2018-annual-report

Narrowing the Data/Compute Gap
with Specialized Hardware
29
Datacenters are facing a challenge because the quantity
of information that needs to be stored and processed is
growing faster than the performance of general purpose
processors. For decades, this has been increasing as per
Moore’s Law, but today it is unclear whether the trend
can be maintained. The shrinking of transistor sizes has
slowed down significantly and even though it is possible
to add more transistors to a central processing unit (CPU),
using them to create additional cores is unlikely to benefit
applications unless they are trivially parallelizable.
In order to change the status quo, we need to investigate
how software interacts with the underlying hardware and
explore ways in which we could tailor the latter to the
application’s needs. As an alternative to adding conventional
cores, we could use part of the chip for specialized
processing elements. When tailored to widely-used application
domains in datacenters, these elements increase
overall processing efficiency and could be used to narrow
the gap between data growth and compute capacity.
databases, blockchains, etc.) and emerging distributed
data-intensive applications that are often bound by the
processing power of CPUs (e.g., business analytics,
machine learning, etc.) The most important challenges of
this research direction revolve around the goal of ensuring
that, while we benefit from the use of novel hardware, the
flexibility, reliability, as well as the security guarantees,
of applications are not impacted. In our exploration we
employ rigorous analysis methods, build proof of concept
software systems and even prototype specialized functionality
in hardware, using Field Programmable Gate Arrays
(FPGAs).
annual report
20
There are already several types of programmable hardware
devices appearing in the datacenter and consumer clouds,
which makes this an exciting time to be working in the
field of systems. Emerging low latency networks with programmable
network interface cards, for instance, allow
distributed applications to change their communication
model and various types of programmable hardware accelerators
allow compute-intensive tasks to be carried out
faster or in a more energy efficient manner. The shift away
from a “CPU-only” view, however, requires us to devise
better methods for software to take advantage of, or even
to directly drive the design of, novel hardware features.
Stagnating CPU performance (approximated here by singlecore
frequency) is limiting our ability to process the increasing
amounts of data we produce. Using more specialized hardware
is one promising direction to close the gap between data and
computation.
At our institute, we explore research questions related to
integrating programmable hardware accelerators in data
management systems that suffer from various forms of
data movement bottlenecks (e.g., large-scale distributed