Cisco Systems caused a major stir with the announcement of its Unified Computing System last month, a blade server based on Intel’s new Nehalem quad-core processors.
Today, the firm opened up with far more detail on how the UCS system works, confirming many of the details InternetNews.com reported earlier despite the company’s official cone of silence.
Cisco (NASDAQ: CSCO) acknowledged the blades will sport a pair of Intel (NASDAQ: INTC) Xeon 5560 processors with a special processor that greatly increases the amount of memory per blade.
The custom chip will make four memory sticks appear as one to the system, allowing for a tripling of memory. Each blade will come with 48 memory slots total for a maximum of 384GB of memory, if customers choose to use the very expensive 8GB memory sticks.
The chassis for the blade sits on Cisco’s Unified Fabric, reducing the need for switches and management software. A lot of blade servers use a management console for each blade, and then the chassis at large. Cisco’s management is done through the fabric interconnect, making up to 320 blades appear as one domain.
In addition, Cisco is introducing Service Profile, which encapsulate all of the identity of a blade; its MAC address, worldwide name, firmware, storage and network policies, boot order, and other identifiers. This profile, in whole or in parts, can be moved to new blades or copied for rapid blade deployment.
“If I have to rapidly scale up an application, I can duplicate that blade profile, put it on new blades and provision them to meet my app’s demands, or we can also replace a blade real quick,” Paul Durzan, director of marketing for the Server Access and Virtualization group at Cisco, told InternetNews.com. He said redeploying a blade can take three to five days without Service Profiles.
The virtualization adapter in UCS improves performance by letting the blade talk straight to the I/O system instead of going through the hypervisor. Normally this would be a recipe for disaster, letting several virtual machines all try to access the same I/O ports, but Durzan said the hypervisor is still in charge.
“It’s just hardware management of traffic instead of software,” he said. “The hypervisor is still the traffic cop but it lets you go through the intersection without having to stop for the light.”
Cisco said it’s working closest with VMware, but is also working with other vendors, including Microsoft and Citrix.
Next page: Buy more, spend less
Page 2 of 2
Buy more, spend less
Cisco is notoriously pricey, but for UCS, it has an argument that the more you buy, the more you save. Here’s why. The first chassis needs to have a fabric interconnect that connects to the Unified Fabric. Each subsequent chassis only needs a fabric extender, which connects to the interconnect.
The interconnects are under $4,000, much cheaper than the five-figure cost of a switch in every chassis that other blades require, argues Durzan. Plus, no management software is needed for each blade or chassis, because everything is managed from the fabric. So even though the chassis costs more and blades cost roughly the same as other blade servers, the savings in software and switches adds up as the server scales up.
Cisco’s connection is built on Fibre Channel over Ethernet (FCoE), whereas Sun’s recent Open Compute Environment, a blade system very similar to UCS announced on Monday, uses Infiniband. Durzan said he’s confident Cisco made the right choice.
“I’ve looked at Infiniband, I’ve looked at Ethernet, and as someone said in the past, you never bet against Ethernet. Bringing in Infiniband is bringing a whole other fabric into your datacenter,” he said.
By reducing the number of switches, Cisco argues its blades reduce cabling by 40 percent, power consumption by 30 percent and allow for 50 percent more physical servers.
More important, Cisco thinks the blades are so powerful that customers will be able to expand their virtualization uses. “I have heard customers say ‘we’re going to start looking at virtualizing apps that we couldn’t virtualize before because of the improved performance and things like hypervisor bypass,'” said Durzan.
Making a dent in the blade server market
Between them, HP and IBM own about 80 percent of the blade market, but Liam McGlynn, senior analyst for systems management at Enterprise Management Associates, thinks Cisco can make a big dent in that.
“The more you scale, the cheaper it gets. If I were an IT manager, how could I overlook this? This will save me money, this will make auto provisioning easier. This will reduce power consumption. I think Cisco is going to make a killing,” he told InternetNews.com.
Specifically, McGlynn says he thinks the reduced network hardware, auto provisioning and massive memory banks will be the big appeals to the blades.
“Those three things together are going to make a difference not because it’s a leapfrog over the competition, but a real step forward,” he said. “I think HP and IBM and the rest will have an interesting time trying to close the window I think Cisco has right now.”
Cisco plans to ship its Unified Computing Systems at the end of this quarter.
