Professional Documents
Culture Documents
Abstract—In the traditional SOA architecture, large-scale the migration process. As to the dynamic migration schemes,
number of services access and execution may result in the poor two important issues should be solved: (1) service code and
performance due to the overload of some nodes, which will data encapsulation and restart at the destination node after
seriously affect the quality of services. Therefore, how to service migration, (2) files configuration, and log
establish high quality and reliable Web Services platform has information and related documents migration [3]. Lu et al.
become a research hotspot, especially in the Cloud computing.
proposed core migration service method to improve the
After analysis of the existing framework in the performance on
load balance, the paper proposed an overall optimal
services viability [4]. In order to meet the flexibility and
framework of Web Services migration for load balance in the scalability in the Web Services framework, services
Cloud systems. In the proposed framework, two strategies of migration can be divided into service replication and services
Web services migration are provided, one is services transfer based on the granularity of the services migration.
replication, and the other is services transfer. Under the Xu adopted services replication scheme to establish a
different situations, the system will dynamically adopt the distributed mutual aid system to alleviate the network
appropriate services migration strategy based on the nodes’ congestion [5]. After analysis the performance of grid
workload. The experiments results show that the proposed services, considering the dynamical grid services and their
Web Services migration framework and the corresponding status, a stateful grid service replication mechanism based on
strategies can obtain better performance on the load balancing virtual workspace was presented [6]. Shen et al. discussed
the services replication by using the middleware technology
Keywords-Cloud computing; Web Services; Service in the Web Services framework [7]. In addition, one
migration; Service replication; Service transferring framework for Web services migration was proposed based
on SOA [8]. The above research emphasized service
I. INTRODUCTION replication or services transfer separately without
considering two scheme together. In this paper, service
Large-scale development of Web Services and Cloud
replication and service transfer were provided. Under the
Computing bring enterprises and users new experience. In
different situations, the appropriate service migration
the traditional service system architecture, large-scale
strategy is adopted based on the current workload.
number of service access and execution make some of nodes
overload and seriously affect the quality of services. III. SERVICE MIGRATION FRAMEWORK
Compared with the traditional computing model, emergence
of cloud computing greatly changed the traditional software A. Load Measurement and Prediction
usage model, and also changed the software development
model. Establishing a high quality and reliable Web services The life cycle of the service instance starts from the
framework in the cloud can better support the Web services request generation. If the node has received the large number
management and maintain the quality of Web services. This of user requests, the thread pool will dispatch the extra
paper proposed an optimal service migration framework to requests to the request queue and wait for the successive
support the prediction and analysis of the nodes’ workload. execution. When the request queue is full, the node refuses
Based on the different situation, the system executes the Web the successive requests arrival. Therefore, the number of the
services replication or transfer. receiving requests in one unit of time can be used to estimate
the workload of the node.
II. RELATED WORK A Web service executing on a virtual node, the maximum
accepted requests of the node in one unit of time is the full
Web services migration refers to migrate services from workload of the node. Assume that one Cloud computing
the original location to the destination for the load balancing
system has n virtual nodes, N 1, N 2 ,..., N n , and m Web
in the system [1]. In a dynamic large-scale, and automated
execution system, service migration is a more satisfactory Services, S1 , S 2 ,..., S m . In order to obtain the services
service deployment method [2]. Adopting the appropriate execution ability of the node, it can use one matrix.
service migration scheme, the system can maintain better Suppose there are j services S 1, S 2 ,..., S j executing on
quality of service. At present, there are two categories
node N i , j connections of these services and j requests sent
migration schemes, static and dynamic schemes. The static
migration schemes suspend or terminate the services during via the connections in one unit time with R1 , R2 ,..., R j . For
Authorized licensed use limited to: b-on: Universidade Aberta. Downloaded on November 08,2020 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.
node N i , it can compute the weighted average of the service migration will be triggered. The procedure is shown
maximum of services requests, denoted as the maximum in Figure
g 1.
workload of the node N i at the time t . The maximum
workload of services requests at N i can be calculated:
i
Rk
Load node _ i ( i
n jk ) (1)
k 1
Rk
k 1
R
k 1
k
ACTnode _ i
t
(2)
Load node _ i
From the formula (2), the current workload of virtual
node N i can be calculated via the number of requests.
By adapting the appropriate time interval, it stores the Figure 1. Illustration of load prediction and service migration
real-time workload of each node ACTntode _ i at every interval.
All of the stored real-time workload ACTntode _ i can combined C. Services Migration Selection
into the time sequence of real-time workload to predict the After a Web service migration has been triggered, the
workload at the next interval. overloaded node will decide to replicate or transfer the
Due to the self-similarity of the variances of workload, it service based on the service’s workload. If the overloaded
is feasible to analyze and forecast the nodes’ workload based service takes up the most of node’s workload, it may lead to
on the time sequence of workload. In this paper, exponential the overloading at the destination through transferring the
smoothing method were adapted to predict the workload on heavy service. If the overloaded service just takes up small
the node. The predictive value of the workload at the next parts of workload, the overloaded service will be transferred
time t 1 is denoted as PREnt ode1 _ i and calculated as follows: to the light-loaded node. Thus, one of appropriate service on
the node is selected and transferred to the other node with
_ i ACTnode _ i (1 ∃ ) PREnode
t 1 t t
PREnode _i
(3) light load. Meanwhile, the service on the source node is shut
The service migration model will utilize PRE t 1
to down to reduce the workload on the source node.
node _ i
The detail selection migration procedure as follows.
trigger the service migration. Assume that there are m services S1 , S 2 ,..., S m in one virtual
B. Services Migration Trigger node, the workload of each services are L1, L2 ,..., Ln ,
To further avoid the workload Ping-Pong effect on the respectively. For any service S k , it can compute the
node, the dual-threshold scheme was adopted to trigger the
workload Lk . Firstly, it can store the time sequence of the
services migration. Ping-Pong effect is that services
frequently migrate between the overloaded nodes and the weighted average Lnode of the maximum service requests on
low-loaded nodes. In the dual-threshold scheme, it set one this node, and the time sequence of the real-time requests
high threshold high and one low threshold low . Node N i R1, R2 ,..., Rk ,..., Rm corresponding to each service. Adopting
t 1
compares its prediction value of workload PRE node _ i with the
exponential smoothing methods, it can compute the weighted
average L 'node of the maximum service requests and the real-
high threshold high . If the PREnode t 1
_ i ! high , it means that
time requests to the services R '1, R '2 ,..., R 'k ,..., R 'm at the
the node is overloaded. The Web services executing on this
node should be migrated to mitigate the workload. If next time t 1 . Thus, it can obtain the workload of each
service L1, L2 ,..., Lk , Ln at the time t 1 as follows:
_ i ∀ low , it means that this node is idle. The node
t 1
PREnode
have sufficient resources to execute extra services. Thus, the R 'k
Lk (1 # k # m) (4)
extra Web services can be migrated to the idle node. If L 'node
low # PREnodet 1
_ i # high , it can be considered that the node When an overloaded node triggered the service
has available resources to support the Web services migration, if the current workload of service L k ! high , it
execution without any migration. Adopting the dual- think that the service occupied the large part of resources of
threshold scheme can effectively reduce the workload of the node, the system will trigger the service replication. If
nodes higher than the threshold high as well as avoid the
Ping-Pong effect. Based on the high and low threshold, the
154
155
Authorized licensed use limited to: b-on: Universidade Aberta. Downloaded on November 08,2020 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.
Lk ∀ low , the service can share the resource evenly, then the and the workload of each service are L1, L2 ,..., Ln . For any
system will execute the service transferring.g service S k , its workload can be obtained by formula (4).
It set the parameters a1, a2 ,..., ai ,..., an (1 # i # n) to 1 or 0.
If ai 0 , the service Si cannot be transferred. If ai 1 , the
service Si will be transferred. The transferred workload of
one node Ltrans can be computed as follow:
n
Ltrans a1L1 a2 L2 ... ak Lk ... an Ln ai Li (5)
i 1
Therefore, the amount of transferred data is as follows:
n
Ctrans a1C1 a2 C2 ... ak Ck ... an Cn ai Ci (6)
i 1
On the one hand, it should maximize the amount of
transferred data C trans before the workload reducing below
the high threshold high during the service migration.
Meanwhile, it should minimize the amount of transferred
data C trans after the workload have decreased below the high
threshold high . Therefore, the problem can be converted to
the standard knapsack problem. The optimal solution can be
Figure 2. Illustration of the Web service replication procedure calculated as follows:
) f %m ∃ 1& +( high .−
∋
D. Web Services Replication Stategy C t rans f %m & (+ high −. Max ∗ (7)
), f %m ∃ 1& +( high ∃ Lm .− Cm
In order to compute the number of connections, and
effectively control the workload of source and destination When C trans is set to the maximum value, it can obtain
nodes simultaneously, the Web services replication scheme every ai (1 # i # n) . If ai 1 , the service Si can be migrated
was executed by adopting the sender-driven strategy to to the destination so that the workload can be reduced below
mitigate the Ping-pong effect. the threshold and save the migration g cost.
In this paper, we extended the sender-driven services
replication strategy. The detail procedure is as follow. If the
workload of the node ACTnode _ i ∀ low , the node i is selected
as the destination node for the service replication. After the
service is migrated, the user connections are also migrated.
When the users’ connections are migrated, it will store parts
of connections with the number of requests in the workload
analyzer. If the workload of the source node is greater than
high , the service replication scheme will be activated. Then
the appropriate services will be migrated. After the services
have been migrated, if the workload of source node is still
greater than high and the workload of the selected
destination is greater than low , it will select one node with
the light workload as the destination node, where the extra
Web services are migrated till the workload is lower than
low ultimately. Figure 2 illustrates the execution procedure
of Web services replication strategy. Figure 3. Illustration of the Web service tranferring procedure
E. Web Services Transferring Strategy
If the system adopts the service transferring strategy to IV. SIMULATION EVALUATION
realize the load balance, how to choose the appropriate Web In the simulations, CloudSim is used to evaluate the
services on the overloaded node is one of the most important performance of the proposed service migration strategies. In
issue for service transferring. the CloudSim, 150 identical virtual nodes are set to configure
Assume n virtual nodes N1 , N 2 ,..., N n in the system, the the Web services. Every virtual node has one core CPU,
data provided by each service are C1 , C2 ,..., Cn , respectively, 1,000 MIPS execution rate, 2GB RAM, and 30GB storage.
155
156
Authorized licensed use limited to: b-on: Universidade Aberta. Downloaded on November 08,2020 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.
In addition, European LHC Distributed Computing Center’s that of the other two strategies. From the point of view of the
workload data are used as simulation data set. number of migrated services, the data migration greedy
strategy should migrate more services, which results in the
decrease of the system efficiency.
Figure 4. load balance performance with service replication strategy Figure 7. The average numbert of migrated service with different strategy
156
157
Authorized licensed use limited to: b-on: Universidade Aberta. Downloaded on November 08,2020 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.