When Cisco Systems introduced its Unified Computing System earlier this year, the Extend Memory technology was a focal point, but it was unclear how that technology worked.On Tuesday, a Cisco executive explained how the patented memory technology in some of its Unified
The question is whether Cisco Systems Inc.'s efforts to offer larger memory capacity -- and thus entice users to its UCS system -- will be effective when many IT pros remain hesitant to abandon existing data center investments.Adding capacity
In traditional systems, the CPU memory controller can use only a certain number of DIMMs and, thus, seeks out that number. The latest Intel Xeon 5500, for instance, can address up to 12 or 18 DIMMs, though it poses a performance tradeoff for the latter, David Lawler, Cisco's vice president of platform products, told SearchDataCenter.com.
Cisco's Extend Memory technology makes a CPU see one DIMM as four separate DIMMs, giving a single 12-DIMM server the memory capacity of 48 DIMMs and up to 384 GB of memory.
To address this issue, Cisco engineers placed a high-performance chip on the memory bus between the processor and the DIMM that changes the way the CPU searches for DIMMs, Lawler explained. "So when the CPU searches for an 8 GB DIMM, we can represent that as four 2 GB DIMMs instead," he said.
"Conceptually, it is like virtualization, but this is not a virtual technology," Lawler said.
On the flipside, a startup company called RNA Networks does virtualize memory by creating memory pools that can be used by any server nodes on a given network.
Cisco took a physical approach, though. "We looked at the memory bus as a network and figured out if we changed the addressing that the memory controller looks for, we could allow it to use more DIMMs," Lawler said. "By playing with the variables … we intercept the physical address and say, 'Instead of looking for an 8 GB, we give it four 2 GB DIMMs."
Whether a CPU accesses an 8 GB DIMM or four 2 GB DIMMs makes no difference from a capacity standpoint, but using more, smaller DIMMS versus fewer, larger DIMMs is cheaper. An 8 GB DIMM is significantly more expensive than buying four 2 GB DIMMs because the cost of memory increases exponentially with density. A 2 GB DIMM, for example, costs around $125, but an 8 GB DIMM can cost more than $1,000.
To accommodate virtualization, many blade servers have eight or more DIMM slots. For instance, one of HP's latest blade servers, the ProLiant BL490c server, can support up to 18 DIMMs and up to 144 GB of memory per blade using registered DIMMs. Cisco's memory extender technology gives its servers more than double that capacity, so a 12 DIMM blade server actually has the capacity of 48 DIMMs and up to 384 GB of memory.
Users aren't sold on UCS
Servers hosting virtual machines with large databases can run out of memory far before they run out of CPU power, so having that extra memory capacity to work with could be a strong selling point for Cisco.
"On a normal host, memory usually is the first resource we're short on. When the CPU of a fully loaded ESX box is still at 60%, we often run into memory shortage," said virtualization expert Gabrie van Zanten.
On the Ars OpenForum IT community site, one Unix administrator and virtualization user listed the "huge memory density with full-size blades" on UCS as one reason he may switch from Dell blade servers to the UCS.
But for the virtualization users that aren't experiencing memory constraints, the Extend Memory perk makes no difference.
Armin Heinlein, the head of IT competence at the Panalpina Group, a Morristown, N.J.-based freight transportation company, virtualizes Citrix Systems Inc. applications using VMware on IBM blade servers. Panalpina uses Cisco switches, but Heinlein doesn't see a need for UCS or the memory capacity it offers.
"We have a lot invested in IBM already and are doing well with the memory capacity we have," Heinlein said. "We run our Citrix environment and Windows on IBM BladeCenter blades and didn't have to buy extra memory capacity, so we don't have any need for [Extend Memory]."
Like Heinlein, many IT pros say they won't invest in UCS because they don't want to abandon investments in existing server technologies from trusted x86 server names such as Hewlett-Packard Co. and IBM. The poor economy is also a factor, as is skepticism surrounding Cisco's foray into the server market.
Still, Cisco will begin shipping the UCS system in the next few weeks and there are sure to be some takers. The blade with memory extension technology will not ship until August, and the rack server with memory extender technology will ship in Q4 2009.
Let us know what you think about the story; email Bridget Botelho, News Writer.
And check out our data center blogs: Server Farming, Mainframe Propellerhead and Data Center Facilities Pro.