Hardware Management - Deploying and Managing a Cloud Infrastructure - page 229

Another memory allocation technique is called transparent page sharing (TPS). A host
runs multiple virtual machines that share the same profile, such as the same OS, same hard-
ware specs, and the same workflows. Because of these similarities, there is a fair amount
of identical memory pages that can be shared by multiple virtual machines so that each of
them does not have to keep a copy; they could just access the same page while the hypervi-
sor reclaims the other identical pages. The performance of this method is determined by the
page scan rate and opportunities for shared pages.
Central Processing Unit (CPU)
The CPU, or in this case the vCPU, is at the heart of the VM instance and can be consid-
ered another core resource. Relying on knowing the exact number of vCPUs required for
your VM's workload is not a straightforward approach, and there really is no way to be
exact. You need to determine the amount of CPU time required by a virtual machine, let's
say a web server, to provide satisfactory response times for the end user of the application.
This amount totally depends on the workload that needs to be processed by the web server.
But since workloads and traffic on the Internet significantly differ from time to time, you
cannot simply determine the resource requirement at the peak rate because that amount of
resources will be underused during off-peak times.
The amount of CPU resource that should be allocated to a virtual machine must depend
on two factors: the achievable performance with a given resource allocation and the associ-
ated resource cost to the user. This amount of resource can be determined by creating an
allocation model that establishes the relationship between the amount of CPU resource and
its corresponding level of performance, such as comparing the response time to the amount of
CPU resource. This model will then be used to determine the amount of resource to be allo-
cated to a VM based on the desired performance level. However, it cannot be entirely accu-
rate because of the complexity brought by multiple layers from the hypervisor, the guest OS,
and the subsequent running applications, so there has to be additional checking. There must
be a feedback loop wherein the allocation is fine-tuned based on the performance feedback of
the virtual machine, so any inaccuracies in the model can be compensated by feedback and
adjustment cycles. For example, if a virtual server is supposed to cater to a thousand users at
once, and based on the model, the appropriate amount of computing resource is given to it.
If feedback then indicates that it is not performing satisfactorily, additional resources have to
be allocated; less resources should be allocated if the server is performing above expectations.
This sort of dynamic allocation allows us to provision lean machines that work satisfactorily
as intended, not allocating too much or too little.
Storage and Network Allocation
Storage is easier to allocate because it is the most abundant and easy to scale out of all
the resources, but you still have to be smart and lean with allocation since it is still finite.
During provisioning of a VM instance, you can choose to create a static or dynamic virtual
disk. A static disk's size is no longer expandable, while a dynamic disk will scale along with
the growth of content on the VM. Most providers will set limits in the number of vCPUs
available for each virtual machine for each product set offering. If the customer wants