The Threshold rule is mainly focused on user

The
threshold-based rule is an auto-scaling approach that used across many
platforms such as, , Amazon Web Servers and Microsoft Azure. The Threshold rule
is mainly focused on user defined metrics that analyse the network performance,
such as, CPU Utilization and Average response time etc. Normally there will be
two rules or policies, for scaling up or scaling down virtual machines, to
either decrease or increase the resource pool. Within each rule, there are many
parameters that need to be configured by the user, the upper (thrUP) and lower
threshold (thrDown) which in this approach, would be the limits for a certain
performance metric. This is when a user can set policies which can trigger the
auto-scaling operations if they are met, an example that could be used on an
AWS EC2 Instance, “If CPU Load >70%, for a duration more than 2 minutes,
create new EC2 Instance”.

1.1.1.1      
Review of proposals

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

The
advantages of the Threshold-based rules is the simplicity required with how you
want your network to operate. The policies are very straight forward and makes
it very user friendly, especially with you having full control in which
resources are assigned to different applications 1. However, the fact it is
very user based requires deep understanding to maximise efficiency of workloads
and resources as the user would set numerous upper and lower thresholds for
several parameters. Those parameters could vary from CPU utilization or average
latency of packets; in which the user would need to pinpoint which metric they
want the auto-scaling to take place and when.

A
different approach created by Right Scale 2 is the Right scale voting system.
It is a voting system where each VM has the ability to vote based on its
current load, within it containing two thresholds, upper threshold to scale up
or a lower threshold to down the whole platform. This is a democratic system
where the algorithm takes in the majority vote, 51% of the overall votes,
before it takes action. Right scale allocates voting tags to the servers you
want to participate in the vote, creating a server array which helps with the
overall management of the network. After each scaling action, Rightscale
integrate a 15 minute cooling period, similar to Dutreilh et al 3 in where
the idea of cooldown periods to avoid oscillations in the system, these are
called “Resize calm”. This cool time period is also integrated because of
the fact VM normally take around 10 minutes to start 4.

The voting
system from Right scale’s algorithm still poses the same disadvantages as a
threshold-based rule approach, both approaches are user-defined and are highly
dependent on the user deciding upper and lower thresholds. A proposed tool by
Simmoms et al 5 with the “Strategy Tree”, where he explained that the purpose
of this framework was to address the vulnerability of the current approaches,
where the strategy tool was to be used as a framework for reasoning the
effectiveness of the different strategies, as these strategies are often based
on assumptions, predictions and expectations within the polices that they set.
The strategy tree allows the evaluation of effectiveness of a deployed rule set
to be simplified and the ability to switch strategic approaches into
alternatives using a systematic and hierarchical manner 5. To further improve
effectiveness of strategic approaches, Kupferman et al 6,7 came up with the
idea that you should never terminate a VM even if the load is low before the
hour is over, since you get charged for full hours even if use usage was for a
partial hour.

1.1.2       
Reinforcement Technique

An
alternative approach in auto-scaling is the Reinforcement Technique (figure 1)
which is focused on primarily the agent and its environment. The decision maker
which is the agent, focuses on the best decision (action) to improve the
network environment (state), and in reward it will have the new and improved
current state (reward).

The agent
will be the one interacting with the scalable applications, it will decide when
to scale up or down with its (actions), depending on the network performance
for example, CPU load, response time (state) and it will adapt to the network
performance by adjusting the resource pool, in which the adjusted network is
the reward. This approach is seen as memoryless as future states are only
affected by the current state and action, regardless of the any previous
states.

Below are
the meanings behind the symbols used in the Markov Decision Process (MDP),
which is the typical framework used in decision making in reinforced learning
scenarios. 

·        
S,
is represented as the environmental state, the current network performance

 

·        
A,
represents the action taken place

 

·        
R,
represents the reward which is the current state plus the addition of the
actions to improve it. State + Action

Figure 1

1.1.2.1      
Review of proposals

Due to the
RL technique being defined by three elements, the action A, the state S and
reward R, to apply it to auto-scaling. Dutreilh et al (2011) suggested the concept of  (w,u,p), where w is the total number of user
requests , u is the number of virtual machines allocated to the application and
p is the performance in terms of the average response time. This would allow
the integration of RL in auto-scaling. However, although RL is a possible
auto-scaling technique, there are several problems with it that I believe would
put it beneath a more superior technique such as, The Threshold based rule.

–         
Poor
initial performance. For the algorithm to be optimised and fulfil its potential
it would need a large considerable time to understand the different states and
actions, would take a long training period. Before it explores the environment
and a good solution is found, the performance of the RL would be inaccurate and
unreliable for the responsibility of sensitive decisions. 42 99 92

–         
Adding
onto the previous point, the RL would have to explore the different states and
actions for it to be able to find a workable solution. Therefore, if there are
changes in environmental conditions there would need to be a re-adaption in the
policies of the RL technique.

–         
One
of the main problems of the RL technique is known as “The curse of
dimensionality problem”, which is having a congested state-space. There is
always a change in the current state, growing the number of state variables,
which a lookup table is use to store a separate value for each possible
state-action. This is not scalable as the look-up table would grow exponentially
with the number of state variables, making the look-up table very slow, and
having longer times to access the table and making updates to it.

The last
point is why I believe the threshold based technique is better than the RL, as
it is not restricted to scalability and especially in cloud computing, where elasticity
is key, the RL would fail to be an efficient technique.

x

Hi!
I'm Morris!

Would you like to get a custom essay? How about receiving a customized one?

Check it out