Yellow Dog Linux licks CUDA
Nvidia GPUs sit up and bark
Remember Terra Soft and its Yellow Dog Linux for Power processors?
Well, Yellow Dog is no longer the darling Linux for Apple machines since the latter company switched to Intel Core and Xeon processors for its PCs and servers a few years back. And Terra Soft doesn't exist any more, after it was acquired by a Japanese company called Fixstars in November 2008.
But Yellow Dog is still digging in the back yard to find a cool spot to lay down, and this time around it's playing with Nvidia's CUDA programming environment for its Tesla family of GPU co-processors.
Fixstars says that it has rolled up a special release of its Linux, called Yellow Dog Enterprise Linux for CUDA, optimized to take better advantage of the Nvidia GPUs to handle calculations. Fixstars is not new to the idea of creating tools to make use of co-processors. Terra Soft was the official Linux for the "Cell" Power-derived processor created by IBM, Sony, and Toshiba - Yellow Dog was certified to run on Sony's PlayStation 3 game consoles and on blade servers made by IBM and Mercury Computer that were based on the Cell chips.
While this is all great, there's one problem. IBM killed off the next-generation QSZ2 two-socket Cell blade servers nearly two years ago (and didn't say anything until late last year about it). Big Blue has other plans for hybrid processors aimed at supercomputers after deploying Cell co-processors in a number of high-end machines, including the 1 petaflops Opteron-Cell hybrid running at Los Alamos National Laboratories. With Cell being left in the hands of Sony and Toshiba for future development, all companies that had a stake in the Cell game have to look for other platforms on which to serve their niches.
Nvidia is ramping up its "Fermi" GPU co-processor line right now, so tweaking Yellow Dog to better support Nvidia's GPUs when used as math units for central processors seems like a logical enough choice. Fixstars is not the first company to embed and support the CUDA programming environment, which simplifies the way compilers and operating systems on PCs and servers dispatch work to GPUs in their own product. Grid computing expert Platform Computing announced last August that it was embedding CUDA into its Platform Cluster Manager and HPC Workgroup Manager tools for managing grids.
With Yellow Dog Enterprise Linux for CUDA, Fixstars is embracing the combination of x64 processors and Tesla co-processors as an alternative to Cell-based systems. The distro includes multiple versions of CUDA, which programmers can toggle between as necessary depending on their hardware and applications, and it has everything programmers need to write code that takes advantage of CUDA using Fixstars Eclipse-compatible IDE. The offering also has a single support fee and one mechanism to provide that support, rather than having to deal with finger pointing between Fixstars and Nvidia. If you pay the $400 annual support license for Yellow Dog Enterprise Linux for CUDA, you point the finger at Fixstars.
The other thing that Fixstars is promising is better performance on HPC workloads using its Linux atop x64 servers with Tesla GPUs - as much as 9 per cent more oomph compared to off-the-shelf Linux distros like Red Hat Enterprise Linux or Novell SUSE Linux Enterprise Server. The reason is that Fixstars is tuning its Yellow Dog distro specifically for HPC workloads, not generic Web application and database workloads that RHEL and SLES are tuned for.
As El Reg has previously reported, Nvidia will be rolling out the Tesla 20 series of GPUs based on its "Fermi" engines in the second quarter. Fixstars says the C2050 GPU co-processor (rated at 520 gigaflops of double-precision floating point performance) will come out early in the second quarter, with the C2070 appliance (packing four of these GPUs in a 1U case for 2.52 teraflops) coming in the third quarter. That means there is plenty of time for other Linux distros to figure out how their licenses and support might meld with CUDA. Or, switch to the OpenCL alternative and GPU co-processors from Advanced Micro Devices if that doesn't work out. ®