Although the benefits of virtualising x86 servers have been
pushed relentlessly for the past five years or so, much less
discussed have been the challenges involved in moving to a world
where resources are pooled and everything is linked.
The complexity that such a scenario generates can have a
knock-on effect on issues ranging from infrastructure and licensing
to skills, which means that migrating to the new environment can
end up being an expensive upfront proposition.
Adrian Polley, chief executive at IT services provider Plan-Net,
says, "You are often talking about a complete change in
infrastructure, which is why people who started on this path before
the recession may have continued, but not many have plunged in
since."
A key challenge is that virtualisation involves sharing
resources, whether that relates to hosts, storage or networks, but
changing one element of the whole can have repercussions
elsewhere.
"All of this sharing means that if you give to one thing, you
take away from something else, so it becomes a balancing act to
understand how resources should be properly allocated," Polley
says. "There are always bottlenecks and you can end up just moving
them around. Because things are so interconnected, you can end up
chasing your tail."
As a result, Computer Weekly has come up with a guide to help
you work your way through the mire. Below we look at five of the
biggest challenges relating to x86 server virtualisation and what
you can do about them.
1. Network connections
"If the network is not up to snuff, you are in trouble from the
start. But the bad thing is that, if you have virtualised your
servers without doing your homework, you will not know whether it
is the network that is to blame for performance issues or something
else," says Dan Hidlebaugh, network server manager at Hertford
Regional College.
The educational establishment virtualised its x86 servers about
two years ago in a bid to cut escalating utility bills, reduce its
carbon footprint and improve its disaster recovery provision.
A campus-wide agreement with Microsoft meant that licensing fees
were lower than those of rival vendors. So it agreed to become a
European test site for the supplier's Hyper-V offering, helped by
IBM, which provided the college with a free six-month trial of its
BladeCenters. The organisation has now consolidated its 120
physical servers down to about 55 virtual servers and expects more
to follow.
But Hidlebaugh warns that the success of such projects is not
just dependent on ensuring that the virtualisation software works
effectively.
"You have to look at what hardware you want to use, the storage
area network (San), how you connect the two, how they connect to
the network, how the network reaches the end-user, etc," he says.
"You can have a great virtualisation platform, but if clients
cannot access it due to a network bottleneck, it is useless."
The college had already decided to upgrade its network as part
of a planned move to new premises and undertook a thorough review.
As a result, it introduced an enterprise-class Cisco router, a
dual-band wireless network and 10Gbit network-to-edge switches to
connect the system to users in each classroom. Twelve core fibre
cables were also laid for redundancy purposes and the network was
tested "mercilessly" for a month to push it to its limits.
Another performance consideration, however, related to the
communications backplane of the host.
"We had to ensure that the servers' backplane could handle the
same speeds as the router. If you just throw memory and processing
power at it but are stuck with a 1Gbit network connection, you will
end up with big performance issues," says Hidlebaugh. The
BladeCenters in question have a backplane of 700Gbits.
2. Network storage
A further concern when going down the virtualisation route
relates to storage. Hypervisor suppliers generally recommend
implementing network storage such as Sans for larger production
deployments, particularly if organisations are keen to deploy
high-availability tools such as VMware's VMotion. Direct attached
storage may suffice for smaller development and test environments,
however.
VMotion enables the automatic migration of workloads between
different servers should one crash or need to be taken down for
maintenance. But this activity requires that virtual machines be
stored as disc images in the San. Each host on the network needs to
be able to see each disc image to understand when and where to
assign spare processing capacity should it be required.
But Sans - and personnel with the appropriate skills - are
expensive to acquire, especially if organisations opt for higher
performance fibre channel-based systems rather than cheaper ISCSI
equivalents.
Even if such a system is already in place, it may be necessary
to upgrade it to ensure that performance is adequate and that all
components are certified to run in a virtualised environment, which
is not always the case. Checking suppliers' hardware compatibility
lists is a must, as is following configuration recommendations.
3. Sizing storage capacity
Another must is to size the San adequately, not least to guard
against wasting money by over-provisioning the system. Such a
consideration is also important in light of the fact that some
organisations find their applications run more slowly in the wake
of a virtualisation implementation, despite their use of
server-based memory management techniques such as page sharing.
Hidlebaugh says, "Disc issues tend to be the problem." The
challenge in this context is that virtual machines generate a high
number of I/O requests to be processed each second, but the San's
physical discs may be unable to keep up.
One way of getting around the problem is to use workload
analysis and planning tools such as Novell's Platespin. These tools
evaluate what level of capacity is likely to be required for a
virtualised environment based on the profile of current physical
servers in terms of memory, disc, processor and network bandwidth
usage.
An array that supports mixed workloads can also help.
I/O-intensive applications such as databases and high-throughput
software, such as backup, all appear as a single big workload to
the array despite their different requirements.
But because priority is given to processing big blocks of data,
smaller I/O-based sequential transactions are generally made to
wait, which negatively affects their performance. A system able to
handle both kinds of workloads simultaneously can help to address
the issue, however.
4. Back-up challenges
Many organisations continue to back up their virtualised server
environments in the same way as their physical servers, but this
approach has its downsides. A key challenge relates to the fact
that such activity in a physical environment is often undertaken by
software agents that are installed on host operating systems and
back up both applications and data to either disc or tape.
The problem with doing things this way in a virtual world is
that virtual machines consist of complete logical environments that
include not just the applications and data, but also the VM file
system. Because traditional software does not back up the VM file
system, should the virtual machine go down, it is necessary to
rebuild the file system from scratch. The system must then be
restored and configured and the relevant data and applications
copied over to run on it.
Northern Ireland-based car dealership company Isaac Agnew was
unhappy with the time-consuming nature of this process and so
introduced specialist back-up tools from Veeam.
The organisation initially virtualised a Dell blade server in
the latter half of 2007 to try the technology out, but is now
running 20 virtual machines on three VMware ESX-based machines used
mainly for development and test purposes.
Tim Carter, senior systems administrator at Isaac Agnew, says,
"Before Veeam, we had scripts that one of the team had written to
automatically copy some files from the virtual machine onto one of
the servers that was backed up periodically using CommVault. But we
had to manually choose what to synchronise into the back-up folder,
and if we missed something, we were in trouble."
Backing up each virtual machine would have meant purchasing a
back-up licence for each machine on which they ran, which was
considered too expensive.
But the snapshotting capabilities of the new tools now mean
that, "We can restore the file system in a minute as opposed to
hours of rebuilding the virtual machine and copying files, which
often resulted in staffing having to do overtime in the evenings
and weekends," Carter says.
Although more storage capacity is needed to back up virtual
machines in this way, the compression functionality provided by the
tools mitigates this requirement nicely, he adds.
5. Application support
Although most applications will run in a virtualised
environment, obtaining full support is another matter. There will
be no problem with packages that are certified as
"virtualisation-ready", but some suppliers are unwilling to commit
themselves to this approach either because they have not fully
tested their software on virtualised hosts, or because their
applications have already run into some kind of problem in the
virtualised environment.
Other companies offer a kind of half-way house service in that
users will be requested to reproduce any faults on a physical
server if it is suspected that the issue is associated with the
move to virtualisation.
As a result, Hertford College's Hidlebaugh believes that it is
necessary for organisations to go through "a whole process" to
decide which applications are suitable candidates for migration and
which are not.
"Suppliers of things like domain controllers told us that their
applications were not proven yet and so to please wait. There are
about 30 of our servers that we are not going to virtualise and
about 10 of them relate to applications that have not been tested,"
he says.
"It is crucial to talk to your suppliers and anyone else who is
supporting your applications," Hidlebaugh warns, otherwise you
could end up putting yourself at risk.
He would also be wary about virtualising I/O-intensive
applications such as Hertford College's Microsoft SQL Server
databases and Exchange e-mail servers without heavy amounts of
testing due to San-related performance issues.
Skills
The knock-on effects of moving to a world where everything is
interconnected do not end here. Another important thing to think
about is skills, particularly in large enterprises, where IT staff
tend to specialise in key functional areas such as storage, servers
and networking.
Because all of these areas begin to overlap in the virtualised
world, it is easy to end up in a scenario where support is
duplicated in some areas but falls through the gaps in others. It
is crucial to clearly delineate roles and decide on who is
responsible for what. It may also be necessary to train personnel
across the IT department in new disciplines.
Plan-Net's Polley says, "The skills issue is hard to overstate
because people end up having to have a much greater breadth of
knowledge. They really do need to be expert in a bunch of areas if
they are going to solve problems in a virtualised world
successfully."