Virtualisation systems aim to reallocate resources according to
shifting business needs, maximising CPU usage and saving companies
time and money. But if the systems are to work, separate
departments must be prepared to pool their computing resources,
writes Danny Bradbury.
When the open systems movement started at the turn of the 1990s,
many people thought it would usher in a new era of
interoperability. Industry standards would level the playing field,
they predicted, making it possible to plug anything into anything
and have it work with only the minimum of fuss.
But while systems from different suppliers can be made to work
together in a way that was unfeasible in the supplier-dependent
era, it is still far from easy. Managing the systems is even more
difficult, because IT experts must accommodate the idiosyncrasies
of each platform and device.
Suppliers hope to solve that problem by introducing a new product
category: systems virtualisation. In a virtualised system, a layer
of middleware is introduced between the systems administrator and
the disparate resources that make up the infrastructure; the
administrator gives instructions to the middleware, asking it for
specific resources as and when the need arises. The middleware
deals with the necessary devices, reallocating storage space, CPU
power and other necessary elements under the covers.
Mainframes have always used virtualisation techniques to ensure
that the maximum resource is available to the business at any given
time. The virtualisation products now starting to enter the
non-mainframe market can be broken down into three components:
storage, processing power and network resources. Many companies are
concentrating solely on the storage part of the story, while
others, such as Canadian firm Inkra, are focusing on networking.
Few firms are attempting to bring all three components together;
but there are some, Sun Microsystems, Hewlett-Packard and IBM among
them.
One of the main benefits of bringing systems virtualisation to the
non-mainframe market is economic. With many CPUs and storage
systems under-used, it makes sense to reallocate resources
dynamically according to the need of particular applications or
lines of business. If, for example, a technical support application
is barely ticking over, some of its CPU time could be given to an
accounting application being heavily used by a finance department
coming up to its reporting period. Should the situation reverse,
however, the CPU allocation would need to be adjusted yet again.
Doing this manually would be too time-consuming.
Theoretically, using system resources more efficiently in this way
could help systems administrators squeeze more applications on to
less equipment.
John Mills, technical marketing manager at Sun Microsystems,
stresses the value of "sweating your assets" this way. Using the
company's N1 virtualisation initiative - announced last February
but so far unaccompanied by any products - he hopes to increase
average CPU usage to 85% from a figure that can, he says, be as low
as 15%. HP puts the average CPU usage figure at 35%.
There are three main phases to the N1 strategy. The first,
originally scheduled for late last year but now likely to happen in
the first quarter of 2003, will enable customers to marshall
various parts of their systems infrastructure into a single set of
resources. The second phase, also due to start this year, will
enable systems managers to hook business processes into the
front-end middleware. Sun points to electronic banking as an
example of a business service that can be described as a single
entity to the N1 infrastructure, and served accordingly.
The third phase of the initiative focuses on automating service-led
systems management using policies. Business-level policies can be
set to govern the allocation of resources to specific business
services. A practical example might be switching priorities for
different services to customers depending on the system load - so,
an investment bank providing free quotes to visitors on its website
as a means of generating business might want to throttle back the
CPU time allocated to such quotes as the number of transactions
initiated by paying customers increases.
Blade servers figure heavily in discussions about systems
virtualisation. Blade servers, which began in the
telecommunications sector, are relatively new to the mid-range
server market. Multiple servers on "blades" are slotted into a very
dense chassis, creating a low-footprint array of servers, connected
using a high-bandwidth internal backplane. Sun Microsystems is
scheduled to ship its blade servers early this year, following a
delay to the original ship date.
Egenera, a relatively young company specialising in virtualisation,
uses blades as a pivotal point in its product offering. The company
developed its blade server and associated Processing Area Network
(Pan) Manager virtualisation software from scratch, which gives it
an edge over larger players, according to European services
director John Warnants. He believes he can fit up to 1,000
traditional Intel servers and redeploy them into two of his
Bladeframe servers, taking up one square metre.
The virtualisation part of the system works by rendering blades
stateless, enabling the Pan Manager to dynamically allocate
everything from Mac addresses to storage connections and operating
system images on the fly. All blades assume that resources are
local, but they are actually provided over a fast-switched
fabric.
The case for virtualisation sounds unassailable but, in practice,
there are challenges both suppliers and customers need to consider.
Perhaps one of the most significant issues is interoperability.
Because the product category is so new, no standards have yet been
defined for connecting different suppliers' equipment with
virtualisation middleware. Consequently, Mills says his company
will open up the N1 software infrastructure to third-party
suppliers, enabling them to hook into it.
Egenera's Warnants is less bullish. "Once we've built one of these
virtual servers, it looks like just another server on the network,"
he says. "But what we don't do is use our management software to
virtualise non-Egenera hardware. That's another generation of this
technology that is hugely more complex."
Standards are, at least, being developed in the storage
virtualisation sector. Last August, the Storage Networking Industry
Association launched the Storage Management Initiative in a bid to
turn a specification submitted by several suppliers into a
universal standard for storage virtualisation middleware.
The standard, called Bluefin, was submitted by companies including
Sun, EMC, Dell Computer and Hitachi Data Systems and will go some
way towards easing the concerns of customers. But companies buying
virtualisation products in the processor sharing or network
resource space will find themselves on their own when it comes to
making decisions about interoperability. Everything depends on the
supplier's own APIs and its ability to get other suppliers to work
with its system, or to write its own drivers for other companies'
devices.
And for customers wanting to virtualise their systems, there are
two more, related problems - internal accounting and politics.
"Server hugging" by particular departments unwilling to free up
their own computing resources could create problems for companies
wanting to implement virtualisation software. Business managers may
have to be convinced that security and reliability will not suffer
as a result of throwing computing resources into a central pool.
One way to get around this is by hard-partitioning certain critical
applications.
But even if such resistance is overcome, internal accounting
procedures may present a challenge. IT departments often need to
log the use of computing resources so they can be charged back to
particular departments. Virtualisation systems need to support
these customers.
Most virtualisation suppliers are aware of the internal accounting
challenge and have facilities for logging resource usage. HP, which
has been selling system virtualisation software since March 2002,
also provides this facility. Its Utility Data Centre (UDC), first
announced in November 2001, stems from HP's historical goal of
providing computing power on tap according to variable customer
need. The UDC is interoperable with equipment from multiple
suppliers thanks to HP's internal driver development effort, and HP
is also about to publish its own driver software development kit,
enabling customers or third-party suppliers to write their own
device drivers for the UDC this year.
Bringing up the rear is IBM, which is still researching its
ambitious autonomic computing initiative, which mixes system
virtualisation facilities with software that uses biological
computing algorithms to create self-governing infrastructures. In
addition to redirecting workloads to different sets of system
resources, the company wants to make the interface between the user
and the computing system easier. It wants to abstract the users'
interface with the system even further than its competitors,
letting business managers instruct the back-end infrastructure in
plain English.
In the company's ideal world, a user of the IBM system would be
able to tell a computer to "watch my competitors and adjust pricing
and supply for competitive advantage".
For now, however, even the suppliers of the less ambitious
virtualisation systems have their work cut out. IT departments need
to be convinced that the systems will work reliably. The danger is
that if you abstract your systems management through a piece of
smart middleware and the middleware fails, you could end up with
more administrative overhead than you started with.
Given the relative immaturity of many virtualisation systems on the
market today, it will be a while before the growth curve for such
systems gets steep enough to contribute significantly to most
suppliers' bottom lines.
Sanger Institute goes virtual
Many hands make light work, according to Phil Butcher, head of
IT at the Wellcome Trust Sanger Institute. The institute has
conducted some of the most significant research into the human
genome to date, but to do so it has required heavy computing
resources. The institute maintains a network of 1,100 server
nodes, each of which can be used to help process number-crunching
jobs submitted by end-users.
Server virtualisation was vital to make the most efficient use
of the computing infrastructure and minimise the administration
overhead, according to Butcher.
He has been using Platform Computing's LSF workload management
software to help him administrate his systems. "We have gone from
clustering to distributed resource management to virtualisation,
but the trick's the same," he says. "You want to ensure that you
put your workload on to a number of systems because that is why you
get the efficiencies."
Using the Platform software, Butcher distributes computing jobs
automatically to servers across the network. However, he stops
short of fragmenting a single job across multiple nodes, which is a
fundamental tenet of grid computing, a technology using
virtualisation techniques. He does fragment databases so he can
distribute relevant parts of them across the 1,100 node farm for
pattern matching purposes.
Policy-based administration of his virtual computing resources
is a key part of Butcher's strategy. "If you have a facility for
people to put 10,000 jobs on the system, you can soon end up
creating your own denial of service," he says. The policy
facilities within the Platform software enable him to set rules so
that end-users get a fair share of the resources.
Although internal accounting wasn't on the Institute's agenda,
Butcher nevertheless keeps logs of how much CPU time has been used
by each group, because he needs to allocate resources according to
the importance of each scientific project.
What is virtualisation?
In a virtualised system, a layer of middleware is introduced
between the systems administrator and the disparate resources that
make up the infrastructure; the administrator gives instructions to
the middleware, asking it for specific resources as and when the
need arises. The middleware deals with the necessary devices,
reallocating storage space, CPU power and other necessary elements
under the covers. The virtualisation products now starting to
enter the non-mainframe market can be broken down into three
components: storage, processing power and network resources.