Little servers for big applications

I attended Intel Developer Forum (IDF) in San Francisco last week, one of the premier events for anyone interested in microprocessors, system technology, and of course, Intel itself. Among the many wonders on display, including high-end...


I attended Intel Developer Forum (IDF) in San Francisco last week, one of the premier events for anyone interested in microprocessors, system technology, and of course, Intel itself.

Among the many wonders on display, including high-end servers, desktops and laptops, and presentations related to everything Cloud, my attention was caught by a pair of small wonders - very compact, low power servers paradoxically targeted at some of the largest hyper-scale web-facing workloads.

Despite being nominally targeted at an overlapping set of users and workloads, the two servers, the Dell “Viking” and the SeaMicro SM10000, represent a study in opposite design philosophies on how to address the problem of scaling infrastructure to address high-throughput web workloads.

In this case, the two ends of the spectrum are adherence to an emerging standardised design and utilisation of Intel’s reference architectures as a starting point versus a complete refactoring of the constituent parts of a server to maximise performance per watt and physical density.

The Dell Viking is one of the first systems based on the Server Systems Infrastructure (SSI) specifications. SSI is an industry forum endorsed and sponsored by Intel and a number of other vendors, many of them second-tier server vendors and server technology suppliers.

The SSI spec in question is a specification for a power-efficient dense server, and Dell’s offering certainly look like it meets that goal. The Viking packs up to 12 independent 4-core Xeon 3000 1-socket nodes with multiple (up to four) disks per server, into a 3U chassis, for an aggregate density of .25U per server.

The servers each have multiple (up to 4) disks, and while not a true blade server with a shared midplane and aggregated I/O, they share power and cooling within the 3U chassis, similar to the HP SL-series and the IBM iDataplex.

While some specialised systems such as the HP BL2x220 G7 offer higher density of cores per rack, with a density of up to 156 server/624 cores per 40U rack, the Viking is an excellent candidate for medium to light-weight applications with storage requirements in the TB class per node in large scale-out environments, particularly in light of its claims for power and cooling efficiencies.

The SeaMicro SM10000 takes a completely different design approach, completely re-engineering the standard server module for massively increased density and reduced power. SeaMicro’s system consists of 64 server boards, each of which contains 8 single-core Atom-based servers.

At the heart of SeaMicro’s innovations is the virtualisation (sharing), via their own custom ASICs, of the NICs, SATA controllers, BIOS, and a number of other elements of a conventional server. By sharing common resources, SeaMicro was able to reduce the real-estate of each server to about the size of a credit card and the power to approximately 40W per 8 servers.

Servers are connected via a 1.28 Tb 3D torus internconnect, and each collection of 64 servers shares up to 64 SATA disks (servers do not require disks) and up to 16 10Gb/64 1Gb Ethernet links.

If you do the arithmetic, this works out to 512 single-core servers in a 10U enclosure running on less than 2000 W, for a density of .02U! (although I cannot get the math to work, and keep coming out with around 2,500 W per enclosure rather than the 2000 quoted by the SeaMicro staff).

With a generous but not absurd power budget of 8 1- 10 KW/rack, over 2000 cores can be provisioned in a single rack.

SeaMicro is an early stage privately held company, so they are being cagey about size and customers, but they claim they have major web properties in their existing customer roster, and will be announcing several of these before the end of the year.

Clearly the message about the need for optimised platforms for differing workloads has reached receptive ears, and both of these new server architectures appears to offer strong advantages, depending on the workloads, in very large scale-out environments.

I expect other Tier-1 vendors to follow Dell’s lead with dense single-socket servers, but SeaMicro is likely to be a unique solution for the less risk-averse I&O group.

Blog post by Richard Fichera

"Recommended For You"

AMD releases cloud-optimised Opteron 3200 processors Intel releases new 'smartphone class' Atom server chip