Vendors claim new-generation blades trim server
costs. They’re right—if your power system can take the
heat.
If the name of the data center game is getting more computing
power for less, blades should be the hottest thing since South
Beach. They’re more manageable and deliver better TCO than
their 1U counterparts—our latest testing shows as much as a
four-fold increase in processor density combined with 20 percent to
30 percent power savings.
So why did Gartner Dataquest put this year’s blade shipments
at an anemic 850,000 units this year, just 10 percent of total
server sales?
Because earlier-generation blade servers were like fad
diets—long on hype, short on delivery. Despite vendor
promises, they didn’t represent much of a savings over
conventional devices. Most of the systems we evaluated when we
reviewed blade servers in June 2003 were struggling with
first-generation blues—an 8- or 10-blade chassis used the
same amount of rack space as equivalent IU devices and suffered I/O
bandwidth limitations between blades and backplanes, making them
better-suited for Web server consolidation than running critical
databases.
But even then, one fact came through loud and clear: managing
blades is substantially easier than dealing with individual racked
boxes.
 |
Today, blade server designs have improved, with enough midplane
throughput and modularity at the chassis to provide investment
protection for their three-to-five-year lifespan. Processor density
has increased, and power consumption is lower than you might
expect.
They also deliver incredible flexibility. Instead of limiting the
blade system to particular types of I/O—interconnect, network
or storage—vendors overprovision the I/O channel, providing
sufficient bandwidth for targeted environments, or let IT allocate
I/O as it sees fit. Seems vendors are following the lead of core
switch vendors: make the frame big enough to jam in just about
anything you want for the foreseeable future.
Enterprises are finally catching on. Blade shipments will rise to
2.3 million units by 2011, to account for almost 22 percent of all
server purchases, according to Gartner Dataquest. Although blades
are still more expensive than conventional 1U servers, you should
see operational savings in the 30 percent range, according to Imex
Research. Those changes make the newest generation of blade servers
excellent candidates for high-demand core and server-virtualization
applications.
That’s the blade server story vendors want you to know about.
The less flattering side: even while delivering power savings, the
energy demands of these high-density systems will still tax the
overall infrastructures of many older—and even some
newer—data centers. You may be able to quadruple the
processor density of a rack, but can you power and cool it?
Many data-center managers are saying no. By 2011, 96 percent of
current data-center facilities are projected to be at their power
and cooling capacity limits, according to a recent survey of the
Data Center Users’ Group conducted by Emerson Network Power.
40 percent of respondents cited heat density or power density as
the biggest issue they’re facing. (We examine HVAC and
electrical requirements and offer design suggestions in our
‘This Old Data Center’ special issue at
nwc.com/showitem.jhtml?docid=1610f1 and our Data Center Power
Issues analyst report at nwcanalytics.com)
The Players
Most first-tier server vendors now offer blade lines. Gartner
Dataquest, IDC and our readership agree that IBM, Hewlett-Packard
and Dell, in that order, are the top three players in the market.
IBM holds about a 10-point lead over HP, with Dell a distant third
at less than half HP’s share. No other single vendor is in
the double digits, though judging by our testing, Dell should be
watching over its shoulder for Sun Microsystems.
In an unlikely coincidence, IBM is flexing its muscle by proposing
to standardize blade system modules and interconnects around its
design. Although standardization would benefit the enterprise,
we’re not convinced IBM’s proposal is the best
solution. (See box: ‘IBM and the Quest for
Standardization’)
We asked Dell, Egenera, Fujitsu, HP, IBM,
Rackable Systems and Sun to send a base system, including a chassis
and four x86-based server blades with Ethernet connectivity, to our
new Green Bay, Wisconsin, Real-World Labs, where we had
EqualLogic’s new PS3800XV iSCSI SAN array and a Nortel 5510
48-port Gigabit Ethernet data-center switch online for testing.
We were surprised when only HP, Rackable Systems and Sun agreed to
participate. (See our methodology, criteria, detailed form-factor
and product-testing results at nwcreports.com; our in-depth
analysis of the blade server vendor landscape and poll data can be
found at nwcanalytics.com.)
Not only are the cooling and power improved in our new lab digs,
it’s actually possible to see daylight while testing if you
set your chair just right. It may have been the bright light, but
once we got going, the differences in our three systems came into
sharp focus.
Three for the Money
HP submitted its new 10-U C-Series BL7000c enclosure along with
two full-height and two half-height ProLiant BLc server blades. Sun
sent its recently released 19-U Sun Blade 8000 Modular System with
four full-height Sun Blade X8400 Server Modules. Rackable Systems
provided five of its Scale Out server blade modules.
Rackable blew away the competition in sheer node count. Rather than
basing its design on an eight- or 10-blade chassis, Rackable goes
large with a proprietary, passive rack design that supports 88
blades and has an intriguing focus on DC power. If you’re in
need of processors, and lots of ’em, Scale Outs may be just
the ticket.
What excited us most about the blades from both HP and Sun was
their potential for future expansion to the next generation of
processors and high-speed I/O interfaces. The midplane bandwidth
offered by these systems will be perfectly suited for the rapidly
approaching 10 Gigabit Ethernet and 8-Gb/10-Gb Fibre Channel
adapters and switches. (For more on blades for storage, see box
‘Storage on the Edge of a Blade’) These bad boys have
clearly been designed to provide investment protection.
 |
HP also is on the verge of introducing its new Virtual Connect
technology for Ethernet and Fibre Channel. Exclusive to HP and the
BladeSystem C-Series, Virtual Connect modules allow as many as
four linked BladeSystem chassis to be joined as a Virtual Connect
domain that can be assigned a pool of World Wide Names for Fibre
Channel or MAC and IP addresses for Ethernet. These addresses are
managed internally and dynamically assigned by the Virtual Connect
system to individual blades within those chassis. By managing these
variables at chassis level rather than at blade or adapter level,
individual blades can remain stateless, making it easier for system
administrators to swap out failed modules or assign hot spares for
failover.
Simplify My Life
Clustering for availability as well as monitoring and management
are the top two blade functions, according to our poll. And indeed,
vendors tout their systems’ ability to dispense with
third-party management software and/or KVMs as a key selling
point.
Physically configuring, installing and cabling conventional servers
can be a massively time-consuming process that must be done during
a scheduled downtime to ensure other servers in the rack are not
disrupted accidentally. Even something as simple as replacing a
failed NIC or power supply can be risky business.
But with blade systems, modules are designed to be hot-swapped
without the need to bring down the entire chassis. The time
required for server maintenance can drop from hours to minutes, and
the unified management interfaces offered by most blade systems can
dramatically simplify the process of server administration from a
software standpoint.
All three systems we tested offer onboard, blade-level network
management interfaces as well as front-accessible KVM and USB ports
for direct connections.
The tools Rackable provided offered a degree of control over any
number of remote servers, but the absence of more powerful,
integrated features, such as remote KVM and desktop redirection
over Ethernet, is a noticeable omission in Rackable’s Scale
Out design when compared with chassis-based blade systems. And with
its chassis-less design, Rackable doesn’t offer a unified
management interface.
Conversely, HP and Sun provide extremely detailed, Web-based
management and monitoring capabilities at both server and chassis
level. Both systems offer integrated lights-out management
interfaces, for example, enabling us to do status monitoring and
system configuration across all the blades in our chassis.
But HP’s ProLiant was clearly the class of the management
category. HP went the extra mile by adding to its Web-based
controls a local multifunction LCD interface at the base of each
chassis that supports all the management features found in the Web
interface, without the need to connect a laptop or KVM to the
chassis. At first we were unimpressed with yet another LCD display,
but we were won over by the elegant simplicity and surprising
flexibility offered by that little screen. The HP Insight Display
interface is identical remotely or locally and offers role-based
security, context-sensitive help, graphical displays that
depict the location of problem components, and a chat mode that
lets technicians at remote locations interactively communicate with
those working on the chassis. We could easily imagine how useful
two-way communication capabilities would be in a distributed
enterprise.
Go Green
The costs of power and cooling in the data center are at an
all-time high, and it’s not just the person signing the
utility checks who’s taking notice. In July, the US House of
Representatives passed House Bill 5646 on to the Senate. This
measure directs the Environmental Protection Agency to analyze the
rapid growth and energy consumption of computer data centers by
both the federal government and private enterprise. Sponsors of the
bill cited data-center electricity costs that are already in the
range of $3.3 billion per year, and estimated annual utility costs
for a 100,000 square foot data center at nearly $6 million.
Denser systems run faster and hotter, and the cost of running a
server could well exceed its initial purchase price in as little as
four years. Because the actual energy use of any blade system is
dependent on a number of variables, such as processor, memory,
chipset, disk type and so on, it’s virtually impossible to
establish a clear winner in this category.
All three vendors told us their systems offer an estimated 20-25
percent reduction in energy costs over similarly equipped
conventional servers, and all provide temperature-monitoring
capabilities.
We were intrigued by Rackable’s DC-power option, which could
provide the greatest savings for enterprises that have implemented
large-scale DC power distribution throughout the data center. Most
of the power efficiency provided by the chassis-based blade systems
from HP and Sun stems from the ability to consolidate power and
cooling at rack level.
But HP takes this concept a step further with its Thermal Logic
technology, which continually monitors temperature levels and
energy use at blade, enclosure and rack levels, and dynamically
optimizes airflow and power consumption to stay within a
predetermined power budget.
Expensive Real Estate
Another key benefit of blade systems is the ability to pack a
lot of processing power into the least amount of rack space. Of
course, how well a vendor accomplishes this goes beyond the number
of processors—we looked at how blades were grouped and how
well they’re kept supplied with data.
Rackable’s side-by-side and back-to-back rack provides by far
the greatest processor density per square foot of floor
space—as many as 88 dual-processor servers equate to 176 CPUs
per rack, twice that if you count dual-core processors, putting
Rackable well ahead of HP and Sun in this category.
Coming in second in the density challenge is HP. Four 10U
BladeSystem c7000 enclosures fit in a single rack; each enclosure
holds 16 dual-processor half-height blades, squeezing 128 CPUs into
a conventional rack. The 19U Sun Blade 8000 chassis fit two to a
rack, and each chassis can handle 10 four-processor server modules,
for a total of 80 processors per rack.
But processor density is only part of the story.
When comparing blade systems, it’s important to note the
difference between two- and four-processor blades. The
dual-processor options from Rackable and HP are the basic
equivalent of conventional 1U and 2U servers, while each
quad-processor Sun Blade X8400 matches a traditional 4U box. This
configuration supports the assignment of much larger tasks on a
blade-by-blade basis and makes the Sun system a better candidate
for demanding applications.
Check Out Those Pipes
The modern IT environment is as much about data throughput as
processing capacity. That means for blades to be competitive
against conventional servers, they must keep up with high-speed
fabrics such as 4-Gb Fibre Channel, 4x InfiniBand and 10-Gb
Ethernet while still supporting multiple GbE links for management
and basic network traffic.
When it comes to total backplane bandwidth, we found little
difference between HP’s and Sun’s designs. PCIe, Fibre
Channel, InfiniBand or Ethernet—it’s all serial in
nature. It really comes down to who’s got the pipes, and in
this case it’s Sun. 
What makes this an important IT issue is the fact that, for now, a
decision to buy a given blade system locks you into a single
vendor’s hardware platform for a period of time. Ensuring
that the chassis you purchase can accommodate future high-speed
interfaces and processors provides investment protection.
Did we mention the Sun Blade 8000 has huge pipes? Its midplane can
handle up to 9.6 terabits per second in combined throughput.
According to Sun, this equates to 160 Gbps in usable bandwidth per
blade when you add in protocol overhead and other factors.
HP’s BladeSystem C-Series offers substantial 5-Tbps midplane
bandwidth when measured at pure line speed, more than enough to
support multiple high-speed fabrics from each blade. HP also offers
additional high-speed cross-connects between adjacent blade sockets
designed to improve performance for multiblade clustered
applications as well as to support plans for future
storage-specific blades.
In the case of Rackable Systems, the 2-Gbps offered by the dual GbE
ports is perhaps the weakest link in the Scale Out
design—this much lower bandwidth potential is a shortcoming
that will probably limit use in many high-performance,
high-bandwidth applications.
The other side of the I/O issue is port diversity and flexibility.
Blade systems can potentially be greater than an equivalent sum of
racked servers, thanks to their ability to share support of
multiple fabrics and reduce cabling with integrated switch modules.
The Sun Blade offered the most potential here by virtue of the
remarkable bandwidth of its PCIe midplane architecture. But
HP’s BladeSystem C-Series currently provides by far the
greatest port diversity when it comes to available backplane
switches and pass-through modules.
What’ll All This Cost Me?
Comparing pricing on a purely apples-to-apples basis turned out
to be a fruitless quest because the variety of approaches taken by
our three vendors made direct comparison difficult.
To award our pricing grade, we created a formula that pitted two
dual-processor, Opteron 285-based server blades from HP and
Rackable against a single four-processor Opteron 885-based blade
from Sun. We made sure to price each system on a blade level with
similar memory, SATA storage and dual-port Gigabit Ethernet
connectivity, without consideration for the differences in
chassis/rack capabilities. Not perfect, but at least we’re in
the produce family.
Rackable’s Scale Outs came in at $9,780 for two
dual-processor blades, making them the least expensive option on a
processor-by-processor basis. The HP BladeSystem landed in the
middle, at $11,768 for a pair of ProLiant BL465c half-height
blades. An equivalent Sun Blade would run $24,885 apiece when you
include the cost of the two PCIe Express GbE modules required to
match the port counts of the other two systems.
This was no surprise: The Sun X8400 blade is a quad-processor
system, and it was a foregone conclusion that it would be more
expensive than its dual-processor counterparts. In all fairness to
Sun, this was like comparing a pair of two-way, 1U servers to a
single four-way, 4U system; even though the processor count is the
same across configurations, it costs a lot more to make a four-CPU
server.
See more pricing and warranty details in the accompanying feature,
‘Bullish on Blades.’