Startup 3Leaf Systems has launched its first product offering, the Dynamic Data Center Server (DDC-Server). It is a combination of silicon and software that pools server CPU and memory into dynamically-sized virtual servers. Essentially it aggregates hardware resources so that a cluster farm can be turned into one or more SMP systems. The 3Leaf solution consists of a proprietary ASIC and a set of software that can be inserted into standard x86 server technology.

It is designed to solve multiple problems in the datacenter, including under-utilization of infrastructure and limitations of existing virtualization schemes. Target markets include traditional enterprise customers as well as eCommerce, social media, and high performance computing users -- essentially anyone with a need for scaled up x86 machines. And since the technology enables the cluster nodes to be dynamically split and recombined according to application requirements, it can appeal to organizations that now maintain multiple systems to run different types of workloads.

The company is marketing the technology as an enabler of the "dynamic cloud," but overall, the 3Leaf offering delivers a solution similar to that of ScaleMP's vSMP technology, in that it enables a number of commodity x86 servers to be aggregated into a large shared memory SMP system that runs a single OS image. The idea is to be able to replace much more expensive proprietary SMP machines by using commodity building blocks. However unlike ScaleMP, which uses a software/firmware-only solution, 3Leaf uses a combination of hardware and software to achieve SMP virtualization.

In the case of 3Leaf, their ASIC is placed on the motherboard and enables distributed memory coherence across the cluster's fabric of choice, either Ethernet or InfiniBand. Essentially, the chip acts as a memory coherence controller. The fact that low-latency interconnect switches and adapters are now just commodity server components, rather than custom parts, opens the door to the type of solution 3Leaf is offering.

The downside is that the 3Leaf ASIC must be present on each server in the cluster, so it's not the plug-and-play experience that you would get with a software-only solution. The first 3Leaf product set supports AMD's HyperTransport architecture, where the 3Leaf ASIC is plugged into the Socket F interface. The server being shipped today is built on a two-socket Opteron motherboard supplied by Supermicro. With this solution, up to 16 nodes (192 cores) and 1 TB of memory can be aggregated into a single virtual SMP system. Next year, 3Leaf will offer an Intel version, based on the company's Quick Path Interconnect (QPI) 1.1 and the "Sandy Bridge" processors. That product set will be able to scale up to 32 nodes, many hundreds of cores, and 64 TB of memory.

According to Bob Quinn, 3Leaf founder, chairman and CTO, the rationale for using hardware rather than just software to create a virtual SMP has to do with performance. The ASIC allows a memory page to be read and written simultaneously by an application on two different nodes, since the coherency is hardware enforced at the level of a 64-byte cache line. In a software solution, the OS must get involved, stopping and then restarting one thread to allow another thread to access the same memory page.

"In the case of 3Leaf, we behave like a big old expensive IBM, or SGI, or Sun system," says Quinn. "It really is a traditional cache-coherent shared memory system, with the difference being it's not all custom-designed hardware. It's using existing switches to provide the equivalent of a custom-designed backplane."

But it's not all about just building big SMP machines. The 3Leaf software, which is delivered in firmware, is used to control the way the cluster resources are divvied up. There are three flavors: DDC-Pool, DDC-Range, DDC-Flex. DDC-Pool is for building static SMP systems at the granularity of the cluster node. In this case, resizing the SMP requires a reconfiguration and reboot. DDC-Range is also a static solution, but offers the granularity of allocating compute resources down to the level of an individual core. With this software, a virtual SMP machine can be constructed from various sized slices of one or more physical servers. DDC-Flex provides the granularity of the DDC-Range, but allows the user to reconfigure the cluster while running, rather than requiring a reboot. DDC-Flex is not yet available, but is planned to be released sometime in 2010.

The ability to slice and dice a moderate sized cluster into one or more virtual servers means that users can use a single set of hardware as a platform for heterogenous workloads. For example, in the oil and gas business, seismic data analysis works fine with vanillas clusters in a distributed memory environment, but advanced reservoir simulations are often better run in large shared memory environments. With 3Leaf technology, both applications can be served by the same cluster hardware. That model, says Quinn, can be applied across many application domains.

For the past year, the product has been in the hands of beta customers, including a number of HPC users. Jim Lupo, a researcher at LSU, has been testing the 3Leaf platform with hurricane storm surge prediction and molecular dynamics codes. According to him, performance was comparable to other HPC systems, but since the technology supports both shared memory and distributed memory environments, the 3Leaf system was more flexible and required less admin and programming support.

Although 3Leaf is building the initial AMD-based systems today, the company's market strategy involves partnerships with OEMs and system integrators. The idea is to get vendors like HP, IBM and Dell to take this technology to market as an addition to their x86 server lineups. Quinn says they are currently engaged with all the tier 1 OEMs and a number of tier 2 and 3 OEMs as well.

The challenge here is many system vendors already offer their own proprietary top-of-the-line SMP machines, like the HP Superdome 9000, IBM Power 595, and the SGI Altix 4700. While these are not x86-based machines, they're still aimed at the kind of high-end applications 3Leaf has in its sights. The company is betting that all the major OEMs are looking to offer big x86 shared memory machines, and is hoping that partnering with them will be the most attractive path to get there.

In the HPC space, SGI's upcoming "Ultraviolet" product line, which will move the company's NUMAflex shared memory architecture onto an Intel x86 platform, would perhaps be the most directly threatened by 3Leaf-based platforms. It may come down to a price-performance calculation, but since neither of these products is in the field yet, it's impossible to say how they might match up.

In general, 3Leaf wants to put a lot of daylight between the cost of one of its setups and an equivalent proprietary shared memory system. Pricing on the 3Leaf DDC-Server products shipping today vary from $99,000 for a low-end model (256 GB of shared memory, 96 cores of 2.4 GHz Istanbul processors, and 4 TB of storage) up to $250,000 for a maximum configuration (1 TB shared memory, 192 cores of 2.8 GHz Istanbul processors, and 8 TB of storage). The price includes the InfiniBand switch, cables, Linux operating system, and 3Leaf's DDC-Pool software.

Anyone curious to see 3Leaf systems in action this month can attend the upcoming Supercomputing Conference (SC09) in Portland, Ore., where the company will be demonstrating its technology.