A Dynamic Load Balancing Strategy for a Distributed Biometric
Authentication System
Ana-Maria Pricochi, Andreea Salinca, Sorin Mircea Rusu and Bogdan Ivascu
Research & Development, SOFTWIN, Bucharest, Romania
Keywords: Authentication System, Biometrics, Distributed Computing, Dynamic Signature, Load Balancing, Online
Signature, Service-Level-Agreement, Task Scheduling.
Abstract: Biometrics are a nowadays trend in developing secure systems. Biometric authentication systems must
provide a higher processing capacity, as biometric verifications are complex. Distributed computing relieves
the computational burden, but the efficient scheduling of biometric authentication requests is still an open
problem. This paper presents a distributed system for handling handwritten signature authentication
requests, with a dynamic load balancer ensuring service-level-agreement compliance, low response times,
fair use of resources and high availability. The load balancing method uses the service-level-agreements
information, the classification of processing nodes and the classification of authentication tasks.
1 INTRODUCTION
The need for security constantly increases in the
modern web world. In 2011, the Internet Crime
Complaint Center received over 300,000 complaints,
28,915 being identity theft related crimes.
Biometry emerges as a reliable solution for
security, providing a high level of authentication and
identification. People do not have to remember
passwords or personal identification numbers (PINs)
or to possess something (e.g. smartcards). While
these can be stolen, lost or forgotten, biometric data
authentication implies `something which a person is
or does` (Zhang, 2002).
Biometric authentication methods extract
features and match an input data with a previously
stored template data. Such operations have a greater
complexity than computing the hash of a password
and comparing two password hashes. Biometric
authentication systems demand fast response times
and high availability. In the web environment users
can experience distress when an e-commerce
transaction confirmation or a website login delays.
The unavailability of authentication systems should
not restrict the access to a critical area of a building.
Specialized web-based systems can provide
biometric authentication services to applications
needing to enhance their security. The complexity
and the large number of authentication requests
require high processing capacity systems. A solution
is the usage of parallel and distributed computing.
Considering cost/performance ratio and viability,
distributed computing represents the best option
(Grosu et al., 2002).
Modern distributed systems use heterogeneous
architectures with both multi-core and many-core
processing units (Li et al., 2011), using specialized
co-processors (like Cell Broadband Engine
Architecture – Cell/BE) and CPUs with graphics
processing units (GPUs) or field programmable gate
arrays (FPGAs) (Brodtkorb et al., 2010). Distributed
heterogeneous architectures have proven their
performance. Accordingly to Top 500, 62 of the
world’s best supercomputers use accelerators/co-
processors, 2 being in the first 10.
A challenge in distributed systems is scheduling
tasks while ensuring a balanced load of the
processing nodes and complying with quality of
service (QoS) requirements, particularly when nodes
are heterogeneous (Grosu et al., 2002). In this case,
the optimal mapping of tasks is, in general, a NP-
hard problem (Braun et al., 2001).
The current paper has the following structure:
Section 2 sets the background. Section 3 introduces
our biometric authentication system and section 4
shows our load balancing method. Section 5 details
the implementation. Section 6 discusses the results
obtained so far. Section 7 presents the conclusions.
65
Pricochi A., Salinca A., Rusu S. and Ivascu B..
A Dynamic Load Balancing Strategy for a Distributed Biometric Authentication System.
DOI: 10.5220/0004354400650070
In Proceedings of the 9th International Conference on Web Information Systems and Technologies (WEBIST-2013), pages 65-70
ISBN: 978-989-8565-54-9
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
2 BACKGROUND
In systems with deadlines or other constraints,
finding the optimal task assignment is not feasible.
To simplify the task scheduling problem, studies
consider heuristics, dynamic programming and
relaxation of requirements. Heuristics, popular
approximation techniques, include Opportunistic
Load Balancing, Min-min, Max-min, Minimum
Execution Time, Minimum Completion Time, Tabu,
Genetic Algorithm, Simulated Annealing (Braun et
al., 2001). Another popular choice are directed
acyclic weighted task graphs (DAGs) algorithms.
2.1 Related Work
(Naji et al., 2011) present a fingerprint biometric
system, with a 2.47 seconds average task processing
time and an overall accuracy of 99.48%. Tests are
made on a server, although they prospect the usage
of multiple authentication servers.
(Trevathan and McCabe, 2005); (Trevathan et
al., 2009) studied distributed systems securing
online payments using handwritten signatures. In
their best testing configuration, the authentication
method has a False Acceptance Rate (FAR) of 1.1%
and a False Rejection Rate (FRR) of 1.0%, with no
remarks on system capabilities.
(You et al., 2003) use a centralized scheme for
task scheduling using mobile multi-agents that
dynamically allocate and migrate tasks. The task
scheduling algorithm provides an average round trip
time of 5 to 7 seconds for 100 authentication trials.
The patent of (Yu et al., 1999) describes a web-
based authentication system where the task
allocation considers the type of biometrics specific
servers can process and the content of their local
template database. (Takagi, 2008) presents a load
balancing apparatus using the load evaluation, the
performance property of each authentication device
(clock frequency) and the occupancy rate. (Miller et
al., 2007) propose a biometric authentication system
where the load balancing of biometric matching
engines occurs at router level.
Previously mentioned methods concentrate on
performance goals. The literature does not cover
sufficiently the scheduling biometric authentication
tasks to comply with client contracts. Algorithms
related to machine learning can be optimized to any
goal, but they lead to long scheduling times.
2.2 Proposed Model
The proposed model is a load balancing algorithm in
a heterogeneous distributed biometric authentication
system, using dynamic handwritten signatures. To
comply with client contract, we define load
balancing as a provisioning optimization problem.
Fast mean response times and efficient utilization of
computing resources are secondary goals.
The task distribution method uses worker nodes
and tasks classification to make the assignment. The
load balancing algorithm is dynamic, using
statistical data to make state estimates.
The system uses heterogeneous worker nodes,
assembling a worker farm in a local area network.
Node configurations include CPUs and CPUs with
GPUs. Although machines with GPUs solve faster,
more tasks, simple dual-core CPUs can successfully
process tasks requiring few processing resources.
3 BIOMETRIC
AUTHENTICATION SYSTEM
The proposed method schedules tasks in an
automated system of authentication through
biometric signature (ATHOS). ATHOS is a SoA
(Software Oriented Architecture) based system,
offering online signatures authentication facilities to
applications that improve their security measures.
Multiple clients can concurrently use the system.
The ATHOS system conducts authentication
transactions for its client based on SLAs (Service
Level Agreements), mutually agreed upon and
including terms like: accuracy of authentication,
availability, bandwidth, average response time.
Figure 1: Biometric authentication center.
The ATHOS system comprises authentication
centers with the following modules (Figure 1):
A reversed proxy server, serving as a system
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
66
gateway, filtering and validating requests;
An authentication server, with load balancing
capabilities, formulating tasks based on the
authentication requests and stored templates, and
further distributing them to worker nodes;
A cluster of worker nodes, running processes
executing the authentication tasks received;
A logging module, gathering and storing usage
information during the system’s operation.
A data service, used for the management of the
system databases;
To provide high availability, the ATHOS system
can include several biometric authentication centers
with mirrored databases and back-up servers for
critical components, like the authentication server.
4 LOAD BALANCING SCHEME
The scheme refers to a multi-class dynamic
centralized load balancing algorithm, assigning
authentication tasks with different SLA demands to
heterogeneous machines. The load balancing scheme
must ensure that, for each client, the response time
and the number of requests served complies.
ATHOS system requires an adaptive load
balancing scheme as: the task arrival time is random,
the workflow volume and SLA types vary, and the
worker nodes have dynamic configurations and can
process non-authentication tasks.
4.1 Task Model
The ATHOS system uses online signatures acquired
with a special pen-like device, capturing both
movement and graphic information while a person
signs (Rusu et al., 2011).
The users subscribe templates, consisting of
several specimen signatures. If the system needs to
authenticate a user, the user offers an input signature
to be verified. The authentication method compares
template and input signatures based on the distances
between the feature vectors extracted from the raw
signals of the signatures. A classifier processes the
distances to set the authentication response (Andrei
et al., 2010). More than 1400 feature vectors
comparisons are processed for each authentication.
This method proved a FAR of 1.24% and a FRR of
14.59% on an evaluation of more than 7000
signatures, including skilled forgeries (Salinca et al.,
2012).
Tasks are authentication requests, as they occur
on regular basis and have greater complexity,
influencing system performance. A user usually
registers and deletes signatures once. Tasks have
expiration time, priority and belong to a complexity
class. A task expires when its response time exceeds
the response time defined by its contract.
The priority of the task
j depends on the SLA
of the issuing application
i and additional user data
(Equation 1). The SLA defines an average response
time (
][iT ) and a maximum number of requests per
minute (
][iN ).
min
T and
max
N are the minimum,
respective maximum of the mentioned values, for all
SLAs. The group priority discriminates different
users of the same application (
][iP
group
).
,
and
balance the weight of SLA parameters and group
priority. They are specific to certain system
deployments and have a great influence on the SLA
compliance. Preliminary tests tune these values for
the required performances.
][
][
][
][
max
min
iP
N
iN
iT
T
jP
group
(1)
Handwritten signatures have a complexity,
consisting in loops, intersections, length, direction,
acceleration and pressure variations. The complexity
translates into a number of features and a
computational complexity. Table 1 presents example
processing times (on the same machine) for tasks
belonging to the 5 signature complexity classes used
in ATHOS.
Table 1: Execution times for signature complexity classes.
Very
low
Low Average High
Very
high
22 ms 40 ms 70 ms 110 ms 133 ms
Another classification is urgent and non-urgent
tasks. An urgent task has such an expiration time
that further delaying its assignment leads to
deadline.
4.2 Worker Node Model
Worker nodes are independent computing machines
with multi-core or many-core processing units.
Many-core systems comprise GPUs. A machine runs
multiple authentication processes, depending on its
number of cores and multithreading capabilities.
Processes on multi-core units solve one task at a
time. Processes on many-core units solve multiple
tasks in parallel (packages).
Worker nodes register to the authentication
server, stating the number and type of authentication
processes and receiving a list of load thresholds. The
ADynamicLoadBalancingStrategyforaDistributedBiometricAuthenticationSystem
67
node notifies the authentication server when
reaching the processor or memory load thresholds.
The characteristics stored for each authentication
process include: type and capacity, current load state
of the hosting machine, stability and average
response times for all task classes and, if necessary,
all package sizes. The stability of an authentication
process is the degree of confidence that it completes
processing tasks, within the expected time, without
failure or compromising response. The stability is
the rate of successfully processed tasks, without
deadline violation.
4.3 Load Balancing Method
The load balancing method implies the accumulation
of authentication tasks and, eventually, distributing
them to authentication processes. The distribution is
a continuous repetitive process which, at each step:
Selects a number of tasks;
Classifies tasks as urgent or non-urgent;
Selects the best workers for the current ratio of
urgent and non-urgent tasks;
Assigns urgent tasks to the best workers for this
purpose;
Assigns non-urgent tasks to the best workers for
this purpose.
The major components of the load balancing
scheme are the state estimation and the distribution
module. These components interact with other
modules and data structures (as shown in Figure 2).
Figure 2: The architecture of the load balancing scheme.
Accumulating tasks means storing requests on
arrival in a prioritized structure. The distribution
module selects first the tasks of SLA demanding
applications. To prevent starvation of requests with
less demanding SLAs, a task monitor periodically
upgrades low priority tasks. The number of selected
tasks depends on the worker farm capacity.
The distribution module determines the
proportionality between urgent and non-urgent task
in the selected list. The classifier module uses this
information, along with state estimates, to determine
the scoring of each worker. When most tasks are
urgent, the classifier highlights processes that have
high processing capabilities (multiple authentication
processes and/or parallelization capabilities), low
response times and great stability (redistribution
avoidance). Low performance processes solve the
remaining non-urgent tasks. If non-urgent tasks
dominate, the classifier determines for the few
urgent tasks fast processes with low capacity. It
highlights high capacity nodes, with low stability
and high response time for non-urgent tasks. High
performance processes remain unused, if possible.
Unassigned tasks in the current step return to the
prioritized structure, with an increased priority.
State estimation and control operations run
concurrently with the distribution. The statistical
module stores data of previously solved tasks, for
short periods of time (to reflect the recent state). The
state estimator evaluates authentication processes,
for new worker nodes in the farm and for nodes not
recently used. The evaluation includes test tasks that
cover all complexity classes and package sizes. The
state estimator receives worker node information on
event (processor and memory load changes,
start/stop of processes). State estimates for
authentication processes include average response
times for different complexity classes and package
sizes, operational state and load level.
The state estimation module detects system full
loading: the simultaneous compliance with all SLAs
is not possible with the current worker farm. At full
load, the system degrades all SLA parameters with a
certain percentage. If the performance degradation
exceeds 25%, the system refuses incoming requests,
returning a `System too busy message`.
Equation 2 reflects the condition for full load.
A
#
is the number of applications,
i
T# is the
number of tasks issued by application
i ,
statistic
jt ][
is the average statistic time for task
j
class,
SLA
it ][
is the SLA guaranteed time of application
i and
C
T
is the statistically determined control time.


A
i
i
T
j
CSLA
A
i
i
T
j
statistic
Titjt
#
1
#
1
#
1
#
1
)][(][
(2)
The load balancing scheme allows redistribution,
when a worker node fails during task processing or it
delays the response. We choose redistribution as the
worker farm is reliable and redundancy wastes
computing resources.
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
68
5 IMPLEMENTATION
To evaluate the performance of the proposed load
balancing model, we test real case scenarios for
different system configurations.
We deploy the system, consisting of several Web
Services, on multiple servers. We implement the
Web Services using the Windows Communication
Foundation (WCF) API and the C# programming
language. Internet Information Services (IIS) is the
web hosting solution. We configure the Web
Services and IIS for optimal performances.
We simulate concurrent requests using JMETER
and soapUI. The test model includes multiple SLAs.
The flow of client authentication requests
consists of real signatures, acquired in a previous
data collection stage (Salinca et al., 2012). The
signature database has more than 7000 signatures,
collected from 113 people. The selected signatures
meet the distribution of different signature classes
identified in the database.
6 RESULTS
We run various tests, including different loads and
worker nodes configurations, over various periods.
Table 2 shows the characteristics of different
worker nodes: number of authentication processes
(column 2), number of tasks handled in parallel
(column 3), average response time for medium
complexity tasks (column 4), average number of
tasks processed per minute (column 5).
Table 2: Characteristics of different worker nodes.
Hardware
configuration
No.
proc
.
Task
capacity
Avg.
time
Task
rate
Quad-core
(multithreading)
5 1 510 ms 600
Quad-core
(multithreading)
8 1 551 ms 650
Dual-core with
2 GPUs
2 5 554 ms 700
Quad-core with
3 GPUs
3 7 290 ms 4000
Table 3 shows the SLA compliance for both low and
high system loads, comparing an earlier version of
the load balancing (column 3) and the presented
model (column 4). The earlier version assigns tasks
on arrival to the first available machine capable of
respecting the tasks deadline. The compliance is the
percent of time in which the system respected the
stated SLA parameter, considering each application
separately.
Table 3: Client SLA compliance comparison for a low
load simulation and a high load simulation.
Load SLA parameter
Compliance
in earlier
version
Compliance
in current
version
Low
Maximum
requests number
100% 100%
Average
response time
96.35% 100%
High
Maximum
requests number
98.4% 99.6%
Average
response time
86.54% 90%
The standard deviation of overall performance
degradation for each application, in the high load
simulation, decreased from 16% to 14%. This means
an increased fairness of performance degradation for
applications with different SLAs.
Figure 3: Average response times for different tasks rates.
Figure 3 shows the evolution of system response
times for different system loads (different task
rates). The system comprised two worker nodes, the
second and fourth configuration from Table 2.
Table 4: The task distribution between two machines (one
with a CPU and one with 2 GPUs) and between the GPUs.
Tasks /
min
% tasks on
machine 1
% tasks on machine 2 (% on
GPU 1 - % on GPU 2)
200 50% 50% (50% - 0%)
400 50% 50% (48% - 2%)
600 50% 50% (30% - 20%)
800 52% 48% (23% - 25%)
1000 46% 54% (27% - 27%)
1200 41% 59% (29% - 30 %)
The algorithm also ensures fair distribution between
ADynamicLoadBalancingStrategyforaDistributedBiometricAuthenticationSystem
69
worker nodes, to a certain degree. Table 4 presents
the tasks assignment to different machines and
authentication processes. Machine 1 is a quad-core
with 5 authentication processes. Machine 2 is a
workstation with 2 GPUs, each running an
authentication process executing up to 5 tasks in
parallel. At low loads, the tasks assignment is
balanced among machines, while one authentication
process on machine 2 remains free. As the load
increases, the algorithm assigns more tasks to the
free process, to the point it exploits machine 2 more,
as it has better performance. The assignment
between the 2 processes on machine 2 is balanced.
7 CONCLUSIONS
This paper presents a strategy for load balancing in
heterogeneous distributed biometric authentication
systems. The load balancing algorithm is a SLA
optimization problem. Real case test scenarios,
involving data acquired from potential users of such
a system and actual system deployments show the
high performance of the proposed algorithm.
Furthermore, the experimental results confirm that
such an authentication system can scale while still
offering high QoS.
The algorithm manages to comply with all client
SLAs in normal operation mode and complies with
SLA parameters of 99.6%, respective 90% when the
system exceeds its processing capacity.
Experimental results demonstrate that rates of 6000
request per minute can be reached, establishing the
scalability of the system.
ACKNOWLEDGEMENTS
This work is a result of the research in the project
“Automated system of authenticating through
biometric signature/ATHOS”, co-funded from SOP
IEC PA2: Research, Technological Development
and Innovation for Competitiveness, Operation
2.1.2: “Complex research projects fostering the
participation of high-level international experts”.
REFERENCES
IC3, 2011. Internet Crime Complaint – U.S. Bureau of
Justice, Internet Crime Report, http://www.ic3.gov/
media/annualreports.aspx.
Zhang, D., 2002. Biometric Solutions for Authentication in
an E-World. The Springer International Series in
Engineering and Computer Science.
Grosu, D., Chronopoulos, A., Leung, M., 2002. Load
Balancing in Distributed Systems: An Approach Using
Cooperative Games. In Proceedings of the 16
th
IEEE
International Parallel and Distributed Processing
Symposium, pp. 501-510.
Braun, T., Siegel, H., Beck, N., Boloni, L., Maheswaran,
M., Reuther, A., Robertson, J., Theys, M., Yao, B.,
2001. A Comparison of Eleven Static Heuristics for
Mapping a Class of Independent Tasks onto
Heterogeneous Distributed Computing System. In
Journal of Parallel and Distributed Computing, pp.
810-837.
Li, T., Narayana, V., El-Ghazawi, T., 2011. A Static Task
Scheduling Framework for Independent Tasks
Accelerated using a Shared Graphics Processing Unit.
In IEEE 17
th
International Conference on Parallel and
Distributed Systems, pp. 88-95.
Brodtkorb, A., Dyken, C., Hagen, T., Hjelmervik, J.,
Storaasli, O., 2010. In Scientific Programming, IOS
PRESS, pp. 1-33.
Top 500 Super Computers, 2012. November 2012,
http://www.top500.org/.
Naji, A., Housain, A., Zaidan, B., Zaidan, A., Hameed, S.,
2011. Security Improvement of Credit Card Online
Purchasing System. In Scientific Research and Essays,
Academic Journals, pp. 3357-3370.
Trevathan, J., McCabe, A., 2005. Remote handwritten
signature authentication. In Proceedings of the 2
nd
International Conference on e-Business and
Telecommunication Networks, pp. 335-339.
Trevathan, J., McCabe, A., Read, W., 2009. Online
Payments Using Handwritten Signature Verification.
In 6
th
International Conference on Information
Technology: New generations, pp. 901-907.
You, J., Zhang, D., Cao, J., Guo, M., 2003. Parallel
Biometrics Computing Using Mobile Agents. In
Proceedings of the International Conference on
Parallel Processing, pp. 305-312.
Yu, Y., Stephen, W., Hoffberg, M., 1999. Web-Based
Biometric Authentication System and Method. U. S.
Patent and Trademark Office, US5930804.
Miller, J., Willis, W., Lau, J., Dlab, B., 2007. Multimodal
Biometric Platform. U. S. Patent and Trademark
Office, US7298873.
Takagi, J., 2008. Load Balancing Apparatus. U. S. Patent
and Trademark Office, US0031496.
Rusu, S., Dinescu, A., Diaconescu, S., 2011. Systems and
methods for assessing the authenticity of dynamic
handwritten signature. World Intellectual Property
Organization WO/2011/112113.
Salinca, A., Rusu, S., Diaconescu, S., 2012. An approach
to data collection in an online signature verification
system. In Proceedings of the WebIST’12, pp. 251-
254.
Andrei, V., Rusu, S., Diaconescu, S., 2010. Solutions for
speeding-up on-line dynamic signature authentication.
In Proceedings of the 12
th
International Conference on
Enterprise Information Systems, pp. 121-126.
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
70