Virtualization
Three common misconceptions to overcome
Performance bottlenecks in
virtualized environments can slow
implementations.
by Mike Matchett
While
the economic benefits of IT virtualization,
including lowered capital expense, have been
well touted, over-spending and often
misspending in the virtualized data center
is still common. Often, this is due to
misconceptions about where performance
bottlenecks lie and how to find and solve
them, leading to costly mistakes that limit
the ROI of IT virtualization, or what might
be called the return on virtualization
(ROV). There are several common
virtualization management misconceptions to
avoid.
More resources improve
performance. While 60 percent of
data center performance problems originate
within storage, IT resource managers often
will first invest in more, bigger or faster
servers hoping to solve the issue with brute
computing force. This approach has been one
of the traditional causes of physical server
sprawl.
When storage is identified as a
performance bottleneck, IT often first adds
or allocates more disk space instead of
focusing on adding more spinning disk
spindles. Many storage performance issues
would be most efficiently addressed by
spreading input/output (I/O) across existing
shared disks or onto a greater number of
smaller disks.
Over-allocating unnecessary resources–as
a type of risk insurance to avoid
service-level penalties–leads to
under-utilization in both physical and
virtualized infrastructures. While some
over-allocation to a resource pool can be
recovered by the inherent sharing between
the virtualized clients, over-allocation to
individual virtual clients duplicates the
physical resource sprawl and results in poor
ROV.
Capacity can be determined by
stacking virtual machines. A common
misconception is that physical-to-virtual
server migration is accomplished by stacking
new virtual machines into a server host
until it “hurts,” and then adding to the
infrastructure until it gets “better.” This
performance-by-feel method unnecessarily
creates ongoing performance pain for
virtualized applications and their business
users, and has been the cause of prematurely
curtailed virtualization deployments, and
the loss of any ROV.
Do not virtualize I/O intensive
applications. One of the most
pressing virtualization ROI limitations
comes from deliberately not virtualizing
critical production applications, especially
if the reason is just to avoid any potential
performance impact. There are some
applications that should not be virtualized,
or rather cannot be virtualized as they are
currently implemented, but this group of
applications is shrinking.
Maximizing ROV requires a new “virtually
enabled” approach to IT performance
management and capacity planning. By
implementing the proper “virtually aware”
processes, IT management can provide a
detailed understanding of what resources are
needed when and in what amount, leading to
both optimal utilization of investments and
well-performing applications. Key components
of a maximizing ROV solution include:
- baselining application behavior and
infrastructure workload;
- dynamic performance modeling of each
application’s virtualized logical and
correspondingly allocated physical
infrastructures; and
- analyzing and applying
performance-optimizing recommendations.
In order to manage both application
performance and infrastructure optimization,
each application’s behavior should be
collected and profiled. This includes both
static attributes like its importance and
service thresholds, and its dynamic activity
over time.
The key to producing a good performance
model is observing each application’s actual
physical infrastructure requirements are
utilized, including how much real CPU or how
many actual I/Os. Traditional IT management
solutions often become confused by system
metrics that have become virtualized, like a
virtual machine operating system’s metric
for CPU utilization. Since virtually enabled
solutions should account for the underlying
physical performance dynamics of shared
resource pools, their data collection must
span both virtual and physical resource
perspectives.
In optimizing the whole of the IT
investment infrastructure, as well as
identifying the root cause of specific
performance bottlenecks across these
domains, being able to map logical data
paths across each virtualized layer and IT
domain becomes important, down to the
underlying shared physical resources, with
applications, servers, and storage fully
mapped together.
Mapping an application’s logical data
flow enables IT management to see which
actual resources are being used by an
application. An application’s performance in
nested virtual IT layers can be properly and
automatically analyzed by applying dynamic
queuing models built over the observed
application and infrastructure data.
Cross-domain performance models can
immediately show where to find and how to
remediate performance-constraining
bottlenecks, judge and produce optimizing
recommendations for aligning resources to
applications, and help plan future what-if
scenarios. The immediate results enable IT
to virtualize more performance-critical
applications, not just the low hanging fruit
of development and test applications. By
managing to real application requirements
and actual infrastructure performance,
applications owners get insight and build
trust in IT infrastructure owners that are
hard-pressed to roll out optimally shared
resource pools with virtualization
technologies.
Since virtual environments are, by
definition, highly dynamic in how resources
are allocated to applications, both data
collection and modeling analysis needs to be
automated. System-management tools often
come pre-loaded and running in virtual
appliances.
Longer-term results from dynamic
virtually aware performance management and
capacity planning include investing in the
right resources at the right time.
Appliance-based solutions that are always
analyzing current operations also produce
ways to measure both the effectiveness and
efficiency of IT investment decisions. Not
only can the return on current
virtualization projects be maximized, but
key business-savvy intelligence can be
created for making future IT investment
decisions.
Michael Matchett is director of
product management for Akorri,
Littleton, Mass.
For more information
(click here)