Information Technology Reference
In-Depth Information
Fig. 2.
Resource Management Framework
The sampling time is defined to be the end of each period of the arrival of the feed-
back data. Resource management decisions are made at the beginning of each period
based on the feedback data collected previously. Then
error
for application j could
be calculated with the following formula (4):
error
w
×MissRatio
MissRatio
w
×U
4
U
is the Missed Deadline Ratio in
the
k
sampling instant and Coefficients
w
and
w
are tunable. Based on
error
Where k is the sampling instant and
MissRatio
,
the number of instances of application j to be changed can be calculated in the current
period with the following PID control formula (5):
C
D
×
error
error
DW
C
I
× error
∆Instance
C
×error
5
DW
IW
Where
C
,
C
I
,
C
D
, IW and DW are tunable coefficients. The number of instances
of application j can be changed according to
∆Instance
.
∆Instance
>0
means that
the application should scale up (i.e., the number of instances of application j should be
increased), and
∆Instance
<0
means that the application should scale down (i.e.,
the number of instances of application j should be decreased), and
∆Instance
0
means that current number of instances of application j is appropriate and does not
need change.
4.3
VM Scheduling Policy
How to set up a feedback control system in PoI and how to determine the number of
instances of each application in order to guarantee the QoS of each application are
stated above, then the policy to schedule the rented VMs will be illustrated. The deci-
sions on when to scale up and scale down are made on the objective: Low
Time Usage of VMs. Specifically, as the ∆Instance has been computed, strategies of
