High-available grid services through the use of virtualized clustering.
ABSTRACT Grid applications comprise several components and web-services that make them highly prone to the occurrence of transient software failures and aging problems. This type of failures often incur in undesired performance levels and unexpected partial crashes. In this paper we present a technique that offers high-availability for Grid services based on concepts like virtualization, clustering and software rejuvenation. To show the effectiveness of our approach, we have conducted some experiments with OGSA-DAI middleware. One of the implementations of OGSA-DAI makes use of use of Apache Axis V1.2.1, a SOAP implementation that suffers from severe memory leaks. Without changing any bit of the middleware layer we have been able to anticipate most of the problems caused by those leaks and to increase the overall availability of the OGSA-DAI Application Server. Although these results are tightly related with this middleware it should be noted that our technique is neutral and can be applied to any other Grid service that is supposed to be high-available.
- SourceAvailable from: Kishor S Trivedi[show abstract] [hide abstract]
ABSTRACT: In this paper we present a comparative experimental study of the main software rejuvenation techniques developed so far to mitigate the software aging effects. We consider six different rejuvenation techniques with different levels of granularity: (i) physical node reboot, (ii) virtual machine reboot, (iii) OS reboot, (iv) fast OS reboot, (v) standalone application restart, and (vi) application rejuvenation by a hot standby server. We conduct a set of experiments injecting memory leaks at the application level. We evaluate the performance overhead introduced by software rejuvenation in terms of throughput loss, failed requests, slow requests, and memory fragmentation overhead. We also analyze the selected rejuvenation techniques’ efficiency in mitigating the aging effects. Due to the growing adoption of virtualization technology, we also analyze the overhead of the rejuvenation techniques in virtualized environments. The results show that the performance overheads introduced by the rejuvenation techniques are related to the granularity level. We also capture different levels of memory fragmentation overhead induced by the virtualization demonstrating some drawbacks of using virtualization in comparison with non-virtualized rejuvenation approaches. Finally, based on these research findings we present comprehensive guidelines to support decision making during the design of rejuvenation scheduling algorithms, as well as in selecting the appropriate rejuvenation mechanism.Performance Evaluation. 03/2013; 70(3):231–250.
- [show abstract] [hide abstract]
ABSTRACT: Software aging and rejuvenation research has shown that the application of approaches for software aging modeling, monitoring, and rejuvenation has the potential to significantly increase software resilience. In this chapter, we present an overview of important analytical models and measurement approaches for software aging and rejuvenation. We start by describing the Markov based approaches and renewal process based approaches for software aging and rejuvenation modeling. In addition, we present measurement based approaches using both online and offline methods for software rejuvenation. We conclude by presenting a categorization of the approaches and by presenting a brief overview of applicability of each of the approaches presented in this chapter. Google preview: http://goo.gl/cgHtq01/2012: pages 167-183;
Conference Proceeding: A Generic Approach to the Services Delivery in Enterprise eBusiness Platforms[show abstract] [hide abstract]
ABSTRACT: The contemporary business environment can be seen as a logical effect of eBusiness, advanced integrated networks, and software services. In such an environment application deployment models have revolutionized the way businesses interact, collaborate and transact with customers, suppliers, partners, employees and shareholders. The purpose of our work is to develop a generic model for performance analysis of software services in order to facilitate enterprises to attach their customers more tightly by an effective service delivery which in turn optimizes business processes at different steps. This communications enablement allows not only greater synergy and velocity among vendors, suppliers and partners within the value creation and delivery chain of the business, but also, creates responsive and seamless customer service.Advanced Information Networking and Applications (WAINA), 2011 IEEE Workshops of International Conference on; 04/2011
High-Available Grid Services through the use of
Javier Alonso#, Luis Silva*, Artur Andrzejak+, Paulo Silva* and Jordi Torres#
# Computer Architecture Department, Technical University of Catalonia
Barcelona Supercomputing Center
C/ Jordi Girona 1-3, Building C6, Campus Nord. 08034 Barcelona, Spain
* Dep Eng Informática, University of Coimbra,
+ Computer Science Research Department
Zuse Institute Berlin (ZIB)
Takustr. 7, 14195 Berlin, Germany
Grid applications comprise several components and web-services
that make them highly prone to the occurrence of transient
software failures and aging problems. This type of failures often
incur in undesired performance levels and unexpected partial
crashes. In this paper we present a technique that offers high-
availability for Grid services based on concepts like
virtualization, clustering and software rejuvenation. To show the
effectiveness of our approach, we have conducted some
experiments with OGSA-DAI middleware. One of the
implementations of OGSA-DAI makes use of Apache Axis
V1.2.1, a SOAP implementation that suffers from severe memory
leaks. Without changing any bit of the middleware layer we have
been able to anticipate most of the problems caused by those
leaks and to increase the overall availability of the OGSA-DAI
Application Server. Although these results are tightly related
with this middleware it should be noted that our technique is
neutral and can be applied to any other Grid service that is
supposed to be high-available.
Grid computing deals with the usage of large scale and
heterogeneous resources in geographically dispersed sites.
The target applications which are executed over Grid
environments usually require a huge execution time and
computation capacity. If there is a failure in one or more Grid
services used by the application, all the work that has been
done can be simply lost.
Unfortunately these complex environments are highly
prone to the occurrence of transient failures . Some of these
failures are easy to fix by the system administrator provided
there is a deterministic pattern and it is easy to find the cause
of the error.
However, there are a large number of failures that are quite
difficult to spot due to the indeterminist nature and difficulty
to re-create them during the final acceptance tests of the
software. One particular case is the occurrence of software
aging. This problem has been reported in telecommunication
systems , web-servers , enterprise clusters , OLTP
systems  spacecrafts systems  and even military systems
Since it is impossible to guarantee that the mean time
between failures (MTBF) will be bigger than the total
execution time of a single application, it is needed to provide
some mechanisms to guarantee the availability of the
application despite the occurrence of transient failures, more
than the support for fault-tolerance.
In the last twenty-five years we have seen a considerable
amount of research in the topic of fault-tolerance and high-
availability. Techniques like server redundancy, server fail-
over, watchdogs, replication schemes, checkpoint and rollback
have been extensively studied in the literature. However,
most of those techniques only act when a failure has happened.
If the system anomaly does not result in a crash but rather in a
decrease in the performance levels or some instability in the
1-4244-1560-8/07/$25.00 © 2007 IEEE8th Grid Computing Conference34
quality of service (QoS) metrics most of those techniques are
not even triggered.
We felt it would be important to extend the failure model
and to consider the occurrence of fail-stutter failures . In
this paper we will present a software solution to achieve high-
available and high-responsive application servers that provide
Grid services. Our technique makes use of several concepts: a
virtualization layer that is installed in every application server
that is running a Grid service; the use of primary-backup
scheme for server replication; the implementation of system
surveillance scheme to detect anomalies in a timely-reactive
manner; the adoption of software rejuvenation and a clean
migration mechanism to transfer the execution between
servers without losing on-going requests. All of these
concepts and techniques have been merged in one package
and provide a cost-effective solution for high-availability,
without requiring additional servers and load-balancing
machines, which would increase the budget of the
infrastructure. Our technique has been experimentally
evaluated using a well-known package of Grid middleware:
OGSA-DAI. The results are quite promising.
The rest of the paper is organized as follows: Section 2
presents the goals of our technique for high-availability.
Section 3 describes the virtualized clustering architecture.
Section 4 presents the experimental setup that has been used
in our experiments. Section 5 presents a study about the
overhead impact of using a virtualization layer in our solution.
Section 6 presents some results which show the effectiveness
of our solution. Section 7 evaluates the downtime that is
achieved using virtualized clustering. Section 8 concludes the
II. HOW TO ACHIEVE OUR GOALS
The first step in our approach is to achieve a very effective
anomaly detection mechanism. This is achieved by using
external surveillance of QoS metrics and internal monitoring
of system parameters. The surveillance of external QoS
metrics has been proved relevant in the analysis of web-
servers and services  and it has been highly effective in the
detection of software aging and fail-stutter failures . This
is tightly combined with the monitoring of system internal
metrics. All these monitoring probes should be timely-
reactive and able to detect potential anomalies, before the
occurrence of a system failure. Even if this detection would
lead to some false-alarms it would be important to avoid the
occurrence of unplanned crashes.
When a potential anomaly is detected we should apply an
automatic recovery action. In most of the cases we apply a
rejuvenation action. To avoid losing on-going requests when
there is a planned restart or rejuvenation we make use a Grid-
service backup that has been running in the same server
machine of the primary Grid service. To allow the execution
of two replicas of the Grid service in the same machine we
will make use of a virtualization layer and we will launch a
minimum of three virtual machines per server machine. This
way we are not increasing the cost of the infrastructure since
we are exploiting the concept of virtualization and server
Since in our technique we inherit some of the concepts of
server clustering together with virtualization we decided to
name this approach by “Virtualized Clustering”.
The use of Virtualization in every server that is running a
Grid Service may be seen as a source of performance
degradation since we are introducing a new abstraction layer.
However, the results obtained with our proposal are promising
and clearly show that the use of virtualization is clearly
III.VIRTUALIZED CLUSTERING ARCHITECTURE
In this section we describe our proposal to offer high
availability grid services. As explained, our software solution
for high-availability assumes the adoption of a virtualization
layer in the machine that provides the Grid service. In our
prototype implementation we made use of XEN , although
it is also possible to use other virtualization middleware like
VMWare  and Virtuoso , among others.
Our solution requires the creation of three virtual machines:
one virtual machine with our own load balancer module (VM-
LB machine); a second virtual machine where the main grid
service runs; and a third virtual machine where we have a
replica of the grid service which works as a hot-standby grid
service. Figure 1 represents the conceptual architecture of our
- ANOMALY DETECTOR
- S-PROBE COORD
- SRA COORD
- ACTIVE GRID SERVICE
- SRA AGENT
- STANDBY GRID SERVICE
- SRA AGENT
Figure 1: Virtualized Clustering architecture description.
VM1 machine runs a set of components: our Load Balancer
(LB); a module that collects system data from the grid service;
a module that detects potential anomalies; a software
watchdog; and a last module that is the coordinator of the self-
recovery actions. In our prototype we have used some open
source modules modified to achieve our targets. The LB
module was implemented using Linux Virtual Server (LVS)
. LVS is a 4-Layer load-balancer that provides IP fail-over
capabilities and a myriad of balancing policies.
The S-probe coordinator collects data from the Grid
service and the underlying operating system. It requires the
installation of some probes that collect some system metrics
like CPU, memory usage, swap usage, disk usage, and number
of threads, among other system parameters. This module
provides data feed to the last module: Anomaly-Detector.
The Anomaly-Detector is a core module in our system: for
the time being, we are using some simple threshold techniques
based on the observed conditions and system metrics of the
target grid service. In the future, we plan to enhance this
module with statistical learning techniques and time-series
analysis to provide more accurate forecasts of the occurrence
of potential anomalies.
When the anomalies are not detected in time and the Grid
service ends up to crash this is directly detected by the
Watchdog module. For this module we used a modification of
ldirectord tool . This tool modifies LVS tables directly
and it executes a test over a static HTML page to check if the
server is alive. We have enhanced this tool with some
improvements in the time-out policies to avoid the occurrence
In the other two virtual machines, we installed the Grid
service. In every VM we install a SRA Agent. This agent has
the responsibility for the recovery action. This module is
coordinated by the SRA-Coordinator. Another module that is
installed in every VM is the S-probe. This probe collects
system metrics and external QoS metrics like throughput and
latency from the grid service and sends data to the S-probe
coordinator which is running in the VM1 machine.
The SRA-Coordinator works in coordination with the two
SRA-Agents installed in grid services machines. When this
module decides to apply a planned restart to clear up the
potential anomaly of the system we should apply a careful
migration mechanism between the primary and the backup
Grid service. This procedure is implemented by the up-front
Load-Balancer that assures a transparent migration for the
client applications that are assessing that Grid Service. This
operation should also preserve data consistency and assure
that no in-flight request should be lost during the migration
phase. For this reason we have a window of execution where
we have both servers running in active mode: the hot-standby
service is made active and will receive the new requests while
the old grid service should finalize the requests that were still
on-going. This way we do not lose any in-flight request and
we assure the data consistency.
Our software solution has been implemented in Linux, and
in two of the modules we have used some open-source tools,
like LVS or Ldirectord. The deployment of our framework
does not require any change to the grid services or the
middleware containers. They are also neutral to the
virtualization layer. As soon as these software modules are
made robust for third-users they will be made available for the
To evaluate our Virtualized Clustering framework we
elected as benchmark Grid service the OGSA-DAI
OGSA-DAI is a package that allows remote access to data-
resources (files, relational and XML databases) through a
standard front-end based on Web services. The software
includes a collection of components for querying,
transforming and delivering data in different ways, and a
simple toolkit for developing client applications. OGSA-DAI
provides a way for users to Grid-enable their data resources.
The front-end of OGSA-DAI is a set of Web-services that
in the case of WSI requires a SOAP container to handle the
incoming requests and translate them to the internal OGSA-
DAI engine. This SOAP container is Tomcat/Axis 1.2.1 .
The detailed description of the OGSA-DAI internal is out-
of-scope of this paper. At the moment OGSA-DAI
middleware is used in several important Grid projects ,
including: AstroGrid, BIoDA, Biogrid, BioSim-Grid, Bridges,
caGrid, COBrA-CT, Data Mining Grid, D-Grid, eDiaMoND,
ePCRN, ESSE, FirstDIG, GEDDM, GeneGrid, GEODE,
GEON, GridMiner, InteliGrid, INWA, ISPIDER, IU-
RGRbench, LEAD, MCS, myGrid, N2Grid,OntoGrid, ODD-
Genes, OGSA-WebDB, Provenance, SIMDAT, Secure Data
Grid, SPIDR, UNIDART and VOTES.
This list is clear representative of the importance of OGSA-
DAI and its relevance as a Grid Service.
Although OGSA-DAI is being used by several Grid
projects and a considerable community it includes an internal
point of concern, from the point of view of availability.
OGSA-DAI WSI makes use of Apache Axis 1.2.1, as the
SOAP router. This implementation of Axis is highly prone to
internal memory leaks, as has been reported in . When the
service developers of OGSA-DAI make use of session-scope
in their Grid services those memory leaks result in a clear
scenario of software aging, with performance degradation and
even system crashes.
For this reason it is quite reasonable testing our framework
with OGSA-DAI WSI, when the services are configured with
session scope. Without our framework the OGSA-DAI server
would crash or present request failures in that described
scenario and after some time usage. This would undermine
completely the availability of the Grid Application that would
make use of an OGSA-DAI service. To increase the
availability of this service we have installed our framework in
an OGSA-DAI server and we conducted an experimental
evaluation with some synthetic workload.
To simulate a set of clients we have used the QUAKE tool
. This is a tool of dependability benchmarking that was
implemented in our Labs and it facilitates the launching of
simultaneous multiple clients that execute requests in a server
under test with a programmed workload.
In our experiments we used a cluster of 5 machines: 3
running the client benchmark application, one Database server
(Tania) running two virtual machines and our main server
(Katrina) running three virtual machines. All machines are
interconnected with a 100Mbps Ethernet switch. The detailed
description of the machines is presented in Table 1.
DETAILED EXPERIMENTAL SETUP DESCRIPTION
Dual Core AMD64
5.0.28, Axis 1.2.1
OGSA-DAI WSI 2.2
V. VIRTUALIZED CLUSTERING OVERHEAD
As we described before, our solution runs over a
virtualization layer. This new added abstraction layer over the
stack increases the system overhead. For this reason, our first
experiment was focused on calculating the overhead
introduced by our solution. After that, we developed an
accurate mathematical study to measure the advantages versus
the disadvantages of using virtualization.
A. Virtualized Clustering Overhead
Our first experiment was to measure the performance
penalty due to use a virtualization layer (XEN) and our
virtualized clustering technique.
We executed several short time runs (10 min) on burst
mode and we collected the results of three configurations:
OGSA-DAI on top of the Operating System, OGSA-DAI on
top of one virtual machine, and a last configuration where we
executed OGSA-DAI with all the components of the
virtualized clustering framework to measure the overhead of
Figure 2 shows the throughput of the three configurations
in those test-runs of 10 minutes. It can be seen that there is
some overhead for using XEN and our virtualized clustering
Throghput overhead comparison
180 210 240 270 300 330 360 390 420 450 480 510 540 570 600
OGSA-DAI on top of XEN + Virtualized Clustering
OGSA-DAI on top of XEN
OGSA-DAI on top of OS
Figure 2: Throughput Comparison.
Table 2 gives more detailed numbers obtained in these
experiments: average latency, average throughput, total
number of requests and throughput overhead.
OVERHEAD COMPARISON RESULTS
On top of
On top of
By comparing the total number of requests we can see that
the virtualization overhead introduces a 10% of overhead and
our solution adds a 2% additional overhead. These results
have been achieved with a burst workload, so they should be
seen by the reader as the maximum observable overhead.
One could argument that the virtualization middleware
introduces a visible overhead in the performance Grid service.
Due to the introduction of an overhead we decided to make a
study to calculate the advantages of virtualization.
B. Average Performance under Virtualization Overhead
The usage of a virtualization layer obviously decreases
application performance, while on the other hand it reduces
the Mean Time to Recovery (MTTR) in case of a fault or
rejuvenation need. This gives rise to a question what is more
beneficial for the long-term performance: given a large time
interval with many rejuvenation cycles, will an application
Aphys running on a physical server serve more requests than an
identical application Avirt running in a virtual machine, or the
other way around?
To treat this problem we introduce the following definitions.
Assume that under the full load (burst request distribution)
Aphys can serve at most Pp requests per second which we call
its instantaneous performance (and could be called the
maximum throughput), and let Pp depend on the time since
last rejuvenation. Under the same conditions let Pv=Pv(t) be
the number of requests per second served by Avirt. The
average performance Pp is defined and computed by summing
up the number of served requests over a large time interval
and dividing it by the length of this interval (analogously Pv).
The Mean Time to Failure (MTTF) t gives us the expected
time until an application must be rejuvenated, and we assume
that it is equal for both Aphys and Avirt. However, each
application has a different MTTR which we designate as rp
and rv, respectively. We introduce the ratio q= Pv/Pp called the
performance ratio, which is less than 1 due to virtualization
overhead. Even if the instantaneous performances vary over
time, we might assume that q is constant since both servers
First, to compute the average performance under a burst
distribution with a fixed request rate above max(Pp,Pv) it is
sufficient to consider just one failure and recovery cycle while
assuming that MTTF t is not an expectation but a fixed time.
We obtain then:
The major question to be answered is: what is the minimum
level qmin of the performance penalty under which the average
performance of the virtualised application Avirt is still
matching the average performance of Aphys?
By setting Pp = Pv and solving by q we find out that under the
These solutions look different if the applications are not
running under their respective full load. For example, if the
request rate never exceeds Pv, then there is no difference in the
throughput between applications. In the latter case qmin might
take any value above 0 and the ratio of the average
performances depends only on the ratio of the MTTR’s rv and
rp. Furthermore, if the requests arrive according to some more
complex distribution, e.g. the Poisson distribution, then only
time intervals with arrival rates above the momentary levels
Pv(t) and Pp(t) contribute to the differences between average
To compute qmin in the experimental setting of the OGSA-
DAI server we have compared the throughput of the non-
virtualized case (“On top of OS”) vs. the virtualized case
without the rejuvenation service (“On top of XEN”) and with
this service (“Virtualized Clustering”). To safeguard against
the effects of the initialization phase and the aging effects we
compared the service rates after the initial 5 minutes since
start for the duration of 5 minutes. Table 3 shows that the
performance ratio q was 2141/2383 = 0.8984 for the case “On
top of XEN”, and 2090/2383 = 0.8770 for the rejuvenation
MEASURED VIRTUALIZATION OVERHEAD IN THE TIME INTERVAL 5-10 MIN
On top of
On top of
MINIMAL REQUIRED PERFORMANCE QMIN DEPENDING ON THE MTTF T FOR R
=0 AND R=12.5 SECONDS
To confront this value with the computed minimal required
performance qmin to match the average performance in both
modes, we use as values of the MTTR parameters rv = 0 and rp
= 12.5sec. Under the assumption of a variable MTTF t we
obtain Table 4 with the values for qmin. It shows that only for
very short rejuvenation cycles (below 125 sec) the average
performance in the virtualized case does not fall behind the
case of Aphys. Since the aging effects are noticeable after much
longer time (above 30 minutes), the average performance of
Avirt is in general worse than the one of Aphys. However, for
the cost of about 12.3% average performance -with the worst
case only in the burst mode - we completely eliminate outages
caused by software aging, which is a fair trade.
VI.HOW EFFECTIVE IS OUR APPROACH?
After the reasoning about the virtualization overhead, we
have to demonstrate the effectiveness of our approach. The
main goal of this experiment was to demonstrate that our
solution can apply an automatic rejuvenation action without
losing any client request.
We started by conducting several experiments to observe
the behaviour of OGSA-DAI and the underlying system
metrics. We concluded that external QoS metrics were very
unstable and it would be quite difficult to apply some
threshold analysis to trigger automatic recovery action. After
that, we studied in detail the underlying system metrics and
we observed that memory usage was increasing with time
until the maximum memory usage allowed by the
configuration. When this happen the OGSA-DAI server had
an unpredictable behaviour with very unstable performance
and sometimes it resulted in system crashes.
Since we knew the cause of the anomaly was an internal
memory leak on Axis layer, we decided to apply a threshold
trigger associated with memory usage. Thereby, we have
configured our S-Probe to obtain the memory usage every 5
seconds and when it would get higher than a certain threshold
we would apply a rejuvenation action in the main Grid service
to avoid a possible failure or crash.
We ran two configurations: OGSA-DAI without our
solution to obtain reference curves; and then OGSA-DAI with
our solution. In this latter case, we have configured the
threshold to trigger when memory usage gets to a certain limit:
50% of maximum memory. Both experiments were executed
with a workload on burst mode.
Figures 3 and 4 show the throughput and the latency results,
respectively. We can observe that our solution was able to
achieve a higher throughput and a lower latency.
240 450660 870
1080 12901500171019202130 2340 255027602970318033903600
Throughput (normal execution)
Throughput (with virtualized clustering)
Figure 3: Throughput comparison with and without our solution.
In Figure 4 it is clear that the usage of our virtualized
clustering approach was able to provide a very stable latency,
while in the normal case the OGSA-DAI server (with session-
scope) presented a very unstable latency and in some cases we
even registered some missed requests.
Latency (normal execution)
Latency (with virtualized clustering)
Figure 4: Latency comparison with and without our solution.
The reader could ask himself, what is the reason for this
difference between throughput results? Because you can
suppose that in both experiments the throughput will be more
or less the same, because the hardware is the same. However,
this is not enough to achieve the same throughput. Because a
normal execution of OGSA-DAI consumes the memory
quickly and after that the throughput goes down like we can
observe around second 30 in figure 3. Our solution avoids that
fact, improving the results of the OGSA-DAI behaviour.
Table 5 presents more detailed numbers of this experiment.
We can see three main things:
Our approach was able to avoid any missed
request. Without our framework we have
observed 25 missed requests during that
timeframe of the experiment;
Our approach was able to reduce the average
latency by a factor of 3;
Our approach was able to improve the average
throughput by a factor of 3.
Even if we had to pay the overhead of virtualization we
were able to improve the performance of OGSA-DAI server
by a considerable factor, just by applying some planned
restarts to avoid performance failures. With this data we are
able to argument about the cost-effectiveness of our solution
for high-available Grid services.
DETAILED VALUES FROM 1 HOUR OF EXECUTION
VII.DOWNTIME ACHIEVED WITH THE VIRTUALIZED
After showing effectiveness of our solution to obtain a high
available and highly-stable grid service, we wanted to
evaluate the potential downtime for the grid service when a
recovery action was executed as well as the number of failed
requests during this process.
This experiment was very important to achieve one of our
main goals. As mentioned we wanted to develop a solution
that when a recovery action is done any work-in progress
would not be lost to guarantee that client application would
not see any request failure and would notice the OGSA-DAI
server as always available.
We run four different experiments to compare our
technique with other simpler restart solutions:
• One runs where we applied a recovery action at the time
of 300 seconds.
• Another run where we triggered an OGSA-DAI restart
at exactly that time.
• A third runs, where we applied a restart in the XEN
Virtual machine where the main server was running, at
the same time.
• And finally a last run, where we executed a full
machine reboot at that time.
The results are presented in Figure 5, which presents an
execution window of 600 seconds. We adopted that graph
format to show the client perception about the availability of
the grid service when a “restart” is done.
Figure 5: Client perceived availability for different restart mechanisms.
We can observe that when our solution is triggered there is
no perceived downtime: from the point of view of the clients
the Grid service is always available. This is achieved due to
the execution window where both grid services (primary and
hot-standby) are running at simultaneous time.
The restart of the OGSA-DAI server had a visible
downtime of 25 seconds. A restart of the XEN Virtual
Machine resulted in a downtime of 92 seconds. Finally, a
restart of the full operating system incurred in a downtime of
The next step was to measure the impact of every “restart”
process according to the number or requests lost. The results
are presented in Table 6. We included the average throughput
obtained in every experiment during a test-run of 10 minutes.
Assuming that there is one unique restart action in those 10
minutes we present the total requests during the execution, the
number of failed requests, the average throughput in that test-
run of 10 minutes and the observed downtime.
COMPARING DOWNTIME AND FAILED REQUESTS WITH DIFFERENT RESTART
When applying our mechanism we were able to achieve the
minimum MTTR as possible: zero. The Grid service was
always available and no downtime was perceived by the client
applications. We were also able to register zero failed requests.
Other restart solutions incurred in several failed requests
(from 70 up to 2353) and a visible downtime.
This experiment has shown that our virtualized clustering
approach is a quite promising solution to achieve a high
available grid service, with zero downtime and zero failed
requests and without incurring any additional cost in terms of
servers and IT infrastructure.
In this paper, we have presented a software solution for is
highly effective for software aging, fail-stutter and
performance failures. Our solution makes use of a
virtualization layer that although it introduces some overhead
it pays-off in terms of consolidation and ease of deployment.
We have proved the effectiveness of our solution in the case
of OGSA-DAI, but are convinced that similar results can also
be observed in other Grid Services that present some latent
Another important achievement is the fact that our solution
is completely independent the server under test. We can use
our framework with legacy software without requiring any
The results presented in this paper encouraged us to
consider that this approach for high-availability can be a
partial contribution for the deployment of dependable grid
services and we are currently enhancing our solution with
support for the real-time forecast of failures and critical
This research work is supported by the FP6 Network of
Excellence Core-GRID funded by the European Commission
(Contract IST-2002-004265) and the Ministry of Science and
Technology of Spain and the European Union (FEDER funds)
under contract TIN2004-07739-C02-01.
Oppenheimer, D., Archana Ganapathi, and David A. Patterson. “Why
do Internet Services fail, and What can be done about it?” 4th USENIX
Symposium on Internet Technologies and Systems (USITS’03), March.
A. Avritzer, E. Weyuker, Monitoring Smoothly Degrading Systems for
increased Dependability, Empirical Software Eng. Journal, Vol 2, No 1,
pp. 59-77, 1997.
L.Li, K.Vaidyanathan, K.Trivedi. “An Approach for Estimation of
Software Aging in a Web-Server”, Proc. of the 2002 International
Symposium on Empirical Software Engineering (ISESE'02)
V. Castelli, R. Harper, P. Heidelberg, S. Hunter, K. Trivedi, K.
Vaidyanathan, W. Zeggert, “Proactive Management of Software
Aging” IBM Journal Research & Development, Vol. 45, No. 2, Mar.
K. Cassidy, K. Gross, A. Malekpour. “Advanced Pattern: Recognition
for detection of Complex software Aging phenomena in Online
Transaction Processing Servers”, Proc. of the 2002 Int. Conf. on
Dependable Systems and Networks, DSN-2002.
A. Tai, S. Chau, L. Alkalaj, H. Hecht. “On-Board Preventive
Maintenance: Analysis of Effectiveness an Optimal Duty Period”, Proc.
3rd Workshop on Object-Oriented Real-Time Dependable Systems,
E. Marshall. “Fatal Error: How Patriot Overlooked a Scud”. Science,
p. 1347, Mar. 1992.
R. Arpaci-Dusseau, A. Arpaci-Dusseau. “Fail-stutter Fault Tolerance”.
Proc. 8th Workshop on Hot Topics in Operating Systems. (HOTOS-
D. Menascé. “QoS Issues in Web Services”. IEEE Internal Computing,
L. Silva, H. Madeira and J.G. Silva. “Software Aging and Rejuvenation
in a SOAP-Based Server”. IEEE-NCA: Network Computing and
Applications, Cambridge USA, July 2006.
Renato J. Figuereido, Peter A. Dinda, José A. B. Fortes. “A Case For
Grid Computing on Virtual Machines”. Proc. of the 23rd Int. Conf. on
Distributed Computing Systems, p. 550, May 19-22, 2003.
R. Figuereido, P. Dinda, J. Fortes, “Resource Virtualization
Renaissance” IEEE Computer, 38(5), pp. 28-69, May 2005.
(2007) Xen Source website [Online]. http://www.xensource.com/
(2007) VMWare website. [Online]. http://www.vmware.com/
(2007) Virtuoso website. [Online]. http://www.virtuoso.com/
(2007) LVS website. [Online]. http://www.linuxvirtualserver.org/
(2007) Ldirectord website. [Online].
(2007) OGSA-DAI website. [Online]. http://www.ogsadai.org.uk/
(2007) Apache Axis. [Online]. http://ws.apache.org/axis
(2007) OGSA-DAI Projects, [Online].
W. Hoarau, S. Tixeuil, N. Rodrigues, D. Sousa and L. Silva.
“Benchmarking the OGSA-DAI Middleware”. CoreGrid Technical
Report, No. TR-0060. October 5, 2006. http://www.coregrid.net
S. Tixeuil, W. Hoarau, L.M. Silva, “An Overview of Existing tools for
Fault-Injection and Dependability Benchmarking in Grids”, CoreGrid
Technical Report TR-0041, http://www.coregrid.net