On Tuesday, Intel had a briefing in San Francisco to discuss their plans for the micro server marketplace. Not all server work loads are 24/7 intensive number crunching that requires a rack of 6U multi-socket servers in a single giant data center. In an Internet data center, the challenge is to handle millions of relatively small, independent tasks like those needed for searching, social networking, viewing web pages, and checking email. Noteworthy were comments from Chief of Facebook Labs, Gio Coglitore.
This mismatch between volume servers and the now dominant Internet workload is the primary cause of the rapid increase in server power consumption and is responsible for the multi-billion dollar power problem in the data center.
Power consumption is one of the major factors that distinguishes the micro-server from the traditional form factor server. Micro-servers try to pack a lot more single-chip boards into a smaller space with much lower power levels and a smaller overall footprint. Obviously, Intel’s move into this arena is to protect their x86 architecture.
Intel has been working on providing their server customers with lower-power servers since 2008. Last week, we showed a photo of their original micro-server reference design. Presently Intel micro-server CPUs power range is from 45 watt TDP (thermal design power) E3-1260L Xeon to 20 watt TDP, E3-1220L Xeon. Intel’s roadmap shows in second half 2011 they will have a 15 watt TDP Sandy Bridge version of the Xeon CPU.
Tuesday, Intel claimed that their customers are placing an emphasis on 64-bit capability, virtualization, ECC memory, and software compatibility.
The micro-server form factor calls for support of ECC (Error Checking and Correcting) memory which is a critical feature for server dependability. ECC memory, helps protect servers from failure due to errors in the transfer of data to and from the memory. Unlike standard memory, ECC memory can actually detect and correct single-bit errors. With standard memory, if even a single-bit memory error occurs, your server stops functioning.
Gio Coglitore, Director of FaceBook Testing Lab, explained why they are evaluating the micro-server form factor. Clearly, FaceBook’s Internet model is lots of short time frame data connections. Coglitore favors the micro-server over the virtual multi-socket server model. He said that with a few large centralized virtual servers too much of their infrastructure can be taken down when a virtual server node goes offline.
Coglitore said the other item not often discussed with virtualization is the necessity of having a standby replacement server attached to a large pipe to take care of a failure. He said their usage models show that they lose the so-called cost saving advantage of virtualization. Instead, they let a load-balancer handle the process of redirecting data flow when a node drops offline.
Intel’s VP of the Architecture Group, Boyd Davis, said that they project the micro-server will grow to 10 percent of server sales over the next 4 to 5 years. He did not say how much of that will be their Atom processor. Davis said that the vast majority of Intel-based servers are single and dual core.
SeaMicro backed by Draper Fisher Jurvetson and Khosla Ventures, is located about 2.5 miles from Intel’s Santa Clara headquarters. They have built a fabric approach around the Atom processor. On Monday, SeaMicro announced a low-power server that includes 256 of Intel’s latest Atom N570 dual-core processors. The SM10000-64 has 512 Atom processing cores running at 1.66GHz, which combine to deliver 850GHz of processing power, the company said. Each core is capable of running two threads simultaneously, which helps boost application performance.
ARM is not about to be left eating Intel’s dust. There are options like the ARM-based server maker Calxeda and CPU designer Tilera, both of which are aimed at the same small form factor server marketplace.
The small form factor server has a lot going for it and may eat a lot more of the larger server market than Intel is projecting.