Database Reference
In-Depth Information
Given an abstract service S
i
,ifS
i
belongs to more than one data-flow graphs,
then there are many solutions can be used to execute S
i
. The proposed approach
will select the most frequently used solution (from execution history), or ask end
users to select a preferable solution.
8.3
Experiment and Evaluation
Our experiments consist of two parts. First, comparisons are conducted between the
proposed approach and other approaches in small-scale scenarios. Second, compar-
isons are conducted in large-scale scenarios. All the experiments are conducted on
computers with Intel Core 2 Duo 6400 CPU (2.13 GHz and 2GB RAM).
Creation of Experimental Scenarios
Randomly generated scenarios are used for the experiments. Each scenario contains
a control-flow graph and a data-flow graph. QoS values of different concrete
services, virtual machines, database services and network services for each abstract
service are generated randomly with uniform probability. A scenario generation
system is designed to generate the scenarios for experiments. The system first
determines a root pattern (i.e. sequence, conditional, parallel, loop patterns) with
uniform probability for the control-flow. Within this root, the system chooses with
equal probability to either place an abstract services into it or to choose another
composition pattern as substructure. This procedure ends until the generation system
has spent the predefined number (n) of abstract services. All the conditional patterns
have 2 possible options, either of them has the probability of 0:5. Each loop pattern
will run for twice. There are k candidate concrete services to implement each
abstract service. The number of data transferred between each abstract services in
the flow graph is d . Each concrete service can be lodged in m virtual machines,
p database services and q network services. These variables are predefined and
used as input (denoted as
f
n; k; d; m; p; q
g
) to the generation system. Small-
scale scenarios have the input
f
5; 2; 6; 3; 3; 3
g
. Large-scale scenarios have 100
abstract services. Each abstract service can be executed by 30 concrete services.
120 data items are transferred between services and each concrete service is
suitable to run in 20 different VMs, 20 different database services and 20 network
services. The four QoS attributes and the four QoS constraints have same weight
equals 1. The execution QoS, network QoS and storage QoS were randomly
generated with uniform distribution from the following intervals: Q
1
.T i me/
2
Œ100; 2; 000, Q
2
.P ri ce/
Œ200; 1; 000, Q
3
.A
v
ailability/
2
2
Œ0:9750; 0:9999
and Q
4
.Rep
u
tation/
2
Œ1; 100.
