PMT Power Measurement Toolkit Stefano Corda1 Bram Veenboer2 Emma Tolley1 1EPFL Ecole Polytechnique F ederale de Lausanne2ASTRON Netherlands Institute for Radio Astronomy

2025-05-02 0 0 690.93KB 3 页 10玖币

侵权投诉

PMT: Power Measurement Toolkit

Stefano Corda1, Bram Veenboer2, Emma Tolley1

1EPFL ´

Ecole Polytechnique F´

ed´

erale de Lausanne, 2ASTRON Netherlands Institute for Radio Astronomy

stefano.corda@epﬂ.ch, veenboer@astron.nl, emma.tolley@epﬂ.ch

Abstract—Efﬁcient use of energy is essential for today’s super-

computing systems, as energy cost is generally a major compo-

nent of their operational cost. Research into “green computing”

is needed to reduce the environmental impact of running these

systems. As such, several scientiﬁc communities are evaluating the

trade-off between time-to-solution and energy-to-solution. While

the runtime of an application is typically easy to measure, power

consumption is not.

Therefore, we present the Power Measurement Toolkit (PMT),

a high-level software library capable of collecting power con-

sumption measurements on various hardware. The library pro-

vides a standard interface to easily measure the energy use of

devices such as CPUs and GPUs in critical application sections.

Index Terms—CPU, Efﬁciency, GPU, HPC, Performance

I. INTRODUCTION

Contemporary scientiﬁc applications such as particle hy-

drodynamics [1], radio-astronomical imaging [2] and Earth

digital twin simulation [3] have demanding compute require-

ments, in the order of the ExaFlops or more. Moreover, top

supercomputers [4] aim to reach Zettascale [5] performance

in the current decade. To achieve this goal, increasingly

more powerful hardware is used. However, due to the end

of Dennard’s Scaling law, these advances in compute power

come at the cost of a higher power consumption [6].

Energy efﬁciency is a critical challenge in HPC (High-

Performance Computing) as data centers are struggling to

operate within stringent power budgets [7] and also try to

reduce their carbon footprint [8]. Therefore, it is crucial

to monitor the power consumption of HPC applications in

order to evaluate trade-offs between achieved performance and

energy efﬁciency [9].

In this context, we present the Power Measurement Toolkit

(PMT), a lightweight high-level library capable of collecting

power measurements on various architectures, such as CPUs

and GPUs. PMT can easily be used to probe speciﬁc appli-

cation regions to measure power consumption and evaluate

energy efﬁciency.

II. POWER MEASUREMENT TOOLKIT

PMT’s structure is reported in Fig. 1. The library is written

in C++ and is Linux-only; it interfaces with hardware vendors’

API to collect power consumption. More precisely, for GPUs

it uses NVML [10] for NVIDIA and rocm-smi [11] for AMD.

CPUs are monitored through the RAPL (Running Average

Power Limit) [12] interface, or through LIKWID [13]. PMT

can proﬁle other architectures, such as Xilinx FPGAs, and

it can be easily extended to support new vendors’ hardware.

Some other architectures expose their power usage information

through ﬁles in sysfs (the /sys folder). PMT also has an

interface to physical power sensors such as PowerSensor2 [14].

Power Measurement Toolkit (PMT) NVML

ROCM

RAPL

...

Python Bindings

Python Decorators

Arduino

Fig. 1. Power Measurement Toolkit structure: the library interfaces with

different power measurement back ends. The library also includes Python

bindings and decorators to ease the measurement for application users.

PMT library’s core consists of a background thread to

the proﬁled application that communicates and gathers power

consumption information from the selected back end, such

as NVML. PMT sampling frequency is dependent on the

hardware and back end. For instance, NVML is able to sustain

up to 10 ms and RAPL up to 500 ms.

PMT can be installed using CMake or a Spack [15], [16]

recipe [17]. The library works in two modes: 1) dump-mode

and 2) measurement-mode. The ﬁrst one writes into a ﬁle

timestamps and power measurements to be able to examine

the power consumption over time of an application. The

second one simply provides the proﬁled code’s average power

consumption for quick energy-efﬁciency estimations.

PMT can be integrated into C++ and Python applications.

In C++ the code must be instrumented as shown in Listing 1.

Using PMT in a C++ application is a three-step process:

1) include the PMT header; 2) initialize PMT; 3) surround

the region of interest with PMT activation and deactivation

function calls. This can be done similarly in Python by using

Python bindings. However, we provide a simpler interface

using Python decorators (see Listing 2). As shown in Listing 2,

the Python code only requires an import statement of the

PMT library. In the case mentioned, we are collecting the

measurement by using the NVML back end for NVIDIA

GPUs and RAPL for AMD CPUs. The library support multiple

decorators at the same time (see Listing 2).

Typically, PMT has a small overhead in the order of 1 ms in

C++ and 10 ms in Python. This overhead is cumulative when

multiple decorators are used.

The Listings 1 and 2shows the above-mentioned

measurement-mode and how to programmatically query

the measurement after the region of interest. Some changes

are needed to activate the dump-mode. In Python, we just

need to rename the decorator with dump. Similarly, in C++,

arXiv:2210.03724v2 [cs.PF] 7 Nov 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

PMT:PowerMeasurementToolkitStefanoCorda1,BramVeenboer2,EmmaTolley11EPFL´EcolePolytechniqueF´ed´eraledeLausanne,2ASTRONNetherlandsInstituteforRadioAstronomystefano.corda@ep.ch,veenboer@astron.nl,emma.tolley@ep.chAbstractEfcientuseofenergyisessentialfortoday'ssuper-computingsystems,asenergycostisg...

展开>> 收起<<

PMT Power Measurement Toolkit Stefano Corda1 Bram Veenboer2 Emma Tolley1 1EPFL Ecole Polytechnique F ederale de Lausanne2ASTRON Netherlands Institute for Radio Astronomy.pdf

共3页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

PMT Power Measurement Toolkit Stefano Corda1 Bram Veenboer2 Emma Tolley1 1EPFL Ecole Polytechnique F ederale de Lausanne2ASTRON Netherlands Institute for Radio Astronomy

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: