Software rarely uses all the

potential performance available in a modern microprocessor. For

example, on an Intel Core 2 class workstation—a microprocessor

capable of executing 4 instructions per cycle—the average instructions

per cycle for the single-threaded DaCapo benchmark suite is

0.98. In other words, even in the multi-core era, there is still enormous

potential to increase single-core program performance. Recent

work in PLDI has focused on multi-core performance: by our

count PLDI’10 has around 6 papers on multi-core optimizations

and PLDI’11 has 6 papers as well.

We believe this aggressive focus on multi-core may miss a critical

trend in computing environments: power, or performance per

watt. The issues of power consumption and thermal dissipation are

now major limiting factors in performance, even in environments

with unconstrained power and cooling systems (e.g., desktops or

servers). However, power consumption and performance per watt

are even more critical in mobile computing (e.g., phones or tablets)

and data centers, which are increasingly important computing environments.

We argue that, while research on multi-core optimizations

is valuable, improvements in single-core performance via

improved resource utilization are key to increasing performance

and doing so with minimal impact on power consumption.

In  PLDI FIT

Publisher  ACM

Details