Cluster and grid computing has become extremely popular, yet very few
designers are using modern capacity planning techniques to ensure
Cooking on the Grid
Grid computing has become extremely popular in IT circles mainly because
of the potential computing power and cost savings. But grid computing is
a multifaceted technology that means different things to different people.
Some interpret grids to be a heterogeneous group of desktops and servers,
which other see it as a group of cluster computers connected together
over the internet. Sometimes grids are designed for raw CPU power,
others are designed for raw I/O, while others are a combination of data
and compute power. Many grids are often designed around a specific set of
applications. As we'll see later, this is why grid and cluster design
is strongly tied to its intended purpose.
Yet despite the importance of strategic design of grids and clusters,
very few commercial integrators spend the time or capital to ensure the
system's feasibility and performance. The reason why is many fold;
lack of experience, cost issues, ego, or lack of skill resources.
An important exception to this is the scientific community whose grids and
clusters are often well designed.
A Stitch in Design Saves Nine
So what is really involved in cluster/grid design and planning? First of all,
not every application is suited for a cluster and/or grid environment.
Customers who switch from an SMP environment to clusters with the intension to
someday get into the intra-grid domain are supposed to conduct a feasibility
study prior to getting too deep into the cluster business.
Second, the customer has to
fully understand the current workload behavior, and has to be able to
formulate the goals that have to be achieved in a cluster/grid environment.
Modeling based sensitivity studies allow the customer to compare (from a
relative perspective) design alternatives, and to zoom in on the setup that
is most feasible for the environment. As a modeling based approach is
recommended at this stage, no money has to be spend yet on any hardware
components. In a nutshell, conducting a comprehensive feasibility and
design study early on in any cluster/grid project safes the customer
substantial money, and replaces the common guessing game with a very
pragmatic approach to systems engineering that leads to stable environment
with a high acceptance rate from the user community.
Grid/intra-grid/cluster planning and design studies are fundmental parts of the
implementation process, having the greatest impact when designed into the final
project, encompassing application (workload), network/interconnects, OS, I0,
memory, and CPU subsystems, respectively.
In almost all circumstances, companies can recoup the design costs in the long
run, as fewer firefighting nightmares are necessary.
Fortuitous Technologies provides comprehensive performance, planning, and
design services based on solid mathematical and statistical methods. They can be contacted
at http://Fortuitous.com .