Information Technology Reference
In-Depth Information
4.2
Resource Management Framework
In [9], the resource management framework is mentioned and resource management
problems are simplified into feedback control problems. The first two objectives men-
tioned above can be achieved by feedback control. PID (Proportional-Integral-
Derivative) feedback control is adopted to maintain high CPU utilization. Besides,
VM scheduling algorithm is designed to manage VMs in order to achieve cost effi-
ciency as well as QoS guaranteed applications.
With its three-term functionality covering treatments to both transient and steady-
state responses, PID control offers the simplest and yet most efficient solutions on
many real-world control problems [10]. The reasons why PID control is employed in
this paper include: i) PoI is essentially a dynamic system and PID control technique is
widely accepted technique in dynamic systems; ii) precise mathematical model of the
system is not required in PID control technique. Instead, PID control technique can
achieve satisfactory performance based on an approximate model. The complexity of
PoI makes it difficult to be described precisely by mathematical model, which makes
PID control technique a great candidate for underlying resource management.
A classic feedback control system is composed of a controller, a plant (the object
to be controlled no matter what it is) and sensors (the object to measure the output of
the plant) [11]. Controlled Variables are the variables to be controlled. Set Points
represent the correct and expected values of the Controlled Variables. The difference
between the current value and the Set Point is the Errors. The whole feedback and
control loop is aimed to reduce the Errors.
•
The sensor periodically monitors and compares the
Controlled Variables
to the
Set
Points
to determine the
Errors
.
•
The controller generates control signals through control function based on the
Errors
.
•
The actuator takes actions to control the plant based on the signal generated by the
controller, which is aimed to reduce the
Errors
.
Since Platform Master communicates with all components and collect information
(distribution of applications' instance, workload of each application's instance, re-
source utilization rate of each VM etc.) of the whole PoI and perform application
scheduling, application resource allocation and VM management (by interfacing with
VM Management of IaaS), Platform Master could be reconstructed to support the
feedback-control loop, thus the resource management framework is mainly imple-
mented in Platform Master, shown in figure 3.
In the framework shown in figure 3, Missed Deadline Ratio and average CPU Uti-
lization are adopted as the Controlled Variables. Because Controlled Variables are
application-independent, a small, non-zero value is used as the Set Point of Missed
Deadline Ratio for each application. The Set Point of Missed Deadline Ratio of appli-
cation j is denoted as
MissRatio
. Similarly, an expected percentage is used as the Set
Point of CPU Utilization for each application. The Set Point of CPU Utilization of
application j is denoted as
U
. Note that the workload of each application in PaaS
is unpredictable, it is impossible to achieve 100% CPU utilization and 0% missed
deadline ratio. Therefore, a tradeoff between these two metrics is inevitable.
