the energy cost of data movement, and reduced reliability. Fur-
thermore, this confluence of factors presents a more fundamental challenge to
the traditional POSIX semantics that underpinned parallel file system design
for the past few decades.
325
