SAN JOSE, Calif. – A Microsoft executive called for a new class of multicore system-on-chips to drive the lower power servers needed for tomorrow's data centers. But he poured cold water on ARM-based chip vendors, hoping to get design wins in such systems.
"I've been involved in instruction-set architecture transitions multiple times, and they are extremely painful," said Dileep Bhandarkar, a distinguished engineer in Microsoft's data center group, speaking at the Linley Data Center Conference here.
"The rule of thumb is to make that kind of change you have to have at least a sustainable 2x performance improvement per dollar or per watt—and ARM is not there," he said.
"ARM's interesting to look at, and if it lights a fire under Intel and AMD that makes us happy," said Bhandarkar, an Intel CPU and server design for several years before joining Microsoft four years ago.
The breadth of server software used in today's data centers makes use of ARM unlikely in the foreseeable future, even through Microsoft announced it will support ARM in the next version of Windows, he added.
Instead, Bhandarkar called for 16-core SoCs based on Intel Atom or AMD Bobcat cores. Such chips should integrate all the core logic and I/O functions the companies now put in separate chips, he said. "There's a huge opportunity using these smaller cores to be more energy efficient, and we are talking to both AMD and Intel [about that]," he added.
Microsoft operates between ten and 60 data centers worldwide, each using thousands of servers to run its online business that has hundreds of millions of email, Web and search users.
So far the centers use only x86 chips, typically the lowest power versions. It is currently using dual-socket boards with quad-core chips using a single virtual machine per core.
Bhandarkar described a streamlined x86 server design Microsoft will use for its upcoming fourth-generation data centers, inviting others to adopt the system. Microsoft has not published a specification for the servers, but Bhandarkar said he will reveal more details of the design in a blog in the next few weeks.
The new Microsoft design defines a single node as two low power x86 CPUs in a half-size board that includes one PCI Express slot, 4 DIMM sockets per CPU populated with low-power DRAM and four Serial ATA connectors linked to four hard disk drives. Two of the servers and eight drives total are placed in a 1U rack-mounted unit so that 96 servers can be crammed into a standard rack.
Microsoft currently uses a dual-ported 1Gbit Ethernet link for clustering, with only one port active. It may be two years before it populates its open PCI Express slot to shift to 10Gbit Ethernet.
Microsoft has shared the details of its spec with as many as four server makers. By contrast, its archrival Google is well known for keeping secret its proprietary design for servers it uses in its data centers.
"We don’t restrict vendors from not selling our design to others," said Bhandarkar. "High volume drives lower prices, so to the extent there's a single standard that's good, as long as it's my standard," he said.
To further simplify its design, it is moving to using outside air for cooling with minimal use of air conditioning and moving the temperature of its data centers toward 85 degrees Fahrenheit. It has also defined a larger unit of integration, a 20-foot, 160 KW data-center-in-a-box larger than the shipping containers it currently uses in facilities in Chicago.