System vendor SGI today announced "Cyclone," a cloud service aimed specifically at technical computing. Although the company has sold hardware that ended up in other peoples' clouds, Cyclone represents SGI's first foray into the cloud as a service provider. The idea is to provide a purpose-built HPC cloud, wrapped with third-party application software, and backed up by SGI's considerable HPC expertise. "We expect great things out of Cyclone," says Geoffrey Noer, SGI's senior director of product marketing.

First the hardware. Cyclone will encompass three different SGI platforms: scaled-out InfiniBand/x86-based clusters (Altix ICE), "hybrid" CPU-GPU servers (Altix XE with a GPU option), and SMP-style shared-memory systems. Disk storage will be supplied by SGI's own InfiniteStorage line. For the shared memory systems, SGI will initially offer its Itanium-based system, presumably the Altix 4700s. But with the Nehalem-EX-based Altix UltraViolet (UV) due to be shipped in the second quarter, it's implied that SGI's newest shared-memory machines will be available for renting later this year. In the case of the hybrid platform, both NVIDIA Tesla GPUs and AMD/ATI FireStream graphics hardware will be offered to juice floating point performance. Interestingly, SGI's press release also mentions a Tilera processor option for integer acceleration. The company is not quantifying the in-house resources that will be applied to the Cyclone cloud, but according to Noer, the initial deployment will be on the order of "thousands of CPUs."

The system software relies on a mix of SGI and third party packages. At the OS level, it's Linux, with a choice of either SUSE or Red Hat. Cluster management is provided courtesy of SGI's own ISLE Cluster Manager, with ProPack included to tune performance. Altair's PBS Pro is used for job scheduling. From the top to bottom, it's pretty much what you'd expect from an SGI-equipped HPC environment.

Note that Cyclone provides a non-virtualized environment. That's a model that promises to squeeze maximum performance out of the hardware and eases at least some of the security concerns. On the downside, a non-virtualized model makes the infrastructure less elastic since individual applications may end up wasting resources.

The list price for computer time on Cyclone is $0.95 per core-hour -- that's right, per core not per CPU, so systems based on those upcoming 8-core Nehalem EX chips will be twice as expensive as their quad-core counterparts. Storage is also provided, with 500 GB of disk capacity included with the rental of the first node. More storage can be bought at a rate of $0.20 per GB per month.

Noer says those are the base prices for the infrastructure, with the idea that better rates can be had by buying in bulk. The base rate also doesn't apply to the GPU-accelerated servers (graphic processors have hundreds of "cores" so the per core model doesn't make sense). For the time being, SGI isn't quoting pricing for the hybrid systems, preferring to negotiate those on a customer-by-customer basis. Noer says they intend to offer a "reasonable" deal on the hybrid systems, inasmuch as they view these as development platforms for both ISVs and customers.

For a rough comparison, consider that a High-CPU Instance on Amazon's EC2 cloud offers the equivalent to 8 cores on a 2.5GHz to 3.0 GHz 64-bit x86 processor. That instance is priced at for $0.68 per hour on Linux/UNIX or $1.16 for Windows. Even better deals can be had if you take advantage of the Reserve Instance or Spot Instance models. Obviously, this is much less expensive than what SGI is offering, but with the Amazon infrastructure, you don't get InfiniBand or NUMAlink connectivity, which relegates these setups to HPC applications that are embarrassingly-parallel and don't rely on tight coupling between servers. The use of bona fide HPC systems in the Cyclone cloud is one of SGI's main differentiators from its generic cloud competitors.

But besides the infrastructure-as-a-service (IaaS) model, Cyclone is also providing software-as-a-service (SaaS). Initially SGI will offer 18 HPC applications across five domains, and provide them as a value-add service on top of the infrastructure. Commercial codes will entail an additional fee over and above the IaaS rental, although some of the open source codes may be offered gratis, at least for academic users. The specific pricing on the various SaaS offerings has not been made public and will be a function of SGI's licensing agreements with the various software providers.

The five initial software domains in Cyclone are computational biology, computational chemistry and materials, computational fluid dynamics, finite element analysis, and ontologies (semantic Web and data mining). SGI has partnered with a number of ISVs and non-commercial code providers to offer popular science codes like BLAST, HMMER, Gromacs, Acusolve, LS-DYNA, and about a dozen others. SGI plans to grow this collection of software packages as well as the range of application domains as more providers are signed up.

Although the OEM-as-cloud-provider is not a common route for HPC vendors, SGI is certainly not the first. IBM has been offering a Computing on Demand service for years, which started out as a way to rent time on IBM supers, but has evolved into a more general-purpose cloud offering. Last year Linux cluster vendor Penguin Computing launched its own HPC as a service, called Penguin On Demand (POD). Both of these services still seem to be alive.

A less happy fate befell Sun Microsystems' Network.com service, aka the Sun Grid Compute Utility. Like SGI's Cyclone, Network.com provided both Sun infrastructure and third-party software for HPC customers. The base rate to use the Sun grid was $1 per CPU-hour. Although Sun was able to sign up 40 software providers, the venture was shut down in late 2008.

At best, the model is unproven. In general, though, it seems reasonable to think that doing technical computing in a cloud will require a purpose-built HPC infrastructure, at least for a significant subset of these applications. Whether there's a long-term business model for that by OEMs, large-scale public cloud providers, or anyone else, is still unclear.

"SGI's clear market targeting should help the company differentiate itself from 'cycles for nothing' providers, and its value-added software and service components should lead to higher margin business opportunities for the company," says Chris Willard, chief research officer at InterSect360 Research. "That said, the HPC utility computing market, despite a long and storied past, has not yet captured significant market share, and it is unclear if the addition of Internet-based communications will be enough to tip the balance in favor of cloud computing."