Virtualizing onto mainframes: Analyzing workloads to determine fit

by Andrew Hillier

May 28, 2008 —

As virtualization makes its way deeper and deeper into the data center, organizations
are starting to leverage virtualization not only for the flexibility, efficiency,
and reduced cost of ownership it provides but also because it lifts many of
the constraints that govern which platform an application needs to run on. Different
types of applications possess different workload “personalities” and
these heavily influence how well an application will perform on a given virtualization
model.

Following are some tips for identifying those workloads that are best suited for consolidation onto mainframes and for analyzing
existing IT environments to determine the optimal approach for consolidating
workloads onto the mainframe platform.

Analyzing workloads onto mainframes

Partitioning, hypervisors, containers, and other approaches can all be used
to allow workloads to co-reside on a common platform. Given the relative
strengths and weaknesses among platform classes, considerable diligence is required
when consolidating workloads between them. This involves understanding which
applications will run on the mainframe, which applications will experience the
most benefit by moving, and what they will look like after the transition. In
practice, it is even more important to understand the subtlety of such a transformation,
particularly when the overall goals are to optimize the benefit while minimizing
the risk.

By modeling transformations in terms of the constraints that govern them, it
is possible to chart a course that reaches the end goal. These constraints can
be classified into three categories:

1. Technical constraints – what can go together

2. Business and process constraints – what should go together

3. Workload analysis – what fits together

Technical constraint analysis

Technical constraint analysis generally deals with compatibilities and affinities
between hardware and software components. What can run on what, what talks to
what, etc. For mainframes, this step is concerned with what applications
will work on the mainframe, which will accrue the most benefit by moving, and
how they should be organized to optimally leverage these benefits.

Hardware and software compatibility - In addition
to understanding which applications talk to one another and which have an affinity
or aversion to one another, it is also important to understand how the applications
and their usage drive middleware to behave. Likewise, any cross-platform transition
must take into account whether the source systems employ any specialty hardware
such as token rings, faxes, and USB devices that may be difficult to move with
them. If any of these things influence how the application is configured in
the mainframe environment, it should be ruled out as a consolidation candidate.

Network connectivity and latency - Any shifts in the topological
location of an application may introduce changes in communication latency that
can impact application performance. If the changes introduce latency in the
wrong places, such as between application servers and databases, performance
will suffer. This is why targeting applications whose databases are already
on the mainframe is a good initial strategy.

Application connectivity - Rather than just looking for applications
that already talk to the mainframe, organizations should look for applications
that talk to each other. By consolidating onto the same system, they can realize
tremendous benefits by promoting “crosstalk” that doesn’t ever
hit the network. A note of caution though: this strategy can sometimes bypass
important network security controls, so collapsing application tiers and communication
paths should be approached carefully.

Business and process constraint analysis

These constraints help organizations determine what can and can't be done from
a non-technical perspective due to regulatory requirements, internal politics,
or other real-world considerations.

Process-oriented constraints - The proper functioning of production
IT environments often relies on tight process controls. Change freezes, maintenance
windows and other controls must be respected, and any transformation of IT environments
must take into account the rules that have been put in place.

This is good news for mainframe consolidation, as the mainframe platform has
traditionally been subject to a level of process rigor that is only now making
its way the data center.

Security & regulatory compliance - To protect intellectual
property and other sensitive information, many environments segregate data types
and apply security and access control policies accordingly. Businesses should
avoid virtualization scenarios that may cause any security vulnerabilities.
Even if the security model of the mainframe supports policies that mirror the
distributed environment, it may be more practical to try simpler approaches
that don’t cross security zones and introduce potential regulatory or security
challenges.

General business constraints - Organizations are often resistant
to sharing infrastructure between departments, partly due to the lack of chargeback
models that enable cross-department hosting. Another issue that may impact virtualization
is operational environments. Combining production and non-production environments
on a shared infrastructure may not be desirable. Placing the unpredictable patterns
of development and testing on the same system as critical production workloads
requires a great deal of finesse, particularly with respect to workload management.

Workload analysis

One of the most challenging aspects of analyzing mainframe virtualization is
the proper normalization of system utilization patterns. Because of the diverse
workload personalities present in the data center, and the varying characteristics
of the platforms themselves, getting the right answer is no easy feat. CPU utilization
must be properly normalized and must take into account I/O rates, memory utilization,
and context switching to ensure accurate results.

To compound this problem, many published benchmarks tend to steer clear of
the high bandwidth, high synchronization segment of the workload population,
necessitating the creation of specialized benchmarks that accurately capture
the relative merits of the different platform classes. To put this in
a different way, using standard integer benchmarks to normalize workloads between
midrange and mainframe platforms simply will not give the right answer.

This problem is not unique to the mainframe --many virtualization technologies
are plagued by the fact that proper benchmarks are either non-existent or only
just emerging. In any case, what is required is a flexible benchmarking
approach that allows different workloads to be normalized independently based
on their specific personalities and the relative performance of the target platforms
with respect to that personality.

CPU Normalization and Benchmarking - Although the theory behind
what makes a workload run well on a mainframe can be daunting, the practical
approach to getting the answer is a little more straightforward. By employing
multiple benchmarks that represent the relative performance of source and target
systems for a given workload personality, it is possible to construct a single
analysis that optimizes utilization across multiple source/target combinations
in a single step. Each source workload is then able to be normalized independently
based on its distinct personality, and the target endpoints (mainframe partitions)
filled to capacity based on a reasonably accurate projected utilization. Although
the advantages of this approach are numerous, the challenge becomes the determination
of exactly what the personality is for each source workload. Fortunately
there are some key indicators that can be used to separate out the different
types of activity.

I/O Analysis - Although one of the strengths of the mainframe
platform is its I/O performance, it is still important to analyze the I/O characteristics
of the source workloads. In any consolidation scenario it is desirable to combine
workloads that dovetail in a way that promotes a balanced use of resources.
Because of this, I/O constraints should be analyzed in parallel with CPU and
other metrics in order to give a balanced result that makes optimal use of the
target resources.

Memory Analysis - Memory is often a deciding factor when determining
how many applications can be virtualized onto a single system. Yet this can
be tricky to analyze, as the truly memory requirements of an application are
not always made evident by the physical memory in use on the system. Using the
measured memory utilization to determine what can fit on a target will provide
a safe answer but might require the use of more memory than is actually required.

Conclusion

Ultimately, all of these constraints must be analyzed together to determine
the optimal paths to virtualization. When considering this array of constraints
against a series of source and target servers, the analysis becomes a three-dimensional
optimization problem.

Consolidating application workloads onto the mainframe platform is a great
strategy to consider in many IT environments. By properly assessing the suitability
of applications to consolidate, the results they will give, and the overall
TCO of the solution, it is possible to uncover significant opportunities to
simplify IT infrastructure, improve reliability, increase resilience, decrease
power consumption and ultimately drive down the costs associated with servicing
these workloads.

CiRBA