Users and suppliers alike are waking up to the idea that
there is more to networking than adding fatter and fatter pipes.
And this is happening on both sides of the Atlantic, for
once.
I have previously written about the work that the likes of Zeus
in the UK and F5 Networks in the US have been doing on the Layer
4-7 stage, taking server load balancing and network traffic
management into new dimensions.
Two further, and different, approaches to application
optimisation are network attached processing (Nap) from Californian
start-up Azul Systems, and the Etherfabric server card, developed
in Cambridge by Level 5 Networks.
Azul Systems has taken the idea of network attached storage and
turned it on its head, so that a cluster of processors is made
available on the network for whichever application needs processing
power at a particular time. The aim is to improve application
performance and reduce the number of servers required, thereby
eliminating costly and difficult tasks such as capacity planning
and over-provisioning.
At the same time, side benefits accrue, not least a dramatic
reduction in power requirements and datacentre space, according to
Azul's chief technology officer Scott Sellers. This is not
something that should be taken lightly.
According to an IDC survey, the US server installed base alone
will grow by nearly 50% between now and 2009 to more than 14
million servers installed. IDC estimates that the raw electricity
consumption to run these servers will be in excess of £2.8bn per
year by 2009.
How, then, does the Azul Nap work? Basically, it sits at the
application server level, in front of the server pool, forming what
Azul calls a "compute pool". This processing resource is made
available to all applications and can be prioritised in many
different ways to assign processing power to individual
applications as and when they need it. So there is no 1:1
relationship between server processor and storage, added to which,
Nap is operating system agnostic.
The current limitation is that the Azul technology - sold as a
range of rack-mount appliances with between 96 and 384 processors
in each appliance - is purely designed to support Java-based
applications, though support for Microsoft's .net is coming. The
good news is that there are plenty of Java-based applications.
Offloading the processing requirement from any device at any
point on the network has to be a good thing. At the applications
server, this problem becomes critical - a point identified by Level
5 Networks and which is fundamental in its approach to technology
development with its Etherfabric network interface card. Here
bandwidth provides us with few solutions.
TCP/IP may be the de facto networking protocol, but that does
not mean it is actually any good. In truth, it is alarmingly
inefficient in its basic form, generating enormous amounts of
requests and connections for any single application and single
user. Times this by thousands of users and several applications and
the result is a frighteningly high number of TCP connections at any
one point in time.
All the basic elements of a network server are potential
bottlenecks. At low utilisation levels, all the component parts -
unstressed - are able to cope with delivering satisfactory
performance to the user base. The same applies to high-performance
computers. But when the heat is on, as CPU utilisation goes through
the roof, that is when the relationships between these key elements
on a server become strained to breaking point.
As a result, performance suffers heavily, sometimes to the point
of application unavailability or, at the very least, unacceptable
levels of data delivery.
Here, again, network bandwidth is largely irrelevant. So, with
its Etherfabric technology, Level 5 has effectively reinvented the
server network interface card, with a view to optimising server
components such as CPU and memory.
The company claims that, unlike alternative I/O acceleration
techniques (such as Infiniband or TCP Offload) Etherfabric remains
completely compatible with the existing Ethernet/IP network and
application programming interfaces, so users are not required to
compromise standards-based compatibility for higher performance. In
other words, you do not need to change your network and
applications in order to accommodate this new technology.
So how does it do this? One of the primary features of the
Etherfabric technology is that it creates a separate, virtual IP
stack for each application, so providing more CPU cycles per
application. It enables applications to effectively bypass the
operating system and its kernel - seen by Level 5 as a major
bottleneck in network performance.
According to the company, the Etherfabric hardware is designed
to allow multiple applications to access it directly and
simultaneously, and does not depend on the operating system kernel
to mediate between accesses, thereby avoiding that problem area.
This architecture establishes a direct data path between each
application and the network hardware, bypassing the kernel
entirely, so each application sees its own interface to the
hardware.
The problem of scalability is also seemingly dealt with as the
Etherfabric architecture has been designed to support 10gbps and
beyond. Though bandwidth is not the key element to improving
application performance on a network, the need to support the
forthcoming generations of Ethernet is obviously a primary
requirement for any technology development.
The same applies to compliance with existing Ethernet standards.
Forcing users to throw out what they already have in place has
never been a smart or successful play for a supplier, especially in
the long term. Level 5 Networks therefore believes that its 100%
adherence to Ethernet standards is a key differentiator of
Etherfabric, when comparing it with competing technologies.
"Etherfabric remains completely compatible with the existing
Ethernet/IP network and application programming interfaces, so
users are not required to compromise standards-based compatibility
for higher performance," said Craig Easley, director of product
marketing at Level 5.
What we have here then, is the reinvention of Ethernet without
bending any of the existing rules. Interesting.
About Steve Broadhead
Steve Broadhead runs Broadband-Testing Labs, a spin-off from
independent test organisation the NSS Group.
Author of DSL and Metro Ethernet reports, Broadhead is involved
in several projects in the broadband, mobile, network management
and wireless Lan areas, from product testing to service design and
implementation.
www.broadband-testing.co.uk