A new SeaMicro server based on low-power Intel Atom processors may find a ready and willing audience among service providers and organizations with large Internet-oriented workloads that aren't often virtualized.
Santa Clara, Calif.-based SeaMicro's claims its SM1000 server, which was unveiled this week, can deliver the same performance as 40 1U dual-socket quad-core Intel Xeon or AMD Opteron servers, while consuming only one-quarter the power and space and still running standard x86 applications.
The vice president of engineering operations at a large Internet software company is beta testing the SM10000, and says that it directly addresses his primary constraint: power.
"Everyone's concerned about power, especially out here in the Bay Area," he said. "When talking to the vendors, we ask them to price on power. Nothing else really matters."
Pebbles vs. boulders
The SM10000 collects 512 1.6 GHz Intel Atom processors, some proprietary ASICs chipsets and FPGAs and memory and strings them together with networking and disk into a supercomputer-like compute fabric. At 10U, the system has the same performance-per-watt specifications as a full 40U rack, the company claims.
But the SM10000 is not for everyone, or for every workload, warned Andrew Feldman, SeaMicro's CEO.
The SM10000 is designed primarily for "Internet" workloads, Feldman said, that is, workloads that are "highly partitioned [and] independent and can be executed quickly in a single pipeline." If traditional enterprise workloads are large "boulders" that only heavy equipment can carry, Internet workloads are "pebbles" that need to be carried by the thousands, he said.
At the Internet software maker, the SM10000 is running a number of applications: MapReduce and Apache Hadoop, front-end Web serving, testing, and even Secure Socket Layer. But the SM10000 is not a good fit for the app and database tier because of Atom's 2 GB memory limitation, according to that company's vice president. "Just try running Oracle in 2 GB of memory; it's not going to happen," he said.
"As much as I'd like to, I can't just replace my entire data center with these things," he said.
Extending the amount of memory available to the system beyond 2 GB per processor would dramatically extend SM10000's usefulness, he said. "Once you get to 4 GB, a whole new universe of applications opens up." He said he had talked to SeaMicro about the issue and that the company is working on it.
Calling Linux lovers
Assuming its workloads are a good fit for the SeaMicro box, a prospective customer also needs to determine whether it is willing to run proprietary, first-generation hardware that has yet to figure on an independent software vendor's hardware compatibility list.
That shouldn't pose too much of an issue for open source, engineering-savvy Internet companies and managed services providers, said Pete Sclafani, the CIO and co-founder at 6connect Inc., a San Jose, Calif.-based consulting company focused on data center infrastructure.
"As long as I'm able to access all the information for a server, I'm comfortable with the notion of proprietary hardware," Sclafani said. "If [the vendor] deliberately closes my view into that server, that's where I'm concerned."
Meanwhile, the Internet software vendor vice president said that SeaMicro had been responsive to his company's feedback, adding support for virtual LAN tagging and hot-swappable CPUs, for example.
And as far as software support, most of SeaMicro's target customers run open source applications on Linux, and have less stringent support requirements than do enterprise shops, said SeaMicro's Feldman.
"The beautiful thing about the open source world is no one says anything," he said. "There's no Grand Poobah stamping things saying 'You're approved.' Linux is the dominant OS in 298 of 300 [target] customers, and what they want to run is the LAMP stack."
Specialty servers vs. virtualization
The new SeaMicro server exemplifies an emerging trend: a movement away from commodity servers running virtualization and toward specialized servers optimized for specific workloads on bare metal.
In many ways, the SM10000 aims to solve the same problem as that addressed by virtualization: the inefficiency of today's systems. Average server CPU utilization remains stubbornly below 20%. Virtualization software such as VMware vSphere can increase the efficiency of a CPU by partitioning it into smaller chunks.
But virtualization comes at a cost, architectural and financial, and goes at the problem the wrong way, said SeaMicro's Feldman.
"To us, virtualization is the proof that CPUs are the wrong size," he said. When SeaMicro sees a Xeon system running VMware, "we scratch our heads and say 'Why not just buy smaller CPUs?'"
Sclafani of 6connect said that virtualization has its place, but he also sees a role in data centers for servers like the SM10000.
"Let's say you're serving up Web requests from the front page of Digg. They're small requests, but there are millions of them," Sclafani said. "If you do have these kinds of workloads, [the SM10000] could be part of your server strategy."
That's especially true in power-constrained data centers, he added. "If power and cooling are the main parameters in my data center, and I can kick out a rack of Xeons, I'm going to consider it."
The SeaMicro SM10000 will be generally available July 30, 2010. Pricing starts at $139,000 with 512 processors and eight Gigabit Ethernet interfaces. Disk and memory are additional.