Post-compiler software optimization for reducing energy

Abstract

Modern compilers typically optimize for executable size and speed, rarely exploring non-functional properties such as power efficiency. These properties are often hardware-specific, time-intensive to optimize, and may not be amenable to standard dataflow optimizations. We present a general post-compilation approach called Genetic Optimization Algorithm (GOA), which targets measurable non-functional aspects of software execution in programs that compile to x86 assembly. GOA combines insights from profile-guided optimization, superoptimization, evolutionary computation and mutational robustness. GOA searches for program variants that retain required functional behavior while improving non-functional behavior, using characteristic workloads and predictive modeling to guide the search. The resulting optimizations are validated using physical performance measurements and a larger held-out test suite. Our experimental results on PARSEC benchmark programs show average energy reductions of 20%, both for a large AMD system and a small Intel system, while maintaining program functionality on target workloads.

The next 700 CPU power models

“…Beyond these deployment issues, we also aim at extending this open testbed to consider other power-consuming components, such as GPU [54] and disk, in order to incrementally learn their power model and thus provide wider cartography of the power consumption of a software system. In the future, we believe that PowerAPI can be a cornerstone to new energy-aware scheduling [4,[15][16][17]19], to energyproportional computing [20][21][22][23], to new kind of optimizations [24], and to a better understanding of the power consumption drawn by software [25][26][27].…”

“…23 inst_spec_exec_integer_inst counts the integer data processing instructions speculatively executed. 24 bus_cycles is not officially documented.…”

“…Beyond the CPU power models and experiments we report in this article, we believe that our contribution offers an open testbed to foster the research on green computing. Our open-source solution can be used to automatically infer the power models and to support the design of energy-aware scheduling heuristics in homogeneous systems [4,[15][16][17][18], as well as in heterogeneous data centers [19], to serve the energy-proportional computing [20][21][22][23] and to evaluate the effectiveness of optimizations applied on binaries [24]. It also targets system administrators and software developers alike in monitoring and better understanding the power consumption of their software assets [25][26][27].…”

Optimizing Tradeoffs of Non-Functional Properties in Software

“…We conducted a preliminary investigation into using an energy model, but found that the potential for model inaccuracies can have significant negative impact. Specifically, using the model suggested by Schulte et al [155], we identified a variant of freqmine with over 70% predicted energy improvement, but only 2% actual improvement. To avoid misleading the search to such a significant degree, we decided to use physical measurements in our fitness function.…”

“…We choose to manipulate programs written in assembly language. Assembly language is commonly available as a compiler output format [29], permits optimizations in the spirit of instruction scheduling [73] and superoptimization [82], and has previously been successfully used as the basis for post-compiler optimizations [155].…”