01.12.2012 Views

Architecture of Computing Systems (Lecture Notes in Computer ...

Architecture of Computing Systems (Lecture Notes in Computer ...

Architecture of Computing Systems (Lecture Notes in Computer ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

EnergySav<strong>in</strong>g Cluster Roll: Power Sav<strong>in</strong>g System for Clusters 163<br />

idle <strong>in</strong>terconnections <strong>in</strong> the cluster [2]. Micros<strong>of</strong>t <strong>in</strong>spects the problem from a different<br />

viewpo<strong>in</strong>t and one solution proposed is to share highly efficient power supplies<br />

among several nodes <strong>of</strong> the system, achiev<strong>in</strong>g significant energy sav<strong>in</strong>gs [3].<br />

In this context a well-known energy management technique is DVFS (Dynamic<br />

Voltage and Frequency Scal<strong>in</strong>g). DVFS entails reduc<strong>in</strong>g the system energy consumption<br />

by decreas<strong>in</strong>g the CPU supply voltage and the clock frequency (CPU<br />

speed) simultaneously. This technique has a great impact on the development <strong>of</strong><br />

work aimed at reduc<strong>in</strong>g consumption <strong>in</strong> this research context [4,5,6]. The authors<br />

<strong>in</strong> [7] present an energy-aware method <strong>in</strong> order to partition the workload and reduce<br />

energy consumption <strong>in</strong> multiprocessor systems with support for DVS. Freeh<br />

et al. [8] analyze the energy-time trade<strong>of</strong>f <strong>of</strong> a wide range <strong>of</strong> applications us<strong>in</strong>g highperformance<br />

cluster nodes that have several power-performance states to lower<strong>in</strong>g<br />

energy and power, so that the energy-time trade<strong>of</strong>f can be dynamically adjusted.<br />

In [9], the authors use economic criteria and energy to dispatch jobs to a small set<br />

<strong>of</strong> active servers, while other servers are transitioned to a low energy state.<br />

Alternative strategies to limit power consumption and required cool<strong>in</strong>g <strong>of</strong><br />

HPC clusters are based on switch<strong>in</strong>g on and shutt<strong>in</strong>g down the nodes, accord<strong>in</strong>g<br />

to the needs <strong>of</strong> the users’ applications. An algorithm that makes load balanc<strong>in</strong>g<br />

and unbalanc<strong>in</strong>g decisions by consider<strong>in</strong>g both the total load imposed on<br />

the cluster and the power and performance implications <strong>of</strong> turn<strong>in</strong>g nodes <strong>of</strong>f<br />

is described <strong>in</strong> [10]. Several policies to comb<strong>in</strong>e dynamic voltage scal<strong>in</strong>g and<br />

turn<strong>in</strong>g on or <strong>of</strong>f nodes to reduce the aggregate power consumption <strong>of</strong> a server<br />

cluster dur<strong>in</strong>g periods <strong>of</strong> reduced workload are presented <strong>in</strong> [11]. Rock-solid [12]<br />

and PowerSav<strong>in</strong>g [13] are prototype examples <strong>of</strong> this strategy which provide<br />

little functionality or are still under development. This paper presents a new<br />

application which follows the same strategy, but with a much more complete<br />

functionality. In particular, our tool allows the def<strong>in</strong>ition <strong>of</strong> different conditions<br />

to activate and deactivate and nodes for a full adaption to the requirements <strong>of</strong><br />

the system adm<strong>in</strong>istrator and/or the end user. The tool has been designed and<br />

implemented as a module (Roll) for Rocks R○ [14] and employs the Sun R○ Grid<br />

Eng<strong>in</strong>e (SGE) [15]. A simulation <strong>of</strong> the module under real conditions shows that<br />

its use comb<strong>in</strong>ed with a reasonable policy deliver considerable energy sav<strong>in</strong>gs<br />

compared with a conventional cluster <strong>in</strong> which all nodes are permanently active.<br />

The article is organized as follows: Section 2 presents the hardware and s<strong>of</strong>tware<br />

tools used <strong>in</strong> the development <strong>of</strong> the energy sav<strong>in</strong>g module, and Section 3<br />

reviews its implementation. Section 4 evaluates the performance <strong>of</strong> the energy<br />

sav<strong>in</strong>g module, <strong>in</strong> terms <strong>of</strong> impact on the power consumption and execution time<br />

<strong>of</strong> the applications. F<strong>in</strong>ally, Section 5 describes some future l<strong>in</strong>es <strong>of</strong> work and<br />

summarizes the conclusions.<br />

2 Basic Hardware and S<strong>of</strong>tware Tools<br />

The target hardware platform used <strong>in</strong> this work is an HPC cluster that employs<br />

the NPACI Rocks R○ L<strong>in</strong>ux distribution as operat<strong>in</strong>g system, equipped with a<br />

front-end node that is responsible <strong>of</strong> the queue system and the new energy sav<strong>in</strong>g<br />

module (Roll).

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!