% Razones para alta disponibilidad

MDCSim: A Multi-tier Data Center Simulation
Platform
Seung-Hwan Lim, Bikash Sharma, Gunwoo Nam, Eun Kyoung Kim, and Chita R. Das
Department of Computer Science and Engineering, The Pennsylvania State University
University Park, PA 16802, USA
Technical Report
CSE 09-007
{seulim, bus145, gnam, ekkim, das}@cse.psu.edu

Abstract—Performance and power issues are becoming in- contribute to a considerable portion of the overall delay for
creasingly important in the design of large cluster based multi- web-based services and this delay is likely to increase with
tier data centers for supporting a multitude of services. Design more dynamic web contents. Poor response time has signifi-
and analysis of such large/complex distributed system often
suffers from the lack of availability of an adequate physical cant financial implications for many commercial applications.
infrastructure and the cost constraints especially in the academic User dissatisfaction due to longer response times is a major
community. With this motivation, this paper presents a compre- concern as reported by a survey on the problems with WWW
hensive, flexible and scalable simulation platform for in-depth [2]. Further, the power consumption of data centers is also
analysis of multi-tier data centers. Designed as a pluggable three- drawing much attention, leading to the concept of ”Green Data
level architecture, our simulator captures all the important design
specifics of the underlying communication paradigm, kernel Centers”. As reported in [3], data centers in the U.S. consumed
level scheduling artifacts, and the application level interactions 61 billion kilowatt-hour of electricity in 2006 at the cost of
among the tiers of a three-tier data center. The flexibility of $4.5 billion. This constitutes 1.5% of the total U.S. energy
the simulator is attributed to its ability in experimenting with consumption for that year. It is estimated that data centers’
different design alternatives in the three layers, and in analyzing power consumption will increase by 4% to 8% annually and
both the performance and power consumption with realistic
workloads. The scalability of the simulator is demonstrated with is expected to reach 100 billion kWh by 2011. Thus, future
analyses of different data center configurations. In addition, we data center’s design must focus on three critical parameters:
have designed a prototype three-tier data center on an IBA high performance, power/thermal-efficiency and reliability.
connected Linux cluster to validate the simulator. Using RUBiS In view of this, recent researches focus on the design
benchmark workload, it is shown that the simulator is quite of multi-tier cluster-based data centers, exploitation of high
accurate in estimating the throughput, response time, and power
consumption parameters. We then demonstrate the applicability speed I/O interconnects and power management to meet the
of the simulator in conducting three different types of studies. current demands of fast growing large data centers [4], [5],
First, we conduct a comparative analysis of the Infiniband [6], [7]. Design of distributed web servers is primarily tar-
Architecture (IBA) and 10 Gigabit Ethernet (10GigE) under geted at improving the throughput and reducing the response
different traffic conditions and with varying size clusters for time [8], [9]. Also, with the increasing use of dynamic web
understanding their relative merits in designing cluster-based
servers. Second, measurement and characterization of power contents, a multi-tier architecture provides a clean abstraction
consumption across the servers of a three-tier data center is done. of different functionalities [10], [11]. Further, I/O interconnect
Third, we perform a configuration analysis of the Web server technologies have a significant impact in supporting efficient
(WS), Application Server (AS), and Database Server (DB) for inter/intra cluster communication in a multi-tier data center.
performance optimization. We believe that such a comprehensive In this context, researchers have attempted to examine the ad-
simulation infrastructure is critical for providing guidelines in
designing efficient and cost-effective multi-tier data centers. vantage of employing a user-level communication mechanism
such as the InfiniBand Architecture (IBA) [12] in contrast to
I. I NTRODUCTION the traditional TCP/IP protocol in designing the underlying
Design of high performance, power-efficient and dependable communication infrastructure to reduce the network overhead
cluster based data centers has become an important issue, more by minimizing the kernel involvement in message transfer [6].
so with the increasing use of data centers in almost every sec- Similarly, power management techniques and understanding
tor of our society: academic institutions, government agencies power consumption in data centers have been attempted by a
and a myriad of business enterprises. Commercial companies few researchers [7], [13].
such as Google, Amazon, Akamai, AOL and Microsoft use There are three approaches for the analysis of a system,
thousands to millions of servers in a data center environment namely measurement, modeling and simulation. Measurement
to handle high volume of traffic for providing customized do provide an accurate methodology to analyze systems.
(24x7) service. A recent report shows that financial firms spend However, many times, measuring large scale systems is not
around 1.8 billion dollars annually on data centers for their feasible, given the cost and time constraints. Modeling offers
businesses [1]. However, it has been observed that data centers a simple representation of the construction and working of the

system of interest. However, the cost and overhead involved ing load. Next, power measurement in multi-tier data center
in the reconfiguration and experimenting with a model is was done. Finally, a methodology for the determination of op-
substantially high [14]. Simulation fills in the above defi- timal cluster configurations in terms of performance (latency)
ciencies by providing a flexible, operative working model through simulation experiments was discussed . Our experi-
of a system. A simulator can outline a more detail view mental results for the comparative analysis between 10GigE
of a system than modeling [14]. Simulation of a system is and IBA show that 10GigE performs better than IBA under low
also an economically better option as compared to the real traffic but as the number of clients increases, IBA outperforms
implementation of a system. Moreover, simulation saves the 10GigE both with and without TCP Offload-Engine (TOE).
time spent on configuring test environments. From the power consumption comparison among IBA, 10GigE
Analysis of multi-tier data center has been done with respect with/without TOE connected data centers with different cluster
to modeling and simulation in the past [6], [15]. However, the configurations and with increasing number of clients, it was
scope and scale of analysis is restricted to individual tier or observed that the power consumption increases with the in-
with small size clusters. Further, in most of the studies on crease in the traffic across the cluster. Also, the power usage
data centers, the size of system analyzed is limited because of was found to be varying for different cluster configurations
practical difficulties encountered in testing large scale systems even for the same total number of nodes. Further, simulations
[16], [17]. Analysis done on small scale cannot always guaran- with different cluster configurations established the fact that
tee similar behavior or trend when the scalability is enhanced. increasing the number of nodes in a cluster does not always
Thus, there is a need to analyze system on a larger scale guarantee better performance. There are specific configurations
keeping in view the existing trend in data centers. Moreover, (optimal distribution of nodes across each tier) that can achieve
in spite of the growing popularity of multi-tier data centers, to the desired level of performance.
the best of our knowledge, simulating a multi-tier data center The rest of the paper is organized as follows. In Section
including the functionalities of all the tiers together has not II, we discuss the related works. The design of the simulation
been attempted. Keeping in view the importance of growing platform is described in Section III. Prototype design, system
data centers and the above limitations in the analysis of such configuration and validation of simulator are discussed in
concurrent and distributed systems, we propose a simulation Section IV. Our results and discussions are presented in
platform for the design and analysis of large scale multi-tier Section V, followed by the concluding remarks and future
data centers. directions in Section VI.
Our simulation platform discussed above is comprehensive,
flexible and scalable. The comprehensiveness comes from the II. R ELATED W ORK
fact that each salient feature of a multi-tier data center is A few researchers have addressed the modeling of a multi-
implemented in detail and with much clarity. It is flexible as tier data center. Modeling or simulation of each individual
we can manipulate different features of the data center. For tier separately in a multi-tier data center has been studied in
example, the number of tiers can be changed, the scheduling [15], [19], [20], [21]. Modeling of a multi-tier data center
algorithms can be customized, the type of interconnection used has been done in few works [4], [5]. [4] models a multi-
can be altered and so also the communication mechanisms. tier data center as a network of queues and Stewart et al.
Further, the layered architecture provides the ease in doing [5] models each service tier as a fine-grained component to
modification to any individual layer without affecting other capture practical implementations of each server. However,
layers. Simulation experiments are possible with large number design of a simulation platform for the analysis of multi-tier
of nodes across the cluster, making it scalable and apt to data centers taking into account the effects of all tiers together
existing data centers’ trends. has gone unaddressed in previous works.
For the analysis of data center, we first designed and Balaji et al. [6] have shown the benefits of IBA in multi-
implemented a prototype three-tier data center on an IBA tiered web services. However, the system analyzed in [6] is
and 1GigE connected Linux cluster. The major components restricted in size and system capability in the sense that it does
of this prototype included WS tier, AS tier and DB tier. The not include the Java 2 Enterprise Edition (J2EE) [22] based
web traffic was generated by RUBiS [18] workload, which application servers, which have become the de facto standard
gets distributed using a simple round-robin web switch. Since for designing data centers. Our prototype data center includes
the prototype was limited in terms of the number of clus- J2EE specifications. Further, they conducted the comparison
ter nodes, system configurations, and driving workloads, we analysis with low traffic and small cluster sizes. In this paper,
designed the proposed simulation framework. The simulator as part of an application of our simulator, we perform the
was validated with RUBiS benchmark measurements on the evaluations and comparisons of IBA and 10GigE with/without
above mentioned prototype data center. The average deviation TOE on a much larger scale and under high load conditions
of simulation results from real measurements was found to be typical to the current data centers’ trends.
around 10%. Modeling of power consumption in computer systems has
Using our simulator, we implemented three important ap- been done at various granularity levels, starting from large data
plications. First, comparative analysis of IBA and 10GigE was center to individual components of a system either through
performed under different cluster configurations and with vary- real measurements, estimations from performance counters or

2

End Users Front-end Tier/ Mid Tier/ Back-end Tier/
(B2C) Web Server Application Server Database Server

DtaGnea
a e rl

Web Server
Application Database
Server Server

Storage
Server
Laptop DtaGnea
a e rl

Com m ro
P t SD
Bay et wor ks
N B aySt ac k 450- 12T S wich
t

Upln M e
ik odul 13 14 15 16

1 2 3 4 5 6 7 8 9 10 11 12
1 2 3 4 5 6 7 8 9 10 1112
100
10
F x
D
Actyt
i vi

Webswitch
Computer

Partners
(B2B)

DtaGnea
a e rl

Fig. 1. Functional Overview of the Simulation Platform Fig. 2. A Generic Three-Tier Data Center

using a combination of both [23], [24]. Power management abstraction provides the flexibility and scalability in ana-
for communication protocols has been studied in previous lyzing various design details as described next. Intra-cluster
researches [16], [25], [26]. [16] compared the power based communication is modeled at the communication layer. For
benefits of using RDMA adapters, such as IBA and 10 GigE intra-cluster communication, the simulator supports IBA and
with TOE against TCP/IP Ethernet. However, they conducted Ethernet over TCP/IP as the underlying interconnect technol-
their experiments by using only two servers and with just ogy. The IBA communication modules follow the semantics
micro-benchmarks. Further, energy consumption in server of the IBA specification [12] and support major functional-
farms using simulation is studied in [26]. However, the scale ities of the IBA, including communication with Completion
of simulation is restricted (results for only up to 32 nodes are Queue (CQ) and Send/Receive Queues. Other communication
shown). Our simulation results for power measurement across paradigms/protocols can be incorporated into the communi-
the servers of a data center span over much larger nodes cluster cation layer by simply plugging in the appropriate timing
configuration. Besides, there have been significant researches parameters.
on power management methodologies that focus only on At the kernel layer, we modeled the Linux scheduler 2.6.9,
individual tier such as web server tier without taking into
which maintains a run queue for each CPU in the system.
consideration the effects of all the tiers of a multi-tier data Since we assumed all the CPUs in a node as one single
center. For example, [13], [27] analyze only front end power
resource, we have just one run queue for each server node.
and energy management techniques. [7] discusses local and
Each time the scheduler runs, it finds the highest priority
cluster-wide power management techniques using just a web task among communication and application processes in the
server simulator. Simulator proposed in this paper considers
simulator. If there are multiple tasks with the same priority,
the effects of all tiers together and can accurately estimate
those tasks will be scheduled by the round-robin scheme.
power consumption across the servers of a multi-tier data When calculating the time slice, the scheduler punishes CPU
center.
intensive tasks and rewards I/O intensive tasks as Linux 2.6.x.
To the best of our knowledge, no previous works have kernel does.
looked at multi-tier data center simulator, which can analyze
The high level application/user layer captures the important
a cluster based data center with detailed implementation of
characteristics of a three-tier architecture as shown in Figure
each individual tier. Also, power measurement in a large
2. The user level layer consists of the following categories
consolidated environment like a multi-tier data center has not
of processes - web server (WS), application server (AS), and
been analyzed using a comprehensive multi-tier simulator in
database server (DB) processes; auxiliary processes like disk
previous literatures.
helper for supplementing server processes; communication
processes such as Sender/Receiver processes. The front-end
III. D ESIGN OF S IMULATION P LATFORM
web server is the closest to the edge of a data center and
In this section, we discuss the design and implementation is responsible for serving either static requests or forwarding
issues of our simulation platform. The functional building dynamic requests to application server nodes. To serve static
blocks of our multi-tier data center simulator, written in CSIM requests, it can be read from the memory or disks in a
[28], are shown in Figure 1. node. In our case, we use Apache [29] as the front end web
The simulation is configured into three layers (a com- server interface. The middle tier, the application server tier, is
munication layer, a kernel layer and a user-level layer) for responsible for handling the dynamic web contents, and we
modeling the entire stack starting from the communication use the publicly available JOnAS [30] for our research. On
protocols to the application specifics. Such a three layer receiving the dynamic requests, the application server nodes

3

Requests Web Server Application Server
from Communication Layer
Tier Tier
Web Switch Database Server
Apache Tier
Web Tomcat Web JOnAs
Web Server Application Server Database Server Server Container RMI EJB Container

MySQL
Application
NIC NIC Apache Server
Web
Disk Disk Disk
Sender Receiver Sender Receiver Sender Receiver Server
Helper Helper Helper
Tomcat Web JOnAs
Clients mod_jk Container RMI EJB Container
C-JDBC MySQL
User Layer User Layer User Layer Apache
Web Application
Server Server Application
Server

MySQL
Disk Scheduler Disk Scheduler Disk Scheduler Apache Tomcat Web
RMI
JOnAs
Web Container EJB Container
Server
Kernel Layer Kernel Layer Kernel Layer

Fig. 3. Architectural Details of the Three Layers of the Simulator Fig. 4. A Prototype Data Center

process the business logic. Being in middle between the front- that of each node has been implemented in a lucid manner.
end and back-end in a data center, application servers include For example, we implemented database synchronization and
interfaces to both tiers: translating clients’ messages in HTML load distribution for web requests in the application layer.
to SQL and vice versa. The AS nodes’ functionalities are The task scheduler was deployed in the kernel layer and
based on J2EE specifications. To model EJB session beans the communication primitives are captured in the low level
[20] in our simulator, we implemented three main components communication layer.. Its flexibility comes from the fact that
- web container which takes care of the presentation logic, EJB the simulator provides the ease to modify the behavior of any
container that handles business logic and controls interaction entity in a simulator without affecting other parts by simply
with the database. Finally, the database server, based on plugging different parameters, policies, or even implementing
the SQL semantics [31], is dedicated for complex database different features. For instance, we need to only vary the
transactions. It either primarily accesses the cached data in latency timing parameters to analyze the behavior of both
the database or sends a request to disks. IBA and 10GigE network adapters without changing any other
From measurements on our prototype data center (discussed parts of the simulator. Also, we can analyze the behavior of
later) with the RUBiS workloads, we observe that 97% of several cluster configurations for optimizing various objective
the database requests are satisfied from the cache, thus, only functions without having to rearrange or change our simulator.
3% of the requests are directed to the disks. We use these Moreover, although the current version of the simulator is
cache and disk access rates in simulating the database tier. currently used for performance and power analysis, it can
Since we assumed mirroring as data sharing method among the be extended to study thermal behavior and reliability issues.
database servers, we modeled locking operation to ensure data The scalable aspect of the simulator enables it to simulate
coherence. We also assumed that there is one server process large data centers. There can be many possible applications
per node to simplify the implementation of each tier of our of the simulator such as comparative performance analysis
three-tier data center simulator. of high speed interconnects, power and thermal analysis,
determination of optimal cluster configuration to meet a tar-
The architecture of our simulator is shown in Figure 3. Here,
get performance level, fault tolerance and reliability of data
Network Interface Card (NIC) and Disk are devices and others
centers.
are processes that reside inside the CPU in a server node.
We assumed that CPU, Disk and NIC are basic computing IV. P ROTOTYPE C ONFIGURATION AND VALIDATION
resources in a node responsible to process requests. Each In this section, we present the design and implementation
resource is modeled as an M/M/1 queue in the simulator. In of a prototype data center, the system configuration, and the
the simulator, nodes separately process the requests which they simulator validation. As stated in the introduction, we have
receive through the communication layer. After processing the designed a prototype three-tier data centers to validate our
requests, they send the responses back through the communi- simulator.
cation layer and this procedure follows for each node in the
cluster. Since we simulate each node independently, we can A. A Prototype Design
change the configuration of each tier easily, such as varying A three-tier data center prototype was designed and im-
the total number of nodes in the entire cluster and the number plemented as shown in Figure 4 for conducting actual mea-
of nodes for each tier. surements. The major components of this prototype include
The main advantages of our simulation platform is that it clients, WS tier, AS tier and DB tier. A simple round-robin
is comprehensive, flexible and scalable. It is comprehensive web switch emulation was implemented to distribute client
as the critical behavior of an entire data center as well requests to the web server tier nodes. The web server nodes, in

4

which Apache 2.0.54 [29] is installed, communicate with the
application server nodes through the mod jk connector module
v.1.2.5. The maximum number of Apache processes was set to
512 to maximize the performance of WS tier. There are many
application servers based on J2EE specifications which define
the architecture and interfaces for developing Internet server
applications for a multi-tier data center. Among them, JOnAS
4.6.6 [30], an open source application server, is installed on
our prototype data center. TomcatWeb container 5.0.30 is
integrated with JOnAS as the web container. Remote Method
Invocations (RMI) provide intra-cluster communication among
the AS tier servers. Sun’s JVM from JDK 1.5.0 04 for Linux
was used to run JOnAS. EJB session beans version were
employed to achieve the best performance as reported in [20]. Fig. 5. Estimated Latencies of various 10GigE adapters
Communication between AS-tier and DB tier was achieved
through C-JDBC v.2.0.1 [32], which allows EJB container to
access a cluster of databases. For the back-end tier, we used quests according to the RUBiS benchmark, which follows the
multiple machines running MySQL 5.0.15-0 Max [31]. TPC-W compatible specification [33]. For generating traffic
among different tiers (intra-cluster traffic), we used log-normal
Our prototype is implemented on varying number of ded-
distribution as reported in [11] along with the parameters and
icated nodes of a 96-node Linux kernel 2.6.9 cluster. Each
think time interval of clients shown in Table II.
of the these nodes has dual 64-bit AMD Opteron processors,
In order to correctly simulate the characteristics of the AS
4 GBytes of system memory, Ultra320 SCSI disk drives and
and DB tier with dynamic requests, estimating CPU service
a Topspin host adapter which is InfiniBand v.1.1 compatible.
times for AS and DB servers is essential. We estimate the
The adapters are plugged into PCI-X 133Mhz slots. Each node
CPU service time of each tier node. From measurements on
has 1Gig-Ethernet adapter of Tigon 3 connected to 64 bit, PCI-
the prototype, CPU utilization and throughput of each node is
X 66Mhz slots. The default communication between all the
obtained. We then use the Utilization Law [4] to obtain the
server nodes is through 1GigE connection. In order to make
service time as shown in equation (1).
IBA as our connection backbone instead of Ethernet, we ran
all of the server applications with the SDP library (provided by ρ
T = (1)
Topspin) in our experiment. We used RUBiS [18] benchmark τ
workload to collect actual measurement data for conducting where, T is the service time, ρ is the utilization of the busiest
our experiments since it was difficult to get real workloads resource (CPU) and τ is the throughput of each node.
because of the proprietary nature of Internet applications. Next for simulating 10GigE network adapter, we need the
latency timing equation. Through linear regression analysis
B. System Configuration and Workload (Figure 5), we estimated the latency equations for two kinds
Table I summarizes the system and network parameter of 10GigE implementations (with/without TOE) in terms of
values used in our experiments. These values were obtained message size by obtaining latency from throughput [17] as
from measurements on our prototype data center using micro per the following equation:
benchmarks and hardware specifications. Clients generate re- Data Size
Latency = (2)
T hroughput
TABLE I As reported in [34], the hardware characteristics of 10GigE
T IMING PARAMETERS
is almost the same as 1GigE. Further, the network latency for
Parameter from Measurement Value (ms)
Context switch 0.050 TABLE II
Interrupt overhead 0.01 T RAFFIC C HARACTERISTICS
TCP connection 0.0815
Process creation overhead 0.352
Disk read time 7.89×2+(0.0168 × Size)+ Parameter Value
(Size/368 − 1) × 7.89 WS → AS α = 0.1602, µ = 5.5974
Network latency (IBA) 0.003 + 0.000027 ×Size File Size AS → DB α = 0.4034, µ = 4.7053
Network latency (1GigE) 0.027036 + 0.012 ×Size DB → AS α = 0.5126, µ = 22.134
Parameter from Estimation Value (ms, KB) AS → WS α = 0.5436, µ = 56.438
Network latency (TOE) 0.0005409 + 0.0010891 WS → client α = 0.5458, µ = 50.6502
×Size Dynamic Req/Total Req 0.672
Network latency (10GigE) 0.0009898 + 0.0017586 Avg. DB Req per Dynamic Req 2.36
×Size Think time interval Exponential with a mean of 7s

5

120 450 450
WS WS REAL WS REAL
AS 400 AS REAL 400 AS REAL
100 DB DB REAL DB REAL
WS sim 350 WS SIM 350 WS SIM
AS sim AS SIM AS SIM
80 DB sim 300 DB SIM 300 DB SIM
Latency(ms)

Latency(ms)

Latency(ms)
250 250
60
200 200
40 150 150
100 100
20
50 50
0 0 0
800 1600 2400 3200 4000 4800 5600 6400 4000 4800 5600 6400
number of clients number of clients number of clients
(a) IBA, 2WSs, 6ASs, 3DBs (b) IBA, 4WSs, 16ASs, 6DBs (c) 1G, 4WSs, 16ASs, 6DBs

Fig. 6. Latency comparison between measurement and simulation

(a) 800 Clients (b) 1600 Clients (c) 2400 Clients (d) 3200 Clients

Fig. 7. Service time CDFs from Measurement and Simulation

both 1GigE and 10GigE includes similar hardware overheads power consumption. According to their power model, for a
caused by switches and network interface card. Thus, we can ﬁxed operating frequency, both the power consumption of
safely assume that our simulator can be extended to capture the server and the application throughput are approximately
the behavior of 10GigE. linear functions of the server utilization. The peak power
consumption of the data center was obtained by simple linear
Our simulator is capable of measuring power consumption regression between power and server utilization.
across the servers of a multi-tier data center. To incorporate
power measurements in our simulator, we measured the server C. Simulator Validation
utilization using some modiﬁcations to the heuristics given in
[13]. The server utilization was measured as We validated our simulator with 1GigE and IBA based
prototype data centers with different number of nodes for each
Nstatic Astatic + Ndynamic Adynamic tier. The average latencies of each requests from simulation
Userver = , (3)
Monitor Period × # WS nodes results were compared with those from real measurements
where Nstatic and Ndynamic are the number of static and as shown in Figure 6. Here, we selected the number of
dynamic requests respectively; Astatic and Adynamic are the nodes for each tier in the prototype data center when the
average execution time of static and dynamic requests re- latency of requests was minimized. In Figure 6(a) and 6(b),
spectively; Monitor Period is the time interval over which the underlying network architecture was IBA. Figure 6(c)
we measure server utilization and power. We divided the shows the latencies over 1GigE from actual measurements
numerator of the equation for server utilization from [13] by and simulation. In the above results, the latency of each tier
the total number of web server nodes in the cluster, since from simulation closely matched with the real measurements,
our simulator processes the client requests in parallel at the which implies our simulator’s underlying architecture captures
front end tier [35]. We then computed the server utilization critical paths in each tier. The average latencies from the
using the above equation. Power consumption is expressed simulator match closely with the average deviation of 10%.
as percentage of the peak power across the data center. It should be noted that the above deviation is independent of
We used the power model of Wang et.al [36] to estimate either the number of nodes or the number of clients used in

6

140 90
WS(IBA) AS(IBA)
WS(10G TOE) 80 AS(10G TOE)
120 WS(10G) AS(10G)
70
100
60

Latency(ms)

Latency(ms)
80 50

60 40
30
40
20
20
10
0 0
800 1600 2400 3200 800 1600 2400 3200
number of clients number of clients
(a) Latency in WS tier (b) Latency in AS tier
55 6500
DB(IBA) IBA
50 DB(10G TOE) 6000 10gTOE
Fig. 8. Power comparison between measurement and simulation 45 DB(10G) 5500
10G
40

Throughput(r/s)
5000

Latency(ms)
35
4500
30
4000
25
3500
our simulations. 20
3000
Figure 7 shows the CDF plots of service times of each 15
10 2500
tier for the simulation and real measurements. According to 5 2000
[11], service time of an actual machine for RUBiS workload 0 1500
800 1600 2400 3200
800 1600 2400 3200
follows the Pareto distribution. We notice that the service number of clients number of clients

time of each tier from our simulator closely matched with (c) Latency in DB tier (d) Throughput
the real measurements. It implies that the modeled behavior
for each tier in the simulator is very close to that of actual Fig. 9. Comparison between IBA and 10GigE
servers. From both the latency and CDF comparison, we
confirm that our simulator well captured the critical behaviors
to process requests and showed similar traffic characteristics for three configurations (IBA, 10GigE with TOE and 10GigE
to real environments. without TOE). We used 2 WS nodes, 6 AS nodes, and 3DB
We validated the estimated power consumption using our nodes for this simulation. Notice that the latency of WS
simulator with real power measurements. Figure 8 shows the and AS includes the communication overhead between WS
variation of power consumption with the number of clients and AS and between AS and DB, respectively. We observe
corresponding to real and estimated power measurements. We that 10GigE, regardless of TOE, has a better or very similar
used a power measurement device WT210 [37] to measure performance over IBA in latency measurement, provided the
power usage of nine (the power strip used could only accom- number of clients is under 3200. However, with the increase
modate nine connections at a time) randomly chosen nodes in traffic beyond 3200 clients, the latency of IBA comes out
from our prototype data center. These nine nodes were con- to be 14% smaller than 10GigE with TOE and 26.7% lesser
figured as 2 web servers, 6 application servers and 1 database than 10GigE without TOE.
server. The power consumption from our simulator closely Under high traffic, the communication cost which is deter-
matched with that of real measurements with an average mined by the mean number of jobs in communication layer is
deviation of around 3%. This result indicates that our simulator a dominant factor of the response time. Based on the queuing
is capable of accurately estimating power consumption across theory, the mean number of jobs in a queue, J can be obtained
the servers of a multi-tier data center. by
J = λ × T, (4)
V. P ERFORMANCE E VALUATION
where λ is the inter-arrival rate of requests and T is the mean
In this section, we demonstrate three important applications service time of a single request. Accordingly, given the same
using our simulation platform. First, comparison between two inter-arrival rate of requests, the latency under high traffic is
popular interconnection networks is done in Section V-A. mainly affected by the mean service time of a single request,
Second, power estimation using simulator is elaborated in which is the average latency of a single file for the workload
Section V-B. Finally, performance prediction according to in our simulation. Even though the file sizes vary from 1.5KB
various system configuration is described in Section V-C. to 16KB for the RUBiS workload, most of them are clustered
around 2.5KB [11]. As shown in Figure 10, the latencies of
A. Comparison between IBA and 10GigE 10GigE with TOE and without TOE are greater than that
We conducted comparative performance analysis of 10GigE of IBA with 2.5KB file size. Hence, IBA has the smallest
and IBA for three-tier data centers in terms of latency and response time among all the three network interconnections
throughput. The graphs in Figure 9 depict the average latencies with high communication workload.
and throughput of requests in terms of the number of clients We observe that the latency for the DB tier in case of IBA is

7

0.01
IBA are under utilized, and so do not exhibit large power usage.
0.009 10G TOE
10G However, with the increase in the traffic, server utilization is
0.008 enhanced to make the servers derive more power supply. As we
0.007 observe, the power consumption varies with different cluster
Latency(ms)

0.006 configurations even for the same total number of nodes. This
0.005 implies the dependency of some specific tier in deciding the
0.004 power dominance. These results provide one with the insight
0.003 into determining the particular configuration, which has a
0.002 comparatively less power consumption across the cluster (this
0.001 can be especially important for the designer of a multi-tier
0 data center, where he has to determine the distribution of the
0 1 2 3 4 5
Message Size(KB) number of nodes across each tier based on power efficiency).
Fig. 10. Latency vs message size for different network architectures
C. Server Configuration
To demonstrate the scalability and flexibility of our simu-
lator, we conducted performance analysis of three-tier data
much higher than that of 10GigE with and without TOE under center with increased number of nodes and with different
high traffic (3200 clients as shown in Figure 9(a), 9(b), and interconnection networks. Figure 12 shows the variation of
9(c)). This can be explained due to the fact that the throughput latency with different number of clients for 32, 64, and 128
of IBA in case of 3200 clients is higher than that of 10GigE , nodes respectively. The combination of the number of nodes
which increases the number of requests generated for the DB in each tier was selected randomly. We observe that the
tier and thus degrades the response time of the DB tier. This performance of the data center in terms of latency improves
is due to the fact that the blocking operation, which ensures (with the decrease in latency) with an increase in the total
consistency among databases, can cause system overheads in number of nodes in the system. This is due to the efficient
the database servers when the number of concurrent writing distribution of system load across multiple nodes of the tiers.
operation increases. Also, we notice that the WS part of bar However, we observe that this trend is violated for the case
graphs is negligible in some results. This is because the service with 128 nodes as shown in Figure 12(c). This is due to the
time of WS which includes communication overheads between large overhead caused by the DB tier locking operation, which
AS and WS is less than 10% of the overall latency in most causes the total response time to be dominated by the DB
experiments. portion of latency. These results imply that the performance
might not be enhanced in proportion to the number of total
B. Power Measurement nodes. The underlying mechanisms in each server may cause
As mentioned before, our simulator can estimate power unexpected results if the scalability increases. The above
consumption across the servers of a data center. The flexibility results demonstrate that even without real measurement, we
offered by our simulator allows one to measure the power can suggest, for given system characteristics, the right choice
consumption for any selected number of clients and with any among possible combinations of server configuration by the
chosen cluster configuration. From Figure 8, we notice that simulation.
the power consumption across the cluster nodes increases with Next we examine latency variations for a fixed number of
the increase in the traffic from 800 to 3200 clients for both total nodes in a data center for different cluster configurations.
the prototype data center as well as the simulation platform. Figure 13 depicts the results. We observe that performance of
Figure 11 shows the estimated power consumption using our a data center in terms of latency varies based on the different
simulator across 64 nodes clustered data center connected with configuration of nodes selected for each tier and also with
IBA, 10GigE with TOE and 10GigE without TOE respectively. the type of interconnection used. Also, notice that a better
Here, S1, S2, and S3 represent three different cluster config- configuration of the system for minimizing the response time
urations with varying number of nodes in the WS tier, AS can be achieved by selecting more number of AS nodes
tier, and DB tier respectively for the same total number of than WS and DB nodes. This is because AS nodes have
nodes, 64 nodes. In S1, 8 web servers, 32 application servers, CPU intensive jobs and interfaces with two other ends, thus
and 24 database servers were used. In S2, 8 web servers, 40 increasing them efficiently distributes the load across the
application servers, and 16 database servers were used. S3 cluster. Thus, our simulator can be used to change various
consisted of 16 web servers, 32 application servers, and 16 configurations to determine the desired distribution of nodes
database servers. We expressed power as percentage of peak across the cluster for optimizing performance.
power in Figure 11. As we observe from Figure 11, power
consumed when the number of clients is less, comes out to VI. C ONCLUSIONS
be more or less same for different cluster configurations in In this paper, we first designed and implemented a prototype
all the three cases (IBA, 10GigE with TOE, 10GigE without three-tier IBA and 1GigE connected data center. Since the
TOE). This is because when the traffic is less, the servers prototype was limited in terms of the number of cluster nodes,

8

(a) IBA, 64 nodes (b) 10G TOE, 64 nodes (c) 10G, 64 nodes

Fig. 11. Estimated Power Consumption

140 90 1000
WS(IBA) WS(IBA) WS(IBA)
WS(10gTOE) 80 WS(10gTOE) 900 WS(10gTOE)
120 WS(10G) WS(10G) WS(10G)
AS(IBA) 70 AS(IBA) 800 AS(IBA)
100 AS(10gTOE) AS(10gTOE) 700 AS(10gTOE)
60
Latency(ms)

AS(10G)
Latency(ms)

AS(10G) AS(10G)

Latency(ms)
80 DB(IBA) DB(IBA) 600 DB(IBA)
DB(10gTOE) 50 DB(10gTOE) DB(10gTOE)
DB(10G) DB(10G) 500 DB(10G)
60 40
400
30
40 300
20
200
20
10 100
0 0 0
3200 4000 4800 5600 3200 4000 4800 5600 6400 3200 4000 4800 5600 6400
(a) 32nodes(4,20,8) (b) 64nodes(8,40,16) (c) 128nodes(16,80,32)

Fig. 12. Latency with varying number of nodes (WS, AS, DB)

100 90 160
WS(IBA) WS(IBA) WS(IBA)
90 WS(10gTOE) 80 WS(10gTOE) 140 WS(10gTOE)
WS(10G) WS(10G) WS(10G)
80 AS(IBA) 70 AS(IBA) 120 AS(IBA)
70 AS(10gTOE) AS(10gTOE) AS(10gTOE)
60 AS(10G)
Latency(ms)

AS(10G)
Latency(ms)

Latency(ms)

60 DB(IBA) 100 AS(10G)
DB(IBA) DB(IBA)
DB(10gTOE) 50 DB(10gTOE) DB(10gTOE)
50 DB(10G) 80 DB(10G)
40 DB(10G)
40 60
30
30
20 40
20
10 10 20

0 0 0
3200 4000 4800 5600 3200 4000 4800 5600 6400 3200 4000 4800 5600
(a) 64nodes(8,32,24) (b) 64nodes(8,40,16) (c) 64nodes(16,32,16)

Fig. 13. Latency with varying number of nodes for each tier (WS, AS, DB)

system configurations, and driving workloads, we presented first study detailed a comparative analysis of the performance
a comprehensive, flexible and scalable simulation platform of IBA and 10GigE under varying cluster sizes, network load
for the performance and power analysis of multi-tier data and with different tier configurations. The results from this
centers. The simulator was validated with RUBiS benchmark experiment indicated that 10GigE with TOE performs better
measurements on the above mentioned prototype data center. than IBA under light traffic. However, as the load increases,
Using three application studies, we demonstrated how designer IBA performs better than 10GigE both with and without
of a multi-tier data center might use the proposed simulation TOE. Next, we introduced a methodology to perform power
framework to conduct performance and power analysis. The measurement in a multi-tier data center using our simulator.

9

Finally, using simulation platform, confguration analysis for [15] T. F. Abdelzaher, K. G. Shin, and N. Bhatti, “Performance guarantees
performance optimization in terms of reduced network latency for web server end-systems: A control-theoretical approach,” in IEEE
Transactions on Parallel and Distributed Systems, vol. 13, no. 1, Jan
was demonstrated. 2002.
As a part of our future work, we intend to test our [16] “Evaluating high performance communication: a power pespective,”
simulation platform with diverse set of workloads besides in 23rd International Conference on Supercomputing, to be appeared,
2009.
RUBiS. Further, we plan to extend our simulation platform [17] W. B. Feng, C. P. Baron, L. Bhuyan, and D. Panda, “Performance
for conducting thermal analysis, reliability measurement and characterization of a 10-Gigabit Ethernet TOE,” in High Performance
servers’ provisioning (i.e, determination of the optimal number Interconnects, 2005. Proceedings. 13th Symposium on, 2005, pp. 58 –
63.
of servers in each tier based on performance and power) in a [18] C. Amza, E. Cecchet, A. Chanda, A. L. Cox, S. Elnikety, R. Gil, J. Mar-
multi-tier data center. Also, we are working towards making guerite, K. Rajamani, and W. Zwaenepoel, “bottleneck characterization
our simulator a publicly available tool for use in research of dynamic web site benchmakrs,” in Third IBM CAS Conference, 2002,
2002.
community in near future. [19] D. Villela, P. Pradhan, and D. Rubenstein, “Provisioning servers in the
application tier for e-commerce systems,” ACM Trans. Interet Technol.,
vol. 7, no. 1, p. 7, 2007.
R EFERENCES [20] E. Cecchet, J. Marguerite, and W. Zwaenepoel, “Performance and
scalability of EJB applications,” SIGPLAN Not., vol. 37, no. 11, pp.
[1] FinanceTech, “FinanceTechnology Network,” 2009, available from 246–261, 2002.
http://www.financetech.com. [21] J. Chen, G. Soundararajan, and C.Amza, “Autonomic provisioning
[2] Graphic, Visualization, & Usability Center, “GVU’s 10th of backend databases in dynamic content web servers,” Autonomic
WWW User Survey,” December 1998, available from Computing, 2006. ICAC ’06. IEEE International Conference on, pp.
http://www.gvu.gatech.edu/user_surveys/ 231–242, June 2006.
survey199810/. [22] Sun Microsystems, Inc., “J2EE: Java 2 Platform Enterprise Edition,”
[3] M. Marwah, R. Sharma, R. Shih, C. Patel, V. Bhatia, M. Mekanapurath, 2003, available from http://java.sun.com/j2ee/.
R. Velumani, and S. Velayudhan, “Data analysis, visualization and [23] S. Gurumurthi, A. Sivasubramaniam, M. J. Irwin, N. Vijaykrishnan,
knowledge discovery in sustainable data centers,” in COMPUTE ’09: M. Kandemir, T. Li, and L. K. John, “Using complete machine
Proceedings of the 2nd Bangalore Annual Compute Conference on 2nd simulation for software power estimation: The softwatt approach,” in
Bangalore Annual Compute Conference. New York, NY, USA: ACM, HPCA ’02: Proceedings of the 8th International Symposium on High-
2009, pp. 1–8. Performance Computer Architecture. Washington, DC, USA: IEEE
[4] B. Urgaonkar, G. Pacifici, P. Shenoy, M. Spreitzer, and A. Tantawi, “An Computer Society, 2002, p. 141.
analytical model for multi-tier internet services and its applications,” [24] W. Wu, L. Jin, J. Yang, P. Liu, and S. X.-D. Tan, “Efficient power mod-
in SIGMETRICS ’05: Proceedings of the 2005 ACM SIGMETRICS eling and software thermal sensing for runtime temperature monitoring,”
international conference on Measurement and modeling of computer ACM Trans. Des. Autom. Electron. Syst., vol. 12, no. 3, pp. 1–29, 2007.
systems. New York, NY, USA: ACM Press, 2005, pp. 291–302. [25] C.-H. Hsu and W.-C. Feng, “A power-aware run-time system for high-
[5] C. Stewart and K. Shen, “performance modeling and system manage- performance computing,” in SC ’05: Proceedings of the 2005 ACM/IEEE
ment for multi-component online service,” in NSDI ’05: 2nd Symposium conference on Supercomputing. Washington, DC, USA: IEEE Computer
on Networked Systems Design & Implementation, 2005. Society, 2005, p. 1.
[6] P. Balaji, S. Narravula, K. Vaidyanathan, S. Krishnamoorthy, J. Wu, and [26] V. W. Freeh, F. Pan, N. Kappiah, D. K. Lowenthal, and R. Springer, “Ex-
D. K. Panda, “Sockets Direct Protocol over InfiniBand in clusters: is ploring the energy-time tradeoff in mpi programs on a power-scalable
it beneficial?” in ISPASS ’04: Proceedings of the 2004 IEEE Interna- cluster,” in IPDPS ’05: Proceedings of the 19th IEEE International
tional Symposium on Performance Analysis of Systems and Software. Parallel and Distributed Processing Symposium (IPDPS’05) - Papers.
Washington, DC, USA: IEEE Computer Society, 2004, pp. 28–35. Washington, DC, USA: IEEE Computer Society, 2005, p. 4.1.
[7] R. Bianchini and R. Rajamony, “Power and energy management for [27] M. Elnozahy, M. Kistler, and R. Rajamony, “Energy conservation poli-
server systems,” Computer, vol. 37, no. 11, pp. 68–76, Nov. 2004. cies for web servers,” in USITS’03: Proceedings of the 4th conference on
[8] M. Aron, D. Sanders, P. Druschel, and W. Zwaenepoel, “Scalable USENIX Symposium on Internet Technologies and Systems. Berkeley,
Content-aware Request Distribution in Cluster-based Network Servers,” CA, USA: USENIX Association, 2003, pp. 8–8.
in Proceedings of the USENIX 2000 Annual Technical Conference, June [28] Mesquite Software, “CSIM,” http://www.mesquite.com.
2000, pp. 323–336. [29] The Apache Software Foundation, “The Apache HTTP Server Project,”
[9] V. S. Pai, M. Aron, G. Banga, M. Svendsen, P. Druschel, W. Zwaenepoel, 2003, http://httpd.apache.org.
and E. Nahum, “Locality-Aware Request Distribution in Cluster-Based [30] ObjectWeb Consortuim, “JOnAS,” 2006,
Network Servers,” in Proceedings of the Eighth International Conference http://jonas.objectweb.org.
on Architectural Support for Programming Languages and Operating [31] MySQL, “MySQL,” 2006, available from http://www.mysql.com.
Systems, 1998, pp. 205–216. [32] ObjectWeb Consortuim, “C-JDBC,” 2006,
[10] W. Eckerson, “Three tier client/server architecture: Achieving scalability, http://c-jdbc.objectweb.org.
performance, and efficiency in client server applications,” Open Infor- [33] D. F. Garcia and J. Garcia, “Tpc-w e-commerce benchmark evaluation,”
mation Systems, vol. 10, no. 1, January 1995. Computer, vol. 36, no. 2, pp. 42–48, 2003.
[11] D. Ersoz, M. S. Yousif, and C. R. Das, “Characterizing network traffic [34] G. E. Alliance, http://www.10gea.org.
in a cluster-based, multi-tier data center,” in ICDCS ’07: Proceedings of [35] B. Urgaonkar, P. Shenoy, A. Chandra, and P. Goyal, “Dynamic pro-
the 27th International Conference on Distributed Computing Systems. visioning of multi-tier internet applications,” in Autonomic Computing,
Washington, DC, USA: IEEE Computer Society, 2007, p. 59. 2005. ICAC 2005. Proceedings. Second International Conference on,
[12] InfiniBand Trade Association, “InfiniBand Architecture 2005.
Specification, Volume 1 & 2, Release 1.2,” October 2004, [36] Z. Wang, C. McCarthy, X. Zhu, P. Ranganathan, and V. Talwar,
http://www.infinibandta.org. “Feedback control algorithms for power management of servers,” in
[13] C. Rusu, A. Ferreira, C. Scordino, and A. Watson, “Energy-efficient 3rd Workshop on Feedback Control Implementation and Design in
real-time heterogeneous server clusters,” in RTAS ’06: Proceedings of Computing Systems and Networks (FeBID’08), Jun 2008.
the 12th IEEE Real-Time and Embedded Technology and Applications [37] Yokogawa Electric Corporation, “WT210 Digital Power Meter User’s
Symposium. Washington, DC, USA: IEEE Computer Society, 2006, Manual,” May 1998.
pp. 418–428.
[14] A. Maria, “Introduction to modeling and simulation,” in WSC ’97:
Proceedings of the 29th conference on Winter simulation. Washington,
DC, USA: IEEE Computer Society, 1997, pp. 7–13.

10

% Razones para alta disponibilidad

More Related Content

What's hot (20)

Similar to % Razones para alta disponibilidad (20)

Recently uploaded (20)

% Razones para alta disponibilidad