Multicore software specialist Fixstars Corporation has released Yellow Dog Enterprise Linux (YDEL) for CUDA, the first commercial Linux distribution for GPU computing. The OS is aimed at HPC customers using NVIDIA GPU hardware to accelerate their vanilla Linux clusters, and is designed to lower the overall cost of system deployment, the idea being to bring these still-exotic systems into the mainstream.
Until recently, Tokyo-based Fixstars mainly played in the Japanese market and its software offerings were centered on Power CPUs and the Cell Broadband Engine. The company has established close ties with Sony, Toshiba and IBM (makers of the Cell), and used them as a springboard to build up an HPC customer base. Of the 80 or so employees at the company, just five are based in North America, with the remainder in Japan.
The company’s core competency is in multicore software optimization, porting services and training, with special emphasis on the Cell BE, and now GPUs. Fixstars also offers a range of Cell and NVIDIA GPU accelerator boards, but as a reseller rather than an OEM.
Fixstars’ 2008 acquisition of Colorado-based Terra Soft was a good fit for the company’s market focus and brought the “Yellow Dog” portfolio on board. Yellow Dog Linux was the first general-purpose OS for Sony’s PlayStation3 and offered a standard environment for Cell developers using the relatively inexpensive game platform. The Terra Soft acquisition also gave Fixstars a foothold in the North American market. Today, Fixstars’ customer set includes NASA, Boeing, Argonne National Lab, MIT, Lawrence Livermore Nation Lab, the US Air Force, and Lockheed Martin, among others.
With that in mind, Fixstars is looking anew at the US, believing that multicore software acceleration is well-positioned to make a big impact in financial services (options pricing, risk assessment, algorithmic trading), digital media (image video processing) health care (medical image processing) and the industrial sector (image/signal processing) over the next few years. While each of these verticals is well represented in Japan, the North American market is several times larger.
The problem is that the majority of future HPC accelerated deployments is destined to be GPU-based, rather than Cell-based. While Cell had a brief fling with HPC stardom as the processor that powered the first petaflop system — the Roadrunner supercomputer at Los Alamos National Lab — IBM has signaled it will not continue development of the Cell architecture for HPC applications. With NVIDIA’s steady evolution of its HPC portfolio, propelled by the popularity of its CUDA development environment, general-purpose GPU computing is now positioned to be the most widely used accelerator technology for high performance computing. The upcoming “Fermi” GPU-based boards (Q3 2010) substantially increase the GPU’s double precision capability, add error corrected memory, and include hardware support for C++ features.
Which brings us back to Fixstars. The company’s new YDEL for CUDA offering is aimed squarely at filling what it sees as a growing market for turnkey GPU-accelerated HPC on x86 clusters. Up until now, customers either built their own Linux-CUDA environments or relied upon system OEMs to provide the OS integration as part of the system. That might be fine for experimenters and big national labs who love to tweak Linux and don’t mind shuffling hardware drivers and OS kernels, but commercial acceptance will necessitate a more traditional model.
One of the challenges is that Red Hat and other commercial Linux distributions are generally tuned for mass market enterprise applications: large database and Web servers, in particular. In this type of setup, HPC workloads won’t run as efficiently as they could. With YDEL, Fixstars modified the Red Hat kernel to support a more supercomputing-like workload. The result, according to Owen Stampflee, Fixstars’ Linux Product Manager (and Terra Soft alum), is a 5 to10 percent performance improvement on HPC apps compared to other commercial Linux distributions.
Fixstars is selling YDEL for CUDA as a typical enterprise distribution, which in this case means the CUDA SDK, hardware drivers, and Linux kernel pieces are bundled together and preconfigured for HPC. A product license includes Fixstars support for both Linux and CUDA. The product contains multiple versions of CUDA, which can be selected at runtime via a setting in a configuration file or an environment variable. In addition, the YDEL comes with an Eclipse-based graphical IDE for CUDA programming. To complete the picture, Fixstars also offers end-user training and seminars on CUDA application development.
The company is especially interested in new GPU-based opportunities in financial services and medical/life sciences applications, says Camilo Lopez, Fixstars’ vice president of business development in the US. In these sectors, customers have been kicking the tires on GPU-style HPC for a few years now, with some customers set to deploy in 2010. In the finance sector, GPUs have demonstrated their ability to deliver a big performance boost for number-crunching applications like securities and equities pricing and risk modeling. But despite the “creative” methods banks have been employing with other peoples’ money over the past few years, they are quite risk-averse when it comes to adopting new technology inside their own organizations. In this environment, YDEL for CUDA could offer a convenient path to commercial deployment of GPU-equipped clusters.
Which doesn’t mean Fixstars is casting its Power and Cell BE software support adrift. According to Stampflee, there is still demand for Cell solutions in Japan and Europe. Even in the US, there are a number of Cell-based HPC deployments that are going to need lifecycle support, including the recent Air Force supercomputer installation using PS3 hardware. “A lot of the GPU demand is for upcoming projects,” says Stampflee, “whereas a lot of the Cell demand is for already mature projects.”
For commercial users, the license fee for YDEL CUDA is $400/year per server node, and that includes support. An unsupported version can be downloaded for free by students and faculty at colleges and universities. More information about the new product is available here.
Stampflee adds that support for ATI GPUs is “coming in the future” implying that a YDEL for OpenCL product could be in the offing. AMD is pursuing its own GPU computing strategy (albeit a more conservative one than NVIDIA) and is evolving its corresponding development environment toward one based on OpenCL, the new open software standard that spans multicore CPUs, GPUs, and DSPs. Fixstars has also developed an OpenCL compiler, called FOXC, whose initial implementation is targeted to multicore x86 CPUs. The company’s broader vision is to offer software, middleware and services for CUDA and OpenCL across all its supported multicore platforms.