Power Management
Unified fabrics reduce power use
InfiniBand enables high-speed data
transfers and is suitable for
virtualization.
by Yaron Haviv
An InfiniBand switch can consolidate and
virtualize multiple networks onto a single
high-performance unified fabric, reducing
costs and complexity.
With the growing demand for greater
computing and storage capacity in the data
center, one of the most important and
challenging problems IT faces is how to
address escalating power requirements. To
address the power challenge, data center
architects should look at every data center
component and every new application
environment through the lens of power
efficiency metrics.
The goal is to maximize the power
efficiency of the application. This implies
maximizing the performance of the
application while minimizing the power of
the infrastructure (e.g., server, storage
and networking) supporting that application.
Frequently, application performance is
bottlenecked by inefficient network
connectivity, causing an immediate impact to
overall power efficiency.
The return on a power-efficient data
center fabric infrastructure using currently
available technology can be in the millions
of dollars. By combining various networks
into a data center unified fabric, IT can
significantly reduce overall power
consumption and increase the power
efficiency of the data center.
Unified fabrics provide seamless,
high-performance networking services between
InfiniBand fabrics, Fibre Channel
storage-area networks (SANs) and Ethernet
LANs over a single high-performance fabric,
with multiple virtual interfaces replacing
physical adapters. By consolidating multiple
data center fabrics into one, enterprises
save costs, reduce complexity and enable
further consolidation and virtualization of
the data center.
Unified fabrics eliminate the use of
separate networks for LAN, server-to-server
connectivity and storage. Instead, they
combine them onto a single InfiniBand-based
network that runs all of these traffic types
in parallel lanes. The single InfiniBand
fabric enables connectivity from every
server to every resource and eliminates the
need for additional infrastructure.
InfiniBand is an I/O technology that
enables high-speed data transfers and
ultra-low latencies for computing and
storage over a single fabric. It currently
supports 10-gigbit per second and
20-gigabits per second (Gbps) host
connectivity, with application latency of 1
microsecond end-to-end. InfiniBand also has
built-in mechanisms for memory-to-memory
transactions, enabling one computer to read
the other’s memory in a secured fashion.
InfiniBand is also suited for
virtualization because it allows built-in
network segmentation, multiple network
layers (virtual lanes) per link,
hardware-based high availability, and
multiple virtual I/O adapters (channels) per
host card. This allows dynamic formation of
multiple virtual networks on a single
fabric, without compromising performance,
security or availability.
In order to address data center
scalability, InfiniBand includes features
such as L2 congestion management;
class-based traffic isolation; multipath;
and central management and provisioning. By
implementing a unified fabric based on
InfiniBand, a data center can gain improved
power efficiency and consolidate server I/O,
which creates cost savings and improved
performance through eliminating I/O
bottlenecks.
When defining the power efficiency of an
element within a network, the goal is to
minimize the power each network port
consumes. With 10-Gbps and 20-Gbps
InfiniBand, less than 5 watts of power per
port are used, making it more energy
efficient than 10-Gigabit Ethernet.
Using a high-performance network can
affect application performance and
utilization dramatically. The more
efficiently an application runs, the smaller
the footprint of the data center. The power
efficiency of a data center is determined
not only by how efficient the data center
components are, but also by how many
applications can be run or how many
transactions can be delivered per second.
Even if data center managers choose
power-efficient components, these components
may deliver significantly slower application
performance.
While compute capacity grows
exponentially, Ethernet network capacity has
not managed to keep up at the same pace. In
addition, the increased server utilization
due to server virtualization increases the
load on the network, leading to critical I/O
bottlenecks that significantly degrade
application performance and scalability
In traditional data centers, each server
has multiple redundant I/O adapters to IP
networks, clustering and storage. InfiniBand
can be used to consolidate server I/O by
delivering multiple virtual I/O adapters
over a single unified adapter and multiple
networks over a single fabric architecture
that can provision multiple isolated virtual
lanes, hardware partitioning, quality of
service and high availability. This leads to
better power efficiency when taking into
account the alternative of using multiple
switches and adapters.
In addition to power efficiency
improvements and the indirect influence that
networks and I/O can have on application
power efficiencies, virtualization
technologies also provide additional power
savings. Virtualization technologies allow
the consolidation of multiple applications
onto fewer systems and less infrastructure,
which saves on the power consumed by the
unneeded infrastructure.
Implementing server virtualization
without providing hardware-based I/O
virtualization in parallel, however, will
slow down the applications running over
those servers, again requiring more
resources and reducing the power efficiency.
Fabric and I/O virtualization are key
elements that allow the solution to achieve
full system virtualization, save power and
offer additional benefits including: a
reduction of the number of LAN and SAN
switches in a configuration down to a single
switch; a reduction of the number of network
interface cards and storage adapters down to
one per server, which also eliminates the
need for large and power-hungry servers with
many I/O slots; improved performance of
applications and server virtualization
software; and the ability to prioritize the
I/O and fabric resources and deliver them to
the right applications, eliminating
potential I/O bottlenecks and improving
application performance and utilization.
Yaron Haviv is the chief technology
officer at Voltaire, Billerica, Mass.
For more information
(click here)
Test for future speeds
by Peter Schweiger
The low cost and high
performance of Gigabit Ethernet, or 1,000
Mbps, has caused the copper
structured-cabling market to now only
recognize Category 5e and higher cabling,
which is the minimum necessary to support
Gigabit Ethernet speeds. More than 90
percent of today’s structured-cabling
installations consist of unshielded twisted
pair (UTP) Category 5e and Category 6
cabling. Testing copper cabling for its
ability to handle Gigabit Ethernet requires
test parameters to be measured across a
frequency range of 1 MHz to 100 MHz.
Companies that install
structured cabling to enable telephony, data
and now security communication within a
building expect this investment to support
those applications for at least five to 10
years. These companies can realize longer
cable life if they test copper cabling for
the highest possible speeds during
installation. This means testing cabling for
its ability to handle 10-Gigabit Ethernet
(10GigE) traffic to ensure the copper
cabling will have the longest service life
possible.
Category 5e cabling was
specifically designed for Gigabit Ethernet
transmission. To certify an installation,
tests should be done to a maximum frequency
of 100 MHz. Category 6 testing has better
electrical performance and is certified to
250 MHz, although no popular application
uses that bandwidth. 10GigE requires each
copper pair be characterized to at least 500
MHz and also has an additional specification
for the susceptibility of each pair to
crosstalk generated by pairs in adjacent
cables (alien crosstalk).
The latest generation of
cable certification testers can test
installed cabling to the 500 MHz necessary
to confirm 10GigE will run on that cabling.
Some can even go to 1,000 MHz for future
standards like CAT 7a. These testers can
conduct these tests without additional time
that would increase installation costs and
can also measure alien crosstalk.
Certifying cabling to
higher frequencies requires using the
correct TIA standard. If the cabling is
installed correctly and distance limits are
observed, most cables will pass internal
tests for running 10GigE transmission.
Because signals higher than 300 Mhz begin to
propagate beyond the cable to affect
adjacent cables, this alien crosstalk
interference is often the limiting factor in
running 10GigE on cabling.
Mitigation techniques are
listed in TIA-TSB-155 to help address these
issues, including using alternate patch
panel positions for 10GigE links, upgrading
connector hardware to match the application
and loosely bundling smaller groups of
cables together. Adjacent cabling, however,
is not always the cause of problems at
10GigE speeds.
Electromagnetic
interference present in the building
(external noise) from a variety of sources,
like computing equipment, lighting and
motors, can reduce the signal-to-noise ratio
in a cable. Using a tester with an external
noise spectrum measurement is a solution to
identify and correct this interference.
Peter Schweiger is market development
manager for
Agilent Technologies, Santa Clara,
Calif.
For more information
(click here)