duf: Dynamic uncore frequency scaling to reduce power consumption

Abstract

Reducing the power consumption of applications has become one of the key challenges in high-performance computing. Recent processor architectures differentiate processor core frequency from its uncore frequency. As a consequence, in addition to tuning processor core frequency with Dynamic Voltage and Frequency Scaling (DVFS), power consumption can also be controlled through Uncore Frequency Scaling (UFS).This paper studies how the uncore frequency can be used as a leverage to improve power consumption. We propose DUF, a daemon process that dynamically adapts the uncore frequency to reduce an application power consumption with a user-defined limit on performance degradation.The evaluation of DUF on three different architectures shows that with no performance degradation (less than 0.6 %), DUF can reduce socket power consumption by 7.94 %. We also show that DUF is able to reduce the total energy consumption by up to 18.20 %.

Imitation Learning-Based Performance-Power Trade-Off Uncore Frequency Scaling Policy for Multicore System

“…The next phase is focused on building the UFS policy. The literature [ 12 , 15 , 16 , 17 ] uses a feedback-based single-step tuning-type algorithm for the design, where [ 12 , 15 , 17 ] decides the next action based on the impact of the current time-step uncore tuning action on the processor performance and power consumption. The literature [ 16 ] similarly evaluates the actual impact brought by the current action and decides on the subsequent action based on the feedback state.…”

“…The expert policy consists of three processor behavior prediction models and an optimal selector, and the output of it can be represented by Equation (5). Considering the important impact of uncore on the computational and memory access behavior of the processor, we combine the literature [ 15 , 23 ] to model the performance of the processor under this problem. The role of the optimal selector is to choose the frequency value corresponding to the highest power efficiency that the processor can achieve by changing the uncore frequency in the current state .…”

“…In this section, we will compare the performance of the UFS_IL policy with two advanced UFS policies on fixed load sequences. DUF [ 15 ] and ML_guided [ 13 ] policies represent two different design ideas. They are similar in design objectives, and we briefly describe these two approaches below.…”

Explicit uncore frequency scaling for energy optimisation policies with EAR in Intel architectures

“…Finally, we have compared the average DC node power variation and the average RAPL PCK (package) power when using the ME+eU policy. All the papers in the related work (see VII) use the RAPL PCK power as reference to evaluate energy and power savings as well as for developing their controller-based policy like [18], [19]. The PCK power is a non constant percentage of the DC node power, this is the reason why we think we must evaluate energy policies with DC node power.…”

“…They show results for some class C NPB and class D ECP proxy applications suites, running up to 16 Broadwell with two Intel Xeon CPU E5-2620 v4 @ 2.10GHz with eight cores each. Authors of [19] find a similar approach but this time they use memory bandwidth and FLOPS as metrics for their controller, but also checks for L3 cache bandwidth changes to detect phase changes. They also compare their result with the last mentioned [18], getting similar results on power savings but improving performance penalty on Broadwell and Skylake microarchitectures, using class C for most of NPB on first testbed and class D for the second one.…”