Hardware Reference
In-Depth Information
34.1 Introduction
Many institutions are working toward building exascale
1
computers by
2018{2021. Along the road to exascale, there will likely be inection points
where current strategies for many aspects of a computing system (program-
ming models, resilience and serviceability, power delivery and management,
and mass storage) will no longer be adequate, and new strategies will have to
be adopted. Thought leaders in all of these areas are working to ensure that
we can effectively use exascale machines when they arrive.
Perhaps the largest looming barrier to exascale computing is the potential
power consumption of a machine. The most power-ecient machine at the
time of this writing can accomplish about 4.5 GFLOPS/W [1]. To be viable,
an exascale machine will have to be able to perform 50 GFLOPS/W [30].
Researchers are scrutinizing every subsystem to ensure that no power is being
wasted.
Secondary storage is receiving significant attention for two major reasons.
First, the cost of moving a block of data from disk to memory is comparatively
immense. Moving a byte from disk to main memory requires 100{1000 more
energy than moving a byte from RAM to a CPU cache [20], disregarding
network activity necessary for HPC I/O. Second, many HPC applications
perform significant I/O in the form of bulk synchronous checkpointing, where
the application stops computing to form a restart file on persistent storage.
For applications to make acceptable progress, these (often large) checkpoints
should be performed in a small fraction of the total computation time. This
implies that the I/O system has to appear fast to the user, while consuming
a comparatively reasonable amount of power.
This chapter begins by describing the state of power use of past and present
production supercomputers at a variety of sites. Following, the chapter will
extrapolate today's typical architecture and methods to exascale. Finally, this
chapter describes recent and ongoing innovations that have the potential to
improve I/O power use for exascale machines.
34.2 Power Use in Recent and Current Supercomputers
Historically, system researchers have not emphasized power consumption
for HPC storage systems specifically. Instead, power metrics were gathered in
aggregate, which included storage, compute, networking, and often cooling in
the same gure. To the authors' knowledge, the following subsections discuss
the largest survey of power use in HPC storage systems to date. At the time of
1
10
18
floating point operations per second (FLOPS).
