by activation/deactivation of servers in both closed-
queueing and open-queueing models. In (Wierman
et al., 2009), an optimal speed scaling was proposed
to balance the mean energy consumption and mean
response time under Processor Sharing (PS) schedul-
ing.
All of these work focused on simple single-tier
applications. As we know, modern web applications
are usually using multi-tier architecture (Kamra et al.,
2004; Liu et al., 2006; Liu et al., 2008; Urgaonkar
et al., 2005; Pacifici et al., 2005; Diao et al., 2006;
Liu et al., 2005; Wang et al., 2010). Each tier pro-
vides a certain specific functionality to applications.
A client request will pass through a series of tiers
to attain a complete service. For instance, a typ-
ical e-commerce application consists of three tiers:
web server tier, application server tier, and database
server tier. The multi-tier architecture follows lay-
ered queueing models (Rolia and Sevcik, 1995) and
typically has a cross-tier dependency. The service at a
tier is normally blocked while waiting for the service
from its succeeding tier. Such cross-tier dependency
makes the response time analysis challenging in com-
parison with the single-tier architecture. In (Liu et al.,
2005), an analytical model was proposed for 3-tiered
web service architecture. The concurrency limit was
addressed in (Urgaonkar et al., 2005) for multi-tier
applications. In (Pacifici et al., 2005), an architecture
and underlying model of a performance management
system was presented for multi-tier web applications
on server clusters. In (Diao et al., 2006), a hybrid per-
formance model for differentiated services was pre-
sented for multi-tier applications with cross-tier inter-
action. Among these work, only in (Diao et al., 2006)
the cross-tier dependencywas considered, and the rest
applied a tandem-queue-likestructure and ignored the
cross-tier dependency. In (Wang et al., 2010), an over-
simplified M/M/1 model is used to perform the queue-
ing analysis in multi-tier architecture.
In this paper, we aim to conduct a comprehen-
sive study on power saving in server farms for multi-
tier applications requiring that a given SLA should
be met. We adopt two techniques as used in server
farms for power-aware design: the DVS technique
with variable speeds and the DPM technique with ac-
tivation/deactivation of servers. There are two spe-
cific questions that we need to address in the system
design in order to achieve this goal: (i) How many
servers should be activated at each tier? (ii) What is
the best processor speed (corresponding to the volt-
age/frequency to be used) for each server? For a
single-tier architecture with homogeneous servers, it
is shown in (Gandhi et al., 2009) that the optimal
strategy is to set all servers with the same speed for
all activated servers. However, this does not hold in
the multi-tier architecture due to the cross-tier depen-
dency. We present an efficient power-saving design
strategy called
PowerTier
and study how to choose
the number of activated servers at each tier and the
processor speed for each server to minimize the over-
all power consumption in server farms while meeting
a given mean response time guarantee for multi-tier
applications. We consider both open-queueing and
closed-queueing models for applications.
The rest of this paper is organized as follows: Sec-
tion 2 shows the system model. Section 3 presents a
detailed power consumption and response time anal-
ysis, which is the basis for our power-saving design.
Our power-saving design scheme
PowerTier
is de-
scribed in Section 4. Section 5 presents detailed per-
formance evaluation of
PowerTier
over a various of
platforms. We conclude the paper in Section 6.
2 SYSTEM MODEL
In this section, we define the system model, including
the power consumption model for servers, the multi-
tier architecture, and the client application model.
2.1 Power Consumption Model
We assume that all servers are equipped with the
DVS and DPM techniques for the power manage-
ment. When the server is deactivated by the DPM
technique, its power consumption is negligible. So,
here we focus on the power consumption when the
server is activated.
With the DVS technique, we can choose a proces-
sor speed for a server (with a corresponding choice of
the supply voltage). We define r as the ratio of the
processor speed of the server to its maximum speed.
The speed ratio r is normally bounded by a lower
bound r
l
. Then we haver
l
≤ r ≤ 1. When the server is
activated, either it is (i) in the idle mode at the lowest
speed ratio r
l
without executing any job; or (ii) in the
running mode executing jobs with a processor speed
ratio r. The power consumption in our study is the
system-level power, including the power consumed
by the processor and all other components within the
server such as memory and I/O devices. The power
consumption depends on the mode that the server is
in (idle or running) and the processor speed in use as
well. In this paper, we adopt the power consumption
model in (Gandhi et al., 2009). A server has the fol-
lowing power modes:
• Idle Power Mode. In the idle mode, the server
consumes the static power P
I
;
SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems
138